tamakloepatternrecognitioncrashclusters2025.md 153 KB
category: literaturenote citekey: tamakloepatternrecognitioncrashclusters2025 title: Pattern recognition in crash clusters involving vehicles with advanced driving technologies authors: "Tamakloe, Reuben; Khorasani, Mahdi; Das, Subasish; Kim, Inhi" year: 2025 date: 2025-00-00 2025 doi: 10.1016/j.aap.2025.108072 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2025.108072" zotero_key: EKZSDCAS zotero_storage: GZD83N6G collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Tamakloe, Reuben"
status: converted
Contents lists available at ScienceDirect
Accident Analysis and Prevention
journal homepage: www.elsevier.com/locate/aap
Pattern recognition in crash clusters involving vehicles with advanced driving technologies
Reuben Tamakloe a , Mahdi Khorasani a , Subasish Das b , Inhi Kim a,*
- a *The Cho Chun Shik Graduate School of Mobility, Korea Advanced Institute of Science and Technology, 193 Munji*–*ro, Yuseong*–gu, Daejeon 34051, South Korea
- b Ingram School of Engineering, Texas State University, 601 University Drive, San Marcos, TX 78666, United States
ARTICLE INFO
Keywords: Automated vehicle Advanced driver assistance systems Automated driving systems Crash Machine learning
ABSTRACT
Autonomous Vehicle (AV) technologies, including Advanced Driver Assistance Systems (ADAS) and Automated Driving Systems (ADS), have significant potential to reduce crashes caused by driver errors. However, as AVs become more prevalent on roadways, the number of crashes involving them is also increasing. While considerable research has explored factors contributing to AV crashes, a gap remains in understanding the critical risk factor patterns within clusters of ADAS- and ADS-engaged AV crashes. To address this gap, this study employs the Cluster Correspondence Analysis tool to cluster crash-related factors. The analysis identified three distinct clusters for both ADAS- and ADS-engaged AV crashes. For ADAS-engaged AVs, the most representative cluster involves fatal crashes at intersections, particularly those involving left-turning vehicles. In contrast, ADS-engaged AV crashes most commonly occur in daylight and involve non-motorists. Key differences were observed: when ADAS is engaged, rear-end crashes typically result in property damage only, whereas ADS-engaged rear-end crashes are more likely to cause minor injuries. However, a notable similarity is that high-speed roads (with posted speed limits of 71 mph or higher) frequently feature animals as crash partners in both ADAS- and ADSengaged crashes. Based on these findings, it is strongly recommended to focus on infrastructural improvements alongside enhancing AV algorithms and sensor performance, particularly for non-motorist and animal detection in low-light conditions. Policymakers should prioritize driver education on safe AV operation and interaction while also mandating the installation of external human–machine interfaces to enhance AV communication with other road users and reduce rear-end crashes.
1. Introduction
1.1. Background
According to the National Highway Traffic Safety Administration (NHTSA), over 90 % of road traffic crashes stem from driver errors (NHTSA, 2015). These driver errors are categorized into recognition, decision, performance, and non-performance errors. From a national survey in the United States, recognition errors, like misjudging distance or speed, cause 41 % of crashes. Decision errors, such as speeding or tailgating, account for 34 %. Performance errors, including losing vehicle control, contribute 10 %, while non-performance errors, often health-related, make up 7 % (Bellis and Page, 2008). To mitigate crashes resulting from these errors, automobile companies and researchers have focused on developing and implementing driving automation
technologies (U.S. Department of Transportation, 2021). The concept of driving automation, according to the Society of Automotive Engineers (SAE), revolves around incorporating vehicles with features that assist driving tasks such as adaptive cruise control, lane-changing assistance, and steering operation, thereby reducing human errors and enhancing traffic safety (Antony and Whenish, 2021; Chen et al., 2019; Das et al., 2020; Tamakloe et al., 2023). The SAE International (2021) classifies driving automation into six levels. In the first three levels (Levels 0 to 2), the human drives the vehicle, although driver support features are engaged. The driver must remain attentive and ready to take control at any time. Level 0 indicates no automation, while Level 1 involves automation of either longitudinal (e.g., speed, distance) or lateral control (e.g., lane centering). Level 2 provides automation in both longitudinal and lateral control at the same time. The majority of the vehicles today operate at Level 0. Levels 3–5 assume the vehicle monitors the
E-mail addresses: drtamakloe@kaist.ac.kr (R. Tamakloe), mahdikhorasani@kaist.ac.kr (M. Khorasani), subasish@txstate.edu (S. Das), inhi.kim@kaist.ac.kr (I. Kim).
* Corresponding author.
road. Level 3 is like Level 2, but the driver only intervenes when prompted by the system. Level 4 does not require driver intervention in certain areas (e.g., controlled-access roads), while Level 5 has no operational limitations.
Over the years, the number of vehicles with driving automation, particularly SAE Level 2, has increased significantly. As of 2021, approximately 1/3 of vehicles sold in the US, China, Europe, and Japan were equipped with SAE Level 2 ADAS (Canalys, 2021; SAE International, 2021). Meanwhile, Automated Vehicles (AVs) featuring higher levels of driving automation, including SAE levels 3 to 5, such as ADS, have been undergoing active testing on open roads in the US (Huang et al., 2023b). As the market share of AVs with ADAS and ADS1 technologies grows and public road testing increases, crashes involving these technologies have risen (Huang et al., 2023b; Xu et al., 2021). The National Highway Traffic Safety Administration (NHTSA) reports a significant increase in crashes involving ADAS-equipped vehicles compared to those with ADS (Fig. 1) 2 (NHTSA, 2023). Understanding the distinct factors contributing to ADAS and ADS crashes is essential for improving AV safety and increasing public acceptance (Wicki, 2021; Zhou et al., 2023).
Several studies have explored AV safety performance and factors influencing their crashes in general, utilizing crash data (Boggs et al., 2020; Chen et al., 2023) and traffic simulation tools (Khan et al., 2022; Mihaly ´ et al., 2021). Earlier crash datasets, particularly those from the California Department of Motor Vehicles (DMV), predominantly comprised vehicles equipped with ADS (Huang et al., 2023b). As a result, researchers focused their efforts on comparing crashes involving ADS-equipped AVs with those involving human-driven vehicles (Boggs et al., 2020; Liu et al., 2021). Currently, researchers have utilized clustering tools to group all AV crashes and their sequences into distinct categories without differentiating between those with ADAS and ADS technologies despite their differences (Lee et al., 2024; Yan et al., 2023). Nevertheless, there is very little knowledge about the specific homogeneous3 clusters within ADS- and ADAS-equipped AV crash datasets. Besides, the patterns of factors associated with the crashes are unclear. While safety experts recognize the importance of understanding crash factors associated with these technologies, little focus has been given to analyzing how crash risk factor patterns vary by cluster. Previous transport safety studies suggest that crashes often form homogeneous groups, with distinct patterns for each cluster. Consequently, a comprehensive analysis of these clusters is recommended to develop targeted interventions and policies (Ashifur Rahman et al., 2022; Das et al., 2022a; Tamakloe et al., 2024b, 2023).
1.2. Study objective and contribution
This study seeks to enhance the understanding of critical patterns of factors influencing AV safety. It does so by utilizing a robust unsupervised machine learning tool known as Cluster Correspondence Analysis (CCA) to automatically identify the latent factor patterns in unique clusters existing in crashes involving AVs equipped with ADS and ADAS separately. Particular attention is being paid to analyzing those AV
crashes in which these systems were engaged, activated, or turned on at the time of the crash. Subsequently, the CCA algorithm extracts patterns of attributes most likely associated with crashes in each cluster. By observing these clusters and their patterns across different vehicle automation types, the study seeks to answer the following research questions (RQs):
- 1. What are the homogeneous clusters in the ADS & ADAS-engaged crash data? (RQ1).
- 2. What are the patterns of critical factors associated with each cluster? (RQ2).
- 3. What are the similarities and differences between the critical factor patterns associated with the various AV technologies? (RQ3).
The identification of latent patterns of factors associated with unique clusters for each AV automation type will provide valuable insights for AV researchers and policymakers, enabling them to take more targeted actions to mitigate crashes for each specific cluster and enhance the overall safety of AVs. The findings of this study can guide manufacturers in their quest to optimize AV-based algorithms. Furthermore, the study provides recommendations that can help roadway engineers identify infrastructure improvement needs. By controlling some or all the factors found to be highly associated with each crash cluster, AV safety can be improved for vehicles equipped with both ADAS and ADS technologies within each identified cluster.
The paper begins with a review of the existing literature in Section 2. It then describes the data and methodology used in the study in Section 3, followed by a presentation and discussion of the results of the study in Section 4. In Section 5, the paper concludes by summarizing the key findings and offering policy implications as well as insights into potential directions for future research.
2. Literature review
2.1. Automated driving technologies and safety implications
Automated driving technologies, including ADAS and ADS, are designed to enhance traffic safety through rapid response capabilities (Talebpour and Mahmassani, 2016). Key ADAS features, such as blind spot warning, lane departure/centering warning, and longitudinal control systems, help alert drivers to impending danger, manage vehicle speed, assist with lane changes, and maintain safe distances by adjusting to the speed of lead vehicles (NHTSA, 2018; Sohrabi et al., 2021; Zhu et al., 2020).
These capabilities are made possible through sensor fusion architecture, developed using computer vision, which enables real-time scanning and detection of objects in the vehicle's environment to identify potential risks. Additionally, vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication technologies, based on GeoNetworking, facilitate seamless information exchange, allowing vehicles to detect and avoid potential hazards (Jim´enez et al., 2016). Combined with advanced map-matching algorithms for precise vehicle localization, these automated driving technologies enhance decisionmaking and have demonstrated satisfactory results in various experimental studies (Jim´enez et al., 2016). Further, a study explored how these systems support drivers when overtaking non-motorists also revealed that these systems improve safety with regards to the overtaking behavior between cars and non-motorists (Brijs et al., 2022). When drivers fail to respond appropriately to issued warnings, the system can intervene by adjusting speed or steering, ultimately improving road safety.
Research indicates that drivers following AVs exhibit safer behavior and experience more stable traffic flow compared to those following human-driven vehicles (HDVs), leading to a reduced likelihood of crashes (Wen et al., 2022). Additionally, drivers following AVs often demonstrate significant improvements in time-to-collision (TTC) and
1 Note that, ADS, encompassing SAE Levels 3 to 5, are designed to handle all driving tasks within specific conditions without driver input but are still widely in development and are being tested. In contrast, Level 2 ADAS assist with speed and steering but require the driver to stay fully engaged at all times.
2 ADAS-equipped vehicles are involved in more crashes than ADS-equipped vehicles, as ADAS vehicles are widely available in the US fleet, whereas ADS vehicles are still in development and undergoing testing on public roads. ADS crash recording began in July 2021 (NHTSA, 2023).
3 In clustering analysis, the term homogeneous implies that the observations or elements within a cluster are highly similar to one another based on specific attributes. The objective is to maximize similarity within clusters while ensuring clear distinctiveness from other clusters.
Fig. 1. The number of AV crashes according to the type of system engaged during the crash (2020–2023).
reduced driving volatility in terms of acceleration and speed (Mahdinia et al., 2021). However, while these features offer substantial safety benefits, they may also contribute to decreased driver attentiveness and increased passive fatigue, potentially leading to crashes (Wiedemann et al., 2018; Wu et al., 2019). Furthermore, ADAS-equipped AVs can sometimes issue unnecessary alerts or take control of the vehicle against the driver's intent, leading to driver frustration (Antony and Whenish, 2021).
2.2. Key factors influencing AV crashes and injury severity
Interest in studying crashes involving AVs has been growing rapidly. Table 1 4 provides a summary of various studies examining factors associated with AV-related crashes. Researchers have employed traffic simulations and field experiments to study AV performance and interactions. Simulations, such as those using VISSIM, model HDV behaviors to evaluate the impact of AVs on traffic flow (Katrakazas et al., 2019; Papadoulis et al., 2019). Field experiments, on the other hand, observe real-world AV and HDV interactions but are often limited by controlled environments and small sample sizes (Zhao et al., 2020). While simulations risk oversimplifying traffic dynamics and producing unreliable results, field experiments face constraints due to their limited scope (Mahdinia et al., 2021). In comparison, real-world crash data from sources like NHTSA and the California DMV offer more comprehensive and practical insights into AV crash types and behaviors in urban traffic.
Numerous studies have examined AV crashes using the aforementioned crash data sources, focusing on crash typologies, contributing factors, and implications for road safety, with an emphasis on the interplay between AV decision-making and environmental elements like location and traffic signals. Yan et al. (2023) analyzed 1,849 NHTSA crash cases, finding that lateral maneuvers such as left turns were common in ADAS-related crashes but rare in ADS-related ones, reflecting differences in driver perception and system capabilities. They also identified rear-end crashes as particularly prevalent in ADAS-equipped
vehicles, noting the influence of factors like darkness and road complexity on crash likelihood and severity. Similarly, (Huang et al., 2023a) reported that rear-end crashes are common in both ADAS and ADS scenarios, driven primarily by the ego vehicle's speed and the precrash movements of other vehicles, with motion-related factors outweighing environmental factors. Q. Liu et al. (2024), using the California DMV data on 197 AV-involved crashes, identified 30 pre-crash scenarios, with rear-end crashes and lane changes as the most frequent. The authors observed that AVs are especially vulnerable to rear-end crashes, with 58.88 % caused by HDVs striking from behind. Goodall (2021) further highlights this susceptibility, particularly in urban settings, and underscores the importance of AV decision-making at intersections in shaping crash outcomes.
Novat et al. (2023) analyzed 127 crash incidents and found that AVs are 43 % more likely to be involved in rear-end crashes than HDVs but are 16 % and 27 % less likely to experience sideswipe/broadside and other crash types, respectively. Rear-end crash risk for AVs was higher at signalized intersections and on lanes with speed limits below 45 MPH. Further, distinct crash patterns were also observed from the study: AVs had a lower probability of sideswiping other crashes at unsignalized intersections and were less likely to experience front-end crashes at the segment level compared to intersections.
Biever et al. (2020) found that while AVs are often not at fault for crashes, they are frequently rear-ended during braking, turning, and gap acceptance maneuvers, underscoring the need for improved communication and anticipation with HDVs. Their study also identified a higher occurrence of side-impact crashes attributed to the conservative driving behavior of AVs compared to HDVs. Similarly, Boggs et al. (2020) observed that AV crashes are more likely in mixed land-use environments, which are associated with elevated rear-end crash rates. Das et al. (2020) identified six distinct crash patterns, highlighting the significance of factors such as turning maneuvers, multi-vehicle crashes, dark lighting conditions, and crash type in determining injury severity.
A recent study examined ADS and ADAS crashes using a two-mode social network analysis combined with an N-K model, revealing key differences between the two systems (Dong et al., 2025). The study identified that ADAS-related crashes are more frequent on highways and often involve fixed objects or larger vehicles like trucks and vans, while
4 As this study is not focused on injury severity analysis, studies on this topic were not summarized in the table.
Table 1 Overview of Studies on AV Crash Analysis Utilizing AV Crash Reports.
| Authors | Sample Size | Objective | Method | Findings |
|---|---|---|---|---|
| (Huang et al., 2024) |
246 | This study examines the interrelationships among factors in AV crashes using a graphical network based approach. Factors are categorized into environmental conditions (e.g., lighting, roadway conditions), crash-procedure-related factors (e.g., pre-crash movements, speed gap ratio), and crash outcome factors (e.g., injury severity, property damage). |
Additive Bayesian Network (ABN) Modeling with Post-ABN Regression Analysis |
• Rear-end collisions are the most common crash type involving AVs. • Pre-crash movements of HDVs have a greater impact on crash outcomes than AVs' own actions. • Crash severity is influenced by both environmental factors (e.g., operating entity) and procedural factors (e.g., crash type). • AVs tend to drive conservatively, including slower speeds at intersections, which may unintentionally increase rear-end crash risks. • Waymo-operated AVs showed lower injury rates but higher property damage incidents compared to other AVs. |
| (Lee et al., 2024) |
370 | This study aims to develop a typical crash scenario for assessing AV safety in urban areas by analyzing crash data from the California DMV. |
Hierarchical Clustering, and Association Rule Mining |
• Distinct AV crash patterns emerge, with 70 % of accidents involving rear-end collisions, primarily when AVs were stopped and HDVs were pro ceeding straight. • Sideswipe incidents (13 %) frequently occur during lane changes or unsafe turns. • 77 % of AV crashes happen at intersections, where AVs face complex traffic interactions, rather than highways with fewer obstacles. • AVs are significantly more likely to be rear-ended than to be the initiating vehicle in a crash, unlike HDVs. |
| (Q. Liu et al., 2024b) |
197 | The research aims to enhance AV safety at intersections by analyzing pre-crash scenarios to identify key contributing factors and decision making issues in rear-end and lane-change crashes. |
Association Rule Mining | • 30 AV pre-crash scenarios at intersections were analyzed, with rear-end crashes (58.88 %) and lane change crashes (16.24 %) being the most common. ADS failures in lane change crashes involved inadequate collision avoidance and overly con servative reactions, increasing crash risk. • Rear-end crashes (34.01 % linked to lead vehicle stopped − LVS) were influenced by intersection related areas, traffic signal control, autonomous mode, mixed-use/public land areas, and week day congestion. • ADS decision-making failures in rear-end crashes included abrupt stops, poor deceleration control, and hesitation in restarting. • Lane change crashes (12.69 %) were often caused by on-street parking obstructing AV sensors, peak traffic congestion, and complex four-way intersections. |
| (Sohrabi et al., 2024) |
105 | This study analyzes the accumulated distances between crashes to enable faster assessment of AV safety effectiveness. |
Duration (Survival) Model | • AVs are statistically safer than HDVs at the 95 % confidence level. • AV crash rate: 36.8 crashes per million miles, which is 20 % lower than HDVs (45.6 crashes per million miles). • Miles between crashes: AVs averaged 27,399 miles per crash, compared to 21,634 miles for HDVs, demonstrating a 27 % improvement in crash frequency. |
| (Huang et al., 2023b) |
214 | The research aims to focus on identifying the conditions under which rear-end crashes are more likely to occur in these systems. |
Binomial Logistic Regression | • Rear-end collisions were more common in ADAS crashes (52.4 %) than in ADS crashes (40 %), highlighting differences in crash dynamics. • Speed-gap ratio (SGR) significantly influenced crash likelihood, with higher speed differences increasing risk, particularly for ADS crashes on city streets. • Pre-crash movement of the crash partner (PMCP) played a major role in ADS crashes but had less impact on ADAS-related incidents. • Road type influenced ADAS crashes, with higher rear-end collision likelihood on highways, free ways, and rural roads compared to urban streets and intersections. • ADS vehicles were rear-ended more often when the crash partner was proceeding straight, while ADAS vehicles faced higher rear-end crash rates on highways and freeways. (continued on next page) |
Table 1 (continued )
| Authors | Sample Size | Objective | Method | Findings |
|---|---|---|---|---|
| (Yan et al., 2023) |
ADAS crashes: 92; ADS crashes: 110 |
This study identifies crash patterns, contributing factors, and outcomes, providing insights to improve ADS, driver training, and safety in mixed traffic environments. |
Agglomerative hierarchical clustering algorithm and Logistic Regression Model |
• Five distinct crash patterns were identified: AV maneuvers (except stopping), HDV rule violations, HDV maneuvers, hit object, and AV stopping. • Key factors influencing crash patterns included automation level, speed limit, and vehicle speed. • ADAS-involved crashes exhibited more lateral maneuvers and driver distractions compared to ADS crashes. • AV maneuvers and stopping were more frequent at lower speeds, while hitting objects was more common at higher speeds. • Crash outcomes were influenced by roadway type, model year, and incident time, with wet road conditions significantly increasing injury severity. • ADAS crashes had higher driver distraction rates and weaker perception/motion planning, making lateral maneuvers riskier. • ADS crashes were more strongly associated with |
| (Chen et al., 2023) |
445 | This study identifies factors influencing various crash types, including head-on, sideswipe, rear end, broadside, and hit-object crashes. |
Borderline Synthetic Minority Oversampling Technique, Categorical Gradient Boosting Model, and Bayesian Networks |
rear-end collisions. • AV mode significantly impacts crash types, with autonomous mode AVs more prone to rear-end crashes at intersections. • Taking-over mode AVs face a higher risk of head on crashes in poor lighting conditions. • Vehicle type plays a crucial role, as trucks and buses increase sideswipe crash risk due to visibility obstructions. • Roadway and environmental factors, including |
| (Novat et al., 2023) |
127 | This study investigates the factors contributing to different types of crashes involving AVs and conventional vehicles. |
Bayesian Network Model, and Markov Chain Monte Carlo |
lighting conditions and AADT, influence crash occurrence. • Lower AADT on major roads correlates with more rear-end crashes. • Higher AADT increases sideswipe crash probability. • AVs are 43 % more likely to be involved in rear end crashes than HDVs. • AVs are 16 % less likely to experience sideswipe or broadside crashes. • AVs are 27 % less likely to be involved in other collision types, such as head-on or hitting objects. |
| (Houseal et al., 2022) |
177 | This study examines AV crashes occurring during active autonomous mode or shortly after operator disengagement to conventional control. |
Exploratory Factor Analysis, Binary Logistic Regression, and Decision Tree |
• Key contributors to AV rear-end crashes include abrupt stops at signalized intersections and cautious driving in lower speed limit zones (< 45 mph). • AVs have a higher crash frequency on collector and local roads compared to arterials and freeways. • Bayesian Network analysis highlights that AVs face greater rear-end crash risks at signalized intersections. • Advanced lane-keeping and collision-avoidance systems help AVs reduce sideswipe crash risks. • AVs have a slightly lower probability of rear-end crashes at speed limits ≥ 45 mph compared to lower-speed roads. • Rear-end crashes are the most common AV crash type, accounting for 59.3 % of incidents, primarily occurring when AVs are stopped. • Sideswipe crashes make up 21.5 % of AV crashes. • Intersections, especially signalized ones, are the |
| location for 78.5 % of AV crashes. • AVs were not at fault in 93.8 % of cases, often being rear-ended by human drivers. • Disengagement events preceded 28.8 % of AV crashes. • Sideswipe crashes were 40 % more likely to occur post-disengagement. • EFA highlights key latent factors, including human interference, incorrect maneuvers, adverse weather conditions, and environmental awareness limitations. (continued on next page) |
Table 1 (continued )
| Authors | Sample Size | Objective | Method | Findings |
|---|---|---|---|---|
| • Binary Logistic Regression indicates that manufacturer technology influences crash likelihood. • Moving AVs, such as those changing lanes or slowing down, face higher crash risks than stationary AVs. • Decision Tree analysis reveals that 81.6 % of rear-end crashes occurred in autonomous mode. • Sideswipe and broadside crashes were more frequent after disengagement events. |
||||
| (Goodall, 2021) |
256 | This study aims to investigate and compare the struck-from-behind crash rates of AVs in autonomous mode to national rates to address a gap in research on the predominant crash type accounting for 73 % of autonomous-mode crashes. |
− - | • AVs are rear-ended 4.8 times more per mile than HDVs, with a rear-end crash rate of 17.2 per million miles compared to 3.6 per million miles for HDVs. • Even when excluding late handovers, AVs still experience 14.5 rear-end crashes per million miles. • Urban congestion significantly impacts AV rear end crash rates. • Cruise AVs in dense San Francisco have a rear end crash rate of 43.6 per million miles. • Waymo AVs in suburban Mountain View have a much lower rear-end crash rate of 10.4 per million miles. • Stopped AVs are rear-ended more frequently (8.9 per million miles) than moving AVs (8.4 per million miles). • Waymo AVs had 79 % of their rear-end crashes while stopped. • Cruise AVs had 27 % of their rear-end crashes while stopped. • HDVs had a significantly lower stopped crash rate of 0.6 per million miles (SHRP 2 data). • AVs are more vulnerable to rear-end collisions in |
| (Ashraf et al., 2021) |
251 | This study analyzes pre-crash patterns in AV involved crashes. |
Decision Tree Model, and Association Rule Mining |
traffic compared to HDVs. • Rear-end crashes were the most common AV collision type (62 %), mainly occurring when AVs were stopped or slowing at intersections due to non-AVs failing to stop in time. • Sideswipe crashes (20.2 %) were the second most frequent, often caused by non-AVs misjudging AV movements or making improper lane changes. • Broadside crashes (7.6 %) mostly happened at intersections when AVs moved straight and non AVs made sudden turns. • Head-on crashes (8.6 %) were often due to non AVs traveling in the wrong direction or making aggressive lane changes. • 57 % of AV crashes occurred in autonomous mode, where AVs were more likely to be rear ended. • 15 % of crashes happened immediately after AV disengagement. • A decision tree model identified the top 10 crash patterns, confirming that rear-end crashes occurred at intersections, sideswipes during turns, and broadside crashes in non-autonomous mode (lift value of 13.27). • Association rule mining found that rear-end crashes were 100 % likely when AVs stopped at intersections and non-AVs turned or slowed. |
| (Das et al., 2020) |
151 | This study explores critical variables in AV crashes to identify hidden trends and associated factors from extensive AV crash data and analyzes police crash narratives to accurately determine AV modes prior to crashes. |
Bayesian Latent Class Model | • Rear-end (58 cases) and sideswipe (17 cases) collisions were the most common AV crash types, with most crashes causing minor to moderate damage. • The majority (89 cases) occurred in autonomous mode, and two-vehicle crashes (128 cases) were more frequent than single- or multi-vehicle crashes. • Lighting and weather conditions played a role, with dark lighting linked to higher injury severity. • Peak crash times were Thursdays (7 AM – 12 PM, 16 cases), Fridays (1 PM – 6 PM, 14 cases), and Wednesdays (1 PM – 6 PM, 11 cases). |
Table 1 (continued )
| Authors | Sample Size | Objective | Method | Findings |
|---|---|---|---|---|
| (Boggs et al., 2020) |
124 | This study aims to identify the factors contributing to AV crashes by taking into account detailed roadway and built-environment data. |
Hierarchical-Bayes Fixed Model, and Random Parameter Logit Model |
• The Bayesian Latent Class Model identified six crash clusters, with Class 6 (64 cases) being the largest, mainly involving conventional two vehicle crashes. • Text mining revealed AV crashes were 1.45 times more likely to include phrases like "AV made contact," suggesting potential reporting biases. • Rear-end collisions are the most common AV crash type, accounting for 61.1 % of all AV involved crashes, with the likelihood 42.7 times higher when the ADS remained engaged. • Intersections significantly increase crash risk, as AVs are 6.14 times more likely to be rear-ended at intersections, while higher speed limits further elevate crash probability. • Built-environmental factors influence crash occurrence, with mixed land-use areas experi encing more rear-end crashes, whereas crashes near schools are less frequent due to smoother traffic flows and stricter enforcement. • Injury crashes make up 13.3 % of AV crashes, with higher conventional vehicle speeds increasing injury likelihood. • Marked centerlines and clear weather conditions are associated with fewer injury crashes. • Injury risks are higher at intersections, arterials, and collector roads. |
| Current Study |
2123 | This study discerns unique patterns of critical factors associated with clusters of homogeneous AV crashes by driving automation type and investigates disparities between crashes involving ADAS-activated AVs and ADS-activated AVs. |
Cluster Correspondence Analysis |
ADS crashes are more common on streets and involve vehicles or vulnerable road users. Adverse conditions like darkness with snow or rain with wet roads increase crash risks for both systems, but ADAS crashes under such conditions tend to be less severe, whereas ADS crashes more often result in injuries.
Regarding crash severity, a recent study found that while occupants of traditional vehicles are more likely to sustain severe injuries, those in AVs tend to experience lower injury severity. However, the study also revealed that when AV drivers are inattentive or disregard traffic signs, the risk of severe injuries increases significantly (Ventura et al., 2024). Furthermore, using a classification and regression tree approach to analyze AV crash data, we identified key factors influencing AV crash severity, including the year of the crash, lighting conditions, manufacturer, movement prior to the crash, and collision type (Zhu and Meng, 2022). After analyzing AV crash narratives using natural language processing and explainable artificial intelligence tools, another study found that crashes involving AVs and vulnerable road users were a prevalent topic and were strongly associated with serious injuries (Li et al., 2024). Other studies have shown that factors such as crash type, crash partners, posted speed limits, vehicle movements, manufacturer variations, adverse weather, driver type, high mileage, and dry road surfaces are strongly associated with AV crash severity (Kuo et al., 2024). Specifically, rear-end crashes were found to reduce the likelihood of severe injury for both ADS and ADAS-equipped vehicles. In the case of ADAS-equipped vehicles, airbag deployment and mileage of less than 50,000 miles were associated with a lower probability of minor injuries. However, test drivers and older AV models manufactured after 2020 increased the likelihood of minor injuries. For ADS-engaged vehicles, factors such as test drivers, front impacts, lane changes, and turning maneuvers were found to elevate the probability of minor injuries (Ding et al., 2024). Moreover, injury severity increases significantly in crashes that occur at intersections during inclement weather conditions (Chen et al., 2020).
2.3. Machine learning approaches and clustering techniques for AV crash analysis
Recent studies have begun leveraging machine learning models to explore AV crash types and better understand crash mechanisms. Using a machine learning tool, a study found that rear-end crashes are the most common type involving AVs, while side crashes occur more frequently at night. Additionally, head-on and angular crashes are often more severe, with unsignalized intersections identified as AV crash hotspots (Kohanpour et al., 2024). Focusing on intersections, another study analyzed California DMV AV crash data, revealing that inappropriate vehicle maneuvers significantly increase the likelihood of sideswipe crashes. Rear-end crashes, on the other hand, are more likely when AVs are following other vehicles on major roads with low traffic volumes (Chen et al., 2024).
Researchers have observed that traffic crashes often exhibit similarities in their occurrence and associated features. As a result, many studies have utilized clustering analyses to group crash data into homogeneous categories for further investigation (Das et al., 2022a; Tamakloe et al., 2023). In AV safety research, hierarchical clustering methods have been applied to analyze crashes. One study, for example, used hierarchical clustering to group AV crashes in California as a whole, identifying four distinct clusters. The analysis was followed by the application of an association rules mining algorithm to uncover crash scenarios within the dataset. A key finding from this study was that AV crash patterns differ significantly from those of human-driven vehicles (Lee et al., 2024). Another study encoded AV crash narratives into event sequences and employed agglomerative hierarchical clustering to group these sequences. Logistic regression was then used to examine the impact of various risk factors on AV crashes, revealing that AV speed, posted speed limits, and vehicle automation levels are significant predictors of crash patterns (Yan et al., 2023). More recently, researchers applied latent class analysis—a model-based statistical approach—to segment an AV crash dataset into two distinct clusters based on vehicle status and driving mode (Ren and Xu, 2025). In the first cluster, where
AVs were mostly in motion and operating in conventional mode, factors such as head-on collisions, intersections, and low-light conditions significantly increased the likelihood of moderate to major damage. In contrast, the second cluster, dominated by stopped AVs in autonomous mode, revealed different influencing variables, highlighting the role of unique operating contexts. Notably, the "evening" time indicator emerged as a significant predictor of crash severity only after separating the dataset, underscoring the importance of accounting for heterogeneity to identify latent patterns and improve AV safety strategies. While previous studies have effectively identified unique clusters within AV crash datasets, limited research has focused on analyzing specific clusters involving ADAS- and ADS-engaged AVs prior to crashes. Moreover, the unique patterns of factors associated with these clusters have not been thoroughly explored.
In traffic safety research, various clustering algorithms such as Kmean and hierarchical clustering approaches are commonly used, which rely on a subjective distance measure to identify clusters. As a result, the selection of the optimal number of clusters is not determined by predefined assumptions. These methods do not incorporate dimension reduction. Additionally, it does not allow researchers to identify patterns of influencing factors highly associated with each cluster.
Aside from the recent methodologies for analyzing AV-involved crashes, the CCA offers a promising approach for uncovering latent factor patterns in homogeneous clusters existing in crash data as well as their inter-cluster top associated factor patterns to aid in decisionmaking processes, in contrast to hierarchical clustering and basic Kmeans clustering tools (Tamakloe et al., 2024a; van de Velden et al., 2017). More specifically, CCA is an innovative unsupervised machine learning technique that combines K-means cluster analysis with correspondence analysis to identify and categorize patterns within datasets (Tamakloe et al., 2024b, 2023). This technique has proven to be more robust than other joint dimension reduction clustering methods (Markos et al., 2019), gaining increasing attention from researchers (Das et al., 2022b; Dzinyela et al., 2025; Tamakloe et al., 2024b, 2023). It offers a comprehensive approach to uncovering associations between factors within homogeneous data groups, making it a valuable tool for data analysis. By integrating K-means cluster analysis and correspondence analysis, CCA organizes individual attributes into coherent groups, unveiling critical factor associations within each cluster (Tamakloe et al., 2024b). As demonstrated in the work of van de Velden et al. (2017), CCA continues to revolutionize data analysis in traffic safety research, offering key insights into complex datasets.
2.4. Research gap
In summary, previous studies have examined factors associated with AV crashes and utilized clustering tools to group crashes and their sequences for further analysis (Lee et al., 2024; Yan et al., 2023). Although ADAS and ADS technologies operate differently and are likely to influence crash dynamics in distinct ways, the factors associated with specific crash clusters under these systems remain insufficiently explored. Building on existing research, this study investigates the discrepancies between crashes involving ADAS and ADS, focusing on significant differences in crash dynamics and outcomes. While prior research has highlighted key aspects, such as the prevalence of lateral maneuvers and rear-end crashes in ADAS-related incidents (Yan et al., 2023), the role of environmental factors in crash likelihood and severity (Huang et al., 2023b), and drivers' trust in ADAS technologies (Huang et al., 2023a), few studies have directly compared crash patterns between ADAS and ADS-equipped vehicles. While a prior study applied latent class analysis to identify homogeneous clusters in AV crash data using model-based statistical approaches (Ren and Xu, 2025), to the best of our knowledge, no study has yet identified the critical patterns of factors associated with the unique clusters found in ADAS- and ADS-engaged AV crashes using unsupervised machine learning clustering and intercluster pattern identification tools. This study contributes to the
literature by identifying distinct clusters and critical factor patterns within homogeneous crash groups involving AVs with ADAS and ADS technologies engaged at the time of the crash. Employing CCA, a machine learning methodology combining dimension reduction and clustering techniques, we uncover and evaluate critical factors associated with specific clusters associated with these crashes, deepening the understanding of safety implications across different levels of vehicle automation. CCA presents a robust alternative to traditional clustering methods by incorporating Multiple Correspondence Analysis (MCA) into the clustering process, allowing it to capture relationships among categorical variables. Unlike methods such as k-means, which typically assume numerical inputs or depend heavily on distance metrics, CCA uses MCA to transform categorical data into a low-dimensional space, revealing the underlying associations among variable categories. This transformation enables more meaningful interpretation of clusters, especially when performance measures like standardized residuals are used to highlight over- or underrepresented category levels (relative to global representation) within each cluster.
3. Materials and methods
3.1. Research approach
The study procedure is outlined in Fig. 2. The study utilized NHTSA's crash data, focusing on incidents where ADS or ADAS AVs were engaged at the time of the crash. The dataset was refined by removing duplicates, applying data imputation, and segmenting the data into two groups: ADS-engaged and ADAS-engaged vehicle crashes. Homogeneous clusters within each data group were identified using the CCA tool, allowing the extraction of critical factor patterns associated with each cluster. The results were compared and summarized, providing answers to the three research questions. Policy suggestions and future research directions were then presented based on the study's findings.
3.2. Cluster Correspondence analysis
CCA represents a potent analytical methodology that promises to untangle complex patterns inherent in crash observations and their interplay with decision-making processes. Being an unsupervised machine learning technique, CCA can facilitate the concurrent grouping of individual observations into meaningful clusters solely based on their attributes. This process aims to provide insights into the relationships between factors within homogeneous data groups (van de Velden et al., 2017).
Let us examine a dataset that contains data from n AV crash observations, where each observation is described by p categorical variables. Equation (1) defines Q, which is the total number of categories for all variables. The data is arranged into a comprehensive indicator matrix, Z, which has dimensions of n × Q.
$$Q = \sum_{j=1}^{p} q_j \tag{1}$$
The indicator matrix ZK can be utilized to represent cluster membership, which is considered a categorical variable. Afterwards, to examine the relationship between clusters and categorical variables, a cross-tabulation table, F, is created by comparing cluster memberships and categorical factors. The n × K indication matrix that verifies cluster membership is represented by *ZK*in this application (van de Velden et al., 2017). By applying CCA to this matrix, the analysis obtains the optimal scaling values for the rows (clusters) and columns (categories), which enhances the variance between clusters and ensures that they are accurately identified based on their distributions among the categorical variables. This process allows for simultaneous refinement. To achieve the best cluster allocation in a two-dimensional space, the clusters are divided to efficiently fit the objective function defined by Equation (2).
Fig. 2. Study approach.
$$\max{\text{clusca}}(\mathbf{Z}{K}, \mathbf{B}^{}) = \frac{1}{p} \operatorname{trace} \mathbf{B}^{} \mathbf{D}{z}^{-\frac{1}{2}} \mathbf{Z}' \mathbf{M} \mathbf{Z}{K} \mathbf{D}{K}^{-1} \mathbf{Z}{K}' \mathbf{M} \mathbf{Z} \mathbf{D}_{z}^{-\frac{1}{2}} \mathbf{B}^{*}$$ (2)
Under the stipulation that B*ʹ B* = Ik, where B denotes the column coordinate matrix and B* = 1 ̅̅̅ np √ D 1 zB, and DK indicates the diagonal matrix obtained from F = Zʹ KZ. We introduce M = In − 1*n*1ʹ n/n, and Dz. Furthermore, *Dz*signifies a diagonal matrix such that *Dz1Q* = Zʹ 1n(van de Velden et al., 2017).
The process of determining the optimal cluster allocation, denoted as ZK, revolves around maximizing Equation (1). This equation embodies the essence of the optimization challenge, focusing on maximizing the function ∅(ZK, B*) concerning ZK, which can be construed as a K-means clustering problem given a fixed B*. In this context, G represents a matrix housing cluster means. Consequently, solving the ensuing Kmeans objective, articulated as Equation (3), is necessitated (van de Velden et al., 2017):
$$\min \varnothing{clusca}(\mathbf{Z}{K}, \mathbf{G}) = |\mathbf{Y} - \mathbf{Z}_{K}\mathbf{G}|^{2}$$ (3)
Subject coordinates, encapsulated within Y, are derived through optimal scaling, expressed as ̅̅ n √ MZD− 1 2 z B* . The K-means problem is resolved to obtain matrix G = ( *Z*ʹ KZK )− 1 *Z*ʹ KY. The objective pivots on minimizing the K-means objective function, factoring both G and ZK. This entails maximizing n trace 1 pB*ʹ *D*− 1 2 z Zʹ MZKD− 1 K Zʹ KMZD− 1 2 z B* solely concerning ZK. Thus, when B* remains fixed, the cluster allocation emerges via the application of the K-means algorithm to Y (van de Velden et al., 2017).
The initial cluster allocation matrix *ZK*is generated through random assignments that comply with cluster constraints, thereby initiating the iterative procedure for CCA. Subsequently, the quantifications for the matrix (B*) are obtained through the eigenvalue decomposition of a given statement. Following that, an initial configuration for matrix Y is made. Updates for matrix *ZK*are then extracted from matrix Y using the K-means clustering technique, yielding an initial matrix for cluster updates. Iteratively working toward convergence, the procedure uses the cluster allocation matrix *ZK*before going back to the next step (van de Velden et al., 2017).
The non-parametric machine learning tool, CCA, was used to cluster ADAS- and ADS-engaged vehicle crash datasets into homogeneous groups while identifying critical factor patterns associated with each cluster. All variables in both datasets were sequentially input into the CCA model. The R package 'clustrd' (Markos et al., 2019) was employed to build the CCA model and conduct automated quantitative assessments, while visualizations were created using the 'ggplot2′ package (Hadley, 2016). Determining the optimal number of clusters and dimensions required careful evaluation to ensure accurate results. The Average Silhouette Width (ASW) criterion and the MCA K-means approach were applied to assess clustering solutions, determine the optimal number of clusters, and segregate data effectively. Cluster ranges from 2 to 10 and dimensions from 1 to 9 were tested, with 100 random starts specified in each case for ADAS and ADS datasets to minimize the risk of local minima.
3.3. Research data
The NHTSA provided the dataset for this study in Microsoft Excel's Comma-Separated Values (CSV) format. The dataset covers the period from 2019 to May 2024 (NHTSA, 2023). The dataset includes real-world crashes involving vehicles equipped with ADS and Level 2 ADAS AVs, collected under the Standing General Order issued by the NHTSA in June 2021. The Standing General Order mandates entities (operators and manufacturers) to report crashes involving ADS and Level 2 ADAS technologies promptly and transparently. Under the reporting requirements, entities must report crashes involving ADS and Level 2
ADAS vehicles that were in use within at least 30 s of the crash. Before this regulation, the NHTSA's data sources on crashes involving these vehicles were inconsistent and untimely across entities. By ensuring timely and consistent crash reports, the order enables the NHTSA to respond effectively to crashes, identify safety defects, and conduct further investigations or enforcement actions. If a safety defect is found, the NHTSA takes necessary measures to remove unsafe vehicles from roads or implement appropriate corrective actions. Additionally, the Standing General Order holds manufacturers accountable for meeting statutory safety regulations, ensuring their vehicles are defect-free and do not pose unreasonable risks. It is noteworthy that reporting entities are required to file crash reports promptly after being notified of a crash, even if the information is incomplete. While entities may quickly learn about details pertaining to specific crashes, such as airbag deployments via vehicle telemetry, they may not have details about crash conditions or seatbelt-wearing status. If additional information becomes available, they must update the report. This leads to multiple reports being submitted for the same crash. In line with the NHTSA's guidance, the study prioritizes the latest crash report version, as it provides the most information available (NHTSA, 2023).
The dataset includes crashes involving vehicles equipped with ADS (1,193 observations) and Level 2 ADAS (2,622 observations). In total, there were 3,815 observations (with many duplicates as entities reported different versions for each crash with every new information identified). The dataset contained information related to the name of the reporting entity, the vehicle information (make, model, and model year), unique vehicle identification number (VIN) and crash identifiers (ID), crash information (such as year and month of crash), and crash features (contact area of the subject vehicle or AV and crash party, weather, posted speed limit, roadway surface, highest injury severity alleged, pre-crash movement, airbag deployment, etc.).
After obtaining the dataset, preprocessing (duplicate removal and variable name coding) was necessary to prepare it for analysis. Duplicates were removed using a Python script by identifying observations with the same VIN and Crash ID. The latest report, containing more detailed crash information, was selected, and the remaining reports were discarded. After the data preparation, 2,123 crash cases remained for analysis. Since the entities provided crash reports with certain variables that had missing information, certain crash observations had some unknown crash features. As this study aimed to cluster the data into homogeneous groups to understand the patterns associated with these crashes based on crash features, imputing the missing features was necessary. To address the missing information in the remaining dataset, the study applied the MissForest algorithm, which is known for its effectiveness in data imputation (Stekhoven and Bühlmann, 2012), as supported by previous transportation-related research (Schoeppe et al., 2015; Yan et al., 2023). The MissForest algorithm, a machine learningbased imputation technique, fills in missing values based on observed data patterns. The imputation was conducted using the 'missForest' package in R (Stekhoven and Bühlmann, 2012). The tool evaluates imputation performance using two metrics: normalized root mean squared error (NRMSE) and the proportion of falsely classified entries (PFC), both ranging from 0 to 1. Lower values indicate better performance, while values closer to 1 suggest poor performance. In this analysis, the program returned an NRMSE of 0.268 and a PFC of 0.046, both closer to 0, indicating strong imputation accuracy. Following previous research, the Jensen-Shannon Divergence (JSD ∈ [0,1]) between the originally observed and imputed data was estimated to validate the accuracy of the imputation by measuring the distributional dissimilarity between the two samples. The analysis revealed JSD values of less than 0.10, highlighting a high distributional similarity between the two samples (original and imputed data). Readers interested in reading further about JSD may find further information in the literature (Yan et al., 2023).
The variable names in the dataset were long – as such, they were coded with short forms to enhance the clarity of the visualizations presented in the paper. Variables irrelevant to the study (i.e., not associated with roadway, crash, temporal, environmental, and vehicle characteristics), such as report ID, report version, reporting entity, and report type, were also eliminated from the dataset. A detailed summary of the data is provided in Table 2.
From the dataset, it was established that ADAS-engaged AV crashes constituted 71.4 % of the dataset, significantly surpassing ADS crashes, which made up 28.6 %. Most ADAS crashes involved consumer drivers (70.2 %), whereas ADS crashes were predominantly associated with commercial or test operators (65.7 %) and remote operators (28.3 %). Crashes mostly occurred on non-intersection roadways (77.6 %), with ADAS crashes being more frequent on these roads (86.1 %) and ADS crashes more common at intersections (43.5 %). While both types of crashes typically happened under dry conditions, ADS crashes were more likely to occur on dry surfaces (93.1 %) and in areas with speed limits of 0 to 30 MPH (93.9 %). ADS crashes also frequently occurred in dark but lighted conditions (35.6 %) and often involved vehicles with no passengers (34.3 %). Regarding injury severity, ADAS crashes had higher fatality rates (32.4 %) and more minor injuries (27.8 %), suggesting that ADAS-equipped vehicles often operated in more varied environments, potentially leading to increased risk. In contrast, ADS crashes primarily resulted in no injuries (88.8 %). These differences implied that while ADS may have navigated lower-speed environments safely, it faced challenges in complex urban intersections. Most crashes involved vehicles proceeding straight (72.5 %), with ADS crashes being more likely to involve rear contact and ADAS crashes showing more multiple contact points. These patterns might have indicated sudden stops or misjudgments in braking, highlighting potential areas for improvement in ADS decision-making algorithms.
4. Results and discussions
This study applied the CCA tool to explore the patterns of critical factors associated with both ADAS and ADS-engaged AV crashes. Following multiple iterations, the models for both ADAS and ADS datasets identified that the optimal solution, with the highest ASW values (ADAS: 0.126; ADS: 0.141), consisted of three homogeneous clusters in two dimensions for crashes involving ADAS- and ADSengaged AV crashes. Note that ASW is a distance based statistic ranging from 0 to 1 for measuring the quality of clustering solutions. Higher ASW values indicate greater cluster compactness and better separation. The cluster-specific ASW values for the selected solution using the ADAS dataset were 0.10, 0.19, and 0.15, while those for the ADS dataset were 0.12, 0.17, and 0.15. The objective criterion value for the analysis was 21.2474 for the ADAS model and 21.3702 for the ADS model. Table 3 presents the cluster centroid locations, within-cluster sum of squares (WCSS), and coverage information for clusters in crash data involving ADAS and ADS technologies. WCSS values measure cluster compactness; lower values indicate tighter grouping of data points around the cluster center, signifying well-defined clusters. The coverage shows the size of the clusters.
The results of the CCA tool are presented through biplots (Fig. 3), cluster-specific factor association bar plots (Fig. 4 and Fig. 5), and clustering performance metrics (Table 3). The biplots in Fig. 3(a) and Fig. 3(b)5 display the spatial arrangement of clusters in a twodimensional format, showing their proximity to the plot's centroid for ADAS-engaged and ADS-engaged AV crashes, respectively. Fig. 3 presents the biplot map generated by the program. The origin represents the
5 The dataset's variable names were lengthy, so they were shortened to improve the clarity of the visualizations in the paper. These short forms have been explained in Table 2. Including the full, unshortened names in the figures makes them appear cluttered. In the plots, the shortened attribute/variable names and the category names that are closely associated with the clusters are separated by a "." notation.
(continued on next page)
Table 2 Summary of the data used for the analysis.
| Variables | Categories | ADAS | ADS | ||||
|---|---|---|---|---|---|---|---|
| Obs. | % | Obs. | % | Obs. | % | ||
| 2123 | 100.00 | 1516 | 71.41 | 607 | 28.59 | ||
| Roadway characteristics | |||||||
| Roadway Type (rty) | Intersection | 475 | 22.37 | 211 | 13.92 | 264 | 43.49 |
| Non-intersection | 1648 | 77.63 | 1305 | 86.08 | 343 | 56.51 | |
| Roadway Surface (rsu) | Dry | 1915 | 90.2 | 1350 | 89.05 | 565 | 93.08 |
| Other | 16 | 0.75 | 14 | 0.92 | 2 | 0.33 | |
| Wet | 192 | 9.05 | 152 | 10.03 | 40 | 6.59 | |
| Roadway Description (rde) | No Unusual Conditions | 1908 | 89.87 | 1338 | 88.26 | 570 | 93.9 |
| Unusual Conditions | 215 | 10.13 | 178 | 11.74 | 37 | 6.1 | |
| Posted Speed Limit (PSL) | 0 to 30 | 911 | 42.91 | 341 | 22.49 | 570 | 93.9 |
| 31 to 50 | 531 | 25.01 | 514 | 33.91 | 17 | 2.8 | |
| 51 to 70 | 535 | 25.2 | 515 | 33.97 | 20 | 3.3 | |
| 71 and over | 146 | 6.88 | 146 | 9.63 | 0 | 0 | |
| Temporal and environmental characteristics | |||||||
| Lighting (lgt) | Dark − Lighted | 481 | 22.66 | 265 | 17.48 | 216 | 35.59 |
| Dark − Not Lighted | 263 | 12.39 | 256 | 16.89 | 7 | 1.15 | |
| Daylight | 1269 | 59.77 | 907 | 59.83 | 362 | 59.64 | |
| Other | 110 | 5.18 | 88 | 5.8 | 22 | 3.62 | |
| Weather (wth) | Clear | 1633 | 76.92 | 1122 | 74.01 | 511 | 84.19 |
| Cloudy | 289 | 13.61 | 224 | 14.78 | 65 | 10.71 | |
| Fog/Smoke | 4 | 0.19 | 4 | 0.26 | 0 | 0 | |
| Rain Snow |
52 3 |
2.45 0.14 |
28 2 |
1.85 0.13 |
24 1 |
3.95 0.16 |
|
| Other | 142 | 6.69 | 136 | 8.97 | 6 | 0.99 | |
| Month (mth) | January (JAN) | 153 | 7.21 | 111 | 7.32 | 42 | 6.92 |
| February (FEB) | 159 | 7.49 | 112 | 7.39 | 47 | 7.74 | |
| March (MAR) | 204 | 9.61 | 138 | 9.1 | 66 | 10.87 | |
| April (APR) | 190 | 8.95 | 140 | 9.23 | 50 | 8.24 | |
| May (MAY) | 159 | 7.49 | 116 | 7.65 | 43 | 7.08 | |
| June (JUN) | 129 | 6.08 | 92 | 6.07 | 37 | 6.1 | |
| July (JUL) | 174 | 8.2 | 119 | 7.85 | 55 | 9.06 | |
| August (AUG) | 190 | 8.95 | 105 | 6.93 | 85 | 14 | |
| September (SEP) | 175 | 8.24 | 119 | 7.85 | 56 | 9.23 | |
| October (OCT) | 188 | 8.85 | 137 | 9.04 | 51 | 8.4 | |
| November (NOV) | 184 | 8.66 | 151 | 9.96 | 33 | 5.44 | |
| December (DEC) | 218 | 10.27 | 176 | 11.61 | 42 | 6.92 | |
| Hour of Crash (HOC) | 0 to 5 hrs | 405 | 19.08 | 305 | 20.12 | 100 | 16.48 |
| 6 to 11 hrs | 451 | 21.24 | 311 | 20.52 | 140 | 23.06 | |
| 12 to 17 hrs | 712 | 33.54 | 492 | 32.45 | 220 | 36.24 | |
| 18 to 23 hrs | 555 | 26.14 | 408 | 26.91 | 147 | 24.22 | |
| Year of crash (yr) | 2019 | 2 | 0.09 | 2 | 0.13 | 0 | 0 |
| 2020 | 13 | 0.61 | 13 | 0.86 | 0 | 0 | |
| 2021 | 310 | 14.6 | 241 | 15.9 | 69 | 11.37 | |
| 2022 | 715 | 33.68 | 554 | 36.54 | 161 | 26.52 | |
| 2023 | 827 | 38.95 | 542 | 35.75 | 285 | 46.95 | |
| 2024 | 256 | 12.06 | 164 | 10.82 | 92 | 15.16 | |
| Vehicle characteristics | |||||||
| Mileage Group (MGP) | 0 to 18,214 | 708 | 33.35 | 466 | 30.74 | 242 | 39.87 |
| 18,214 to 42,214 | 707 | 33.3 | 580 | 38.26 | 127 | 20.92 | |
| 42,214 and over | 708 | 33.35 | 470 | 31 | 238 | 39.21 | |
| Crash characteristics | |||||||
| Highest Injury Severity Alleged | Fatality | 491 | 23.13 | 491 | 32.39 | 0 | 0 |
| Serious | 65 | 3.06 | 61 | 4.02 | 4 | 0.66 | |
| Moderate | 128 | 6.03 | 118 | 7.78 | 10 | 1.64 | |
| Minor | 476 | 22.42 | 422 | 27.84 | 54 | 8.9 | |
| No Injuries Reported | 963 | 45.36 | 424 | 27.97 | 539 | 88.8 | |
| Property Damage (pde) | No | 96 | 4.52 | 39 | 2.57 | 57 | 9.39 |
| Yes | 2027 | 95.48 | 1477 | 97.43 | 550 | 90.61 | |
| Were all Passengers Belted in the Subject Vehicle? (svb) | No | 30 | 1.41 | 19 | 1.25 | 11 | 1.81 |
| No Passengers in Vehicle | 209 | 9.85 | 1 | 0.07 | 208 | 34.27 | |
| Yes | 1884 | 88.74 | 1496 | 98.68 | 388 | 63.92 | |
| Crash With (crw) | Animal | 93 | 4.38 | 88 | 5.8 | 5 | 0.82 |
| Bus/van | 64 | 3.02 | 34 | 2.24 | 30 | 4.94 | |
| Fixed object | 348 | 16.39 | 331 | 21.83 | 17 | 2.8 | |
| Heavy Truck | 114 | 5.37 | 67 | 4.42 | 47 | 7.74 | |
| Non motorist | 50 | 2.36 | 19 | 1.25 | 31 | 5.11 | |
| Passenger Car | 610 | 28.73 | 360 | 23.75 | 250 | 41.19 | |
| Pickup Truck | 186 | 8.76 | 130 | 8.58 | 56 | 9.23 | |
| SUV | 300 | 14.13 | 201 | 13.26 | 99 | 16.31 | |
| Other | 358 | 16.86 | 286 | 18.87 | 72 | 11.86 | |
| Consumer | 1491 | 70.23 | 1490 | 98.28 | 1 | 0.17 | |
| Driver/Operator Type (dot) | In-Vehicle (Commercial / Test) | 424 | 19.97 | 25 | 1.65 | 399 | 65.73 |
11
Table 2 (continued )
| Variables | Categories | Overall Obs. 2123 |
% 100.00 |
ADAS Obs. 1516 |
% 71.41 |
ADS Obs. 607 |
% 28.59 |
|---|---|---|---|---|---|---|---|
| Remote (Commercial / Test) | 172 | 8.1 | 0 | 0 | 172 | 28.34 | |
| Other | 4 | 0.19 | 1 | 0.07 | 3 | 0.49 | |
| Crash partners features | |||||||
| Crash Partners Pre-crash Movement (cpm) | Backing | 69 | 3.25 | 2 | 0.13 | 67 | 11.04 |
| Changing Lanes | 144 | 6.78 | 111 | 7.32 | 33 | 5.44 | |
| Lane / Road Departure | 161 | 7.58 | 153 | 10.09 | 8 | 1.32 | |
| Making Left Turn | 192 | 9.04 | 162 | 10.69 | 30 | 4.94 | |
| Proceeding Straight | 479 | 22.56 | 208 | 13.72 | 271 | 44.64 | |
| Stopped | 187 | 8.81 | 172 | 11.35 | 15 | 2.47 | |
| Other | 891 | 41.97 | 708 | 46.7 | 183 | 30.15 | |
| Crash Partners (CP) Air Bags Deployed? | No | 610 | 28.73 | 93 | 6.14 | 517 | 85.17 |
| Not Applicable | 523 | 24.64 | 453 | 29.88 | 70 | 11.53 | |
| Yes | 990 | 46.63 | 970 | 63.98 | 20 | 3.3 | |
| Crash Partners Contact Area (cpc) | Bottom | 20 | 0.94 | 0 | 0 | 20 | 3.29 |
| Front | 111 | 5.23 | 19 | 1.25 | 92 | 15.16 | |
| Left | 53 | 2.5 | 14 | 0.92 | 39 | 6.43 | |
| Multiple | 1471 | 69.29 | 1312 | 86.54 | 159 | 26.19 | |
| Rear | 430 | 20.25 | 163 | 10.75 | 267 | 43.99 | |
| Right | 37 | 1.74 | 7 | 0.46 | 30 | 4.94 | |
| Top | 1 | 0.05 | 1 | 0.07 | 0 | 0 | |
| Subject vehicles characteristics | |||||||
| Subject Vehicles Pre-crash Movement (svm) | Lane / Road Departure | 46 | 2.17 | 39 | 2.57 | 7 | 1.15 |
| Making Left Turn | 49 | 2.31 | 25 | 1.64 | 24 | 3.95 | |
| Proceeding Straight | 1540 | 72.54 | 1299 | 85.69 | 241 | 39.7 | |
| Stopped | 260 | 12.24 | 21 | 1.39 | 239 | 39.37 | |
| Other | 228 | 10.74 | 132 | 8.71 | 96 | 15.82 | |
| Subject Vehicles Air Bags Deployed? (sva) | No | 783 | 36.88 | 187 | 12.33 | 596 | 98.19 |
| Not Applicable | 6 | 0.28 | 3 | 0.2 | 3 | 0.49 | |
| Yes | 1334 | 62.84 | 1326 | 87.47 | 8 | 1.32 | |
| Subject Vehicle Was Towed? (svt) | No | 497 | 23.41 | 34 | 2.24 | 463 | 76.28 |
| Yes | 1626 | 76.59 | 1482 | 97.76 | 144 | 23.72 | |
| Subject Vehicles Contact Area (svc) | Bottom | 16 | 0.75 | 2 | 0.13 | 14 | 2.3 |
| Front | 84 | 3.96 | 54 | 3.56 | 30 | 4.94 | |
| Left | 51 | 2.4 | 20 | 1.32 | 31 | 5.11 | |
| Multiple | 1493 | 70.33 | 1316 | 86.81 | 177 | 29.16 | |
| Rear | 448 | 21.1 | 118 | 7.78 | 330 | 54.37 | |
| Right | 27 | 1.27 | 6 | 0.4 | 21 | 3.46 | |
| Top | 4 | 0.19 | 0 | 0 | 4 | 0.66 | |
| Subject Vehicles Precrash Speed Group (ssg) | 0 to 30 | 526 | 24.78 | 53 | 3.5 | 473 | 77.92 |
| 31 to 50 | 486 | 22.89 | 388 | 25.59 | 98 | 16.15 | |
| 51 to 70 | 1080 | 50.87 | 1054 | 69.52 | 26 | 4.28 | |
| 71 and over | 31 | 1.46 | 21 | 1.39 | 10 | 1.65 |
Table 3 Descriptors of clusters from both ADAS and ADS crashes.
| Data | Cluster | Cluster centroid locations | | Within cluster sum of Squares | Size | Coverage (%) | | |---------------------------------------|------------------------|----------------------------|-------------|-------------------------------|------|--------------|--| | | | Dimension 1 | Dimension 2 | | | | | | ADAS-engaged vehicle-involved crashes | Cluster 1 (C1 – red) | 0.0041 | 0.0255 | 0.0035 | 781 | 51.5 | | | | Cluster 2 (C2 – green) | − 0.0277 | − 0.0233 | 0.0028 | 531 | 35.0 | | | | Cluster 3 (C3 – blue) | 0.0565 | − 0.0369 | 0.0025 | 204 | 13.5 | | | | Sum | | | | 1516 | 100.0 | | | ADS-engaged vehicle-involved crashes | Cluster 1 (C1 – red) | − 0.0026 | 0.0328 | 0.0030 | 355 | 58.5 | | | | Cluster 2 (C2 – green) | − 0.0274 | − 0.0487 | 0.0018 | 199 | 32.8 | | | | Cluster 3 (C3 – blue) | 0.1207 | − 0.0367 | 0.0018 | 53 | 8.7 | | | | Sum | | | | 607 | 100.0 | |
average profile, while all other points indicate deviations from this baseline. The two-dimensional visualization clearly distinguishes three separate clusters. From the results, Cluster 1 is the largest for both ADAS and ADS crashes, while Cluster 3 is the smallest and most compact, showing the lowest variability in data points. Conversely, Cluster 1 exhibits slightly less cohesion and higher variability compared to Cluster 3, evidenced by its higher WCSS values compared to Clusters 2 and 3.
The CCA tool identified key cluster-specific factors for ADAS- and ADS-engaged crashes, as shown in Figs. 4 and 5. These figures display the top 30 factors with the largest standardized residuals (positive and
negative) per cluster. In CCA, standardized residuals measure the deviation of an attribute's observed frequency within a cluster from its expected frequency under the assumption of no association. A large positive residual indicates that the attribute is overrepresented, meaning it strongly characterizes or defines the cluster. Conversely, a large negative residual signifies underrepresentation, suggesting that the attribute occurs less frequently than expected within the cluster.
These figures use bar plots to depict the degree of association between risk factors and each cluster. The x-axis represents feature importance, while the y-axis displays the number of features analyzed.
Fig. 3. Biplots showing cluster centroids in the ADAS and ADS-activated vehicle crash data. Attribute labels corresponding to Table 2, along with their specific categories, are shown following the "." notation. Cluster 1 points are represented by circles, Cluster 2 by triangles, and Cluster 3 by squares, with each cluster visually grouped together.
Fig. 4. Highly important variables in the clusters of ADAS-engaged AV crashes. The y-axis represents the number of bars while the x-axis represents the importance of the variable within the cluster.
Fig. 4. (continued).
The study focused on the top 30 most important features of each cluster. Among the 30 feature bars (labeled with feature names), those extending to the right indicate the feature's association with the cluster, whereas those extending to the left signify disassociation. The farther a bar extends to the right, the stronger the association of the corresponding feature with the cluster, and the farther it extends to the left, the stronger the disassociation.
4.1. ADAS-engaged AV crash clusters
4.1.1. Cluster 1 (51.5 %)
The first cluster of ADAS-involved AV crashes is strongly associated with fatal crashes occurring in the evening hours, particularly involving crash partner vehicles making left turns at intersections (Fig. 4(a)). They typically occur under specific environmental conditions, such as daylight and cloudy weather. Intersections are dangerous places where many vehicles interact. This increased danger, coupled with inclement weather conditions, is likely to amplify the risk of severe crashes, as reduced visibility and complex vehicle interactions can lead to more critical crashes, especially during times of high traffic and lower light levels (Xu et al., 2019).
From the results, speed-related variables indicate that these fatal crashes occurred at moderate speeds, with posted speed limits and precrash speeds of 31 to 50 MPH. As ADAS assists drivers with driving tasks and provides warnings and intervening when executing lateral and or longitudinal control, it is likely that these drivers may become less attentive during the driving tasks (Antony and Whenish, 2021). These behaviors may increase the risk of fatal crashes.
Critical patterns identified from this cluster reveal that fatal crashes involving ADAS-engaged vehicles frequently occur with crash partner vehicles such as passenger cars and SUVs. These partner vehicles often cause fatal crashes in the evening periods after abruptly stopping or departing lanes at intersections. Such sudden maneuvers can elevate the risk of severe crashes, particularly for ADAS-engaged vehicles as drivers of these vehicles may be distracted by secondary tasks, reducing their attention to driving and compromising their ability to respond effectively to unexpected events when the need to take over control arises (Hungund et al., 2021). This finding may also be attributed to limitations in ADAS, such as inaccuracies in object detection under poor lighting or challenging environmental conditions. It is possible that the system failed to detect sudden maneuvers by crash partners, contributing to the higher occurrence of crashes during these periods. The literature suggests that ADAS sensors are less effective at detecting objects in poor weather conditions, such as rain or snow. Additionally, reliance on these systems may lead to driver desensitization or reduced attention to the road, increasing the risk of crashes (Lee and Kwon, 2019).
Overall, this pattern highlights the challenges ADAS AVs face in navigating complex traffic scenarios in adverse weather and managing right-of-way issues during turning maneuvers. To reduce the occurrence of these crashes, it is important to educate drivers on the need to stay focused when driving vehicles with ADAS engaged particularly at intersections. ADAS-equipped AVs require enhanced capabilities for sudden stops, lane departures, and complex traffic in adverse weather. Manufacturers should improve intersection-specific features like object detection and predictive braking and enforce stricter testing standards to
Fig. 5. Highly important variables in the clusters of ADS-engaged AV crashes. The y-axis represents the number of bars while the x-axis represents the importance of the variable within the cluster.
Fig. 5. (continued).
ensure safety and reliability. Roadway engineers might consider implementing advanced intersection management systems, such as smart traffic signals and real-time data-sharing platforms, to enhance vehicle coordination.
4.1.2. Cluster 2 (35.0 %)
The second cluster of ADAS-involved AV crashes is strongly linked to minor injury crashes occurring in dark or unlighted conditions, primarily during midnight and the early morning hours and on high speed roads (71 MPH and over) (Fig. 4(b)). From the critical cluster-specific factors, it was identified that these crashes mostly involve ADASengaged vehicles colliding with animals and occasionally with fixed objects on the road. This pattern aligns with known limitations of ADAS in low-visibility conditions. During midnight and early morning hours, reduced ambient lighting diminishes sensor effectiveness, particularly for vision-based detection systems, making it more challenging for AVs with ADAS to identify and react to unexpected obstacles, such as wildlife crossing the road (Schoner et al., 2024). Additionally, lower traffic volumes at these hours mean fewer visual or radar cues from other vehicles, potentially reducing the system's situational awareness. The impact of high speeds further exacerbates the situation, as the time available for the vehicle to detect, process, and respond to sudden hazards is significantly shortened. This aligns with previous research indicating that nighttime animal-vehicle collisions are more likely to result in minor injuries, primarily due to late detection and limited evasive maneuvering options at high speeds (Al-Bdairi et al., 2020).
Overall, these findings suggest that ADAS-equipped AVs may require enhancements in nighttime object detection and collision avoidance capabilities, particularly on high-speed roads where the consequences of late detection are more pronounced. Encouraging drivers to exercise caution on high-speed roads, particularly when ADAS is activated, is crucial. At the same time, manufacturers should prioritize the development of advanced sensors and adaptive lighting systems to enhance object detection and mitigate crash risks in low-visibility conditions, thereby improving safety on high-speed roads, especially at night. Additionally, improving roadway lighting, integrating advanced obstacle detection technology, and enforcing stricter speed management protocols could help reduce the frequency of these crashes.
4.1.3. Cluster 3 (13.5 %)
Cluster 3, representing 13.5 % of ADAS-engaged AV crashes, features rear-end crashes that cause property damage crashes but no injuries (Fig. 4(c)). These crashes typically occurred during daylight hours, between 12 PM and 5 PM, with the AVs traveling at low speeds (0–30 MPH) and being struck from behind. Many of the vehicles involved are commercial or test vehicles, suggesting professional or testing contexts. The crash partners are usually heavy trucks involved in straight or lanechanging maneuvers, striking the rear of the AVs with their front ends.
The observed pattern aligns with prior studies indicating that AVs are particularly vulnerable to rear-end collisions when stopped in traffic with the automated system engaged, as human drivers in following vehicles may fail to anticipate the AV's stopping behavior (Das et al., 2023). Furthermore, it has been shown that these rear-end crashes are often associated with drivers of ADAS-engaged vehicles over-relying on autonomous technology, engaging in secondary tasks, and failing to take control until critical situations arise. This delay in intervention can result in hard braking, leading to rear-end crashes (Huang et al., 2023b).
The finding that property damage crashes are more likely to occur during daylight hours is plausible due to the improved visibility during this time. With clearer environmental observation, the AV's perception and sensing systems can more effectively detect potential hazards (Tamakloe and Park, 2023; Zhang et al., 2023). Additionally, drivers are better positioned to intervene and take control if necessary, reducing the likelihood of fatal outcomes. This enhanced visibility and the opportunity for driver intervention contribute to the predominance of noninjury, property damage crashes observed in daylight. Previous
research also shows that rear-end crashes at low speeds often result in property damage without severe injuries (Islam, 2016). Further, heavy trucks are likely to be common crash partners in rear-end crashes due to their size and slower acceleration, making them more likely to strike the rear of slower-moving AVs. A policy implication is that greater emphasis should be placed on enhancing ADAS AVs' ability to detect and respond to heavy trucks in dynamic traffic scenarios. This could involve improving sensor capabilities, refining algorithms to better predict heavy truck movements considering their blind spot issues, and prioritizing conducting targeted testing in high-density traffic conditions to reduce the likelihood of rear-end crashes. Drivers following or riding in ADAS-engaged vehicles should be educated to maintain safe distances to prevent crashes during sudden braking.
4.2. ADS-engaged AV crash clusters
4.2.1. Cluster 1 (58.5 %)
The first cluster of ADS-involved AV crashes (Fig. 5(a)), mainly involving commercial or test vehicles, predominantly occurred in 2021 and 2022, with a notable rise in February. These crashes are most common during daylight hours, especially between 12 PM and 5 PM. The passenger(s) in the ADS-engaged vehicle had their seatbelts on during the crash, which may have helped reduce the severity by limiting the forward movement of the occupants during the crash. Previous studies found that lack of seatbelt use is associated with more severe crashes (Kim et al., 1995; Weaver et al., 2015). Besides, minor injuries are expected due to improved visibility in the afternoon, aiding both the AV and other road users in detecting and responding to hazards (P. Liu et al., 2024). This cluster is the largest for ADS-engaged AV crashes – thus, the high frequency of crashes during the afternoon indicates potential areas for improvement in ADS performance even under optimal lighting conditions.
The critical factor patterns also reveal that the ADS-engaged vehicles were mostly rear ended by crash partners. Speed-related data shows that most of these minor injury crashes occur at speeds of 31 to 50 MPH. The findings suggest that the following vehicle (crash partner) may have misjudged the speed of the subject vehicle (ADS-engaged vehicle), resulting in a rear-end crash with the AV (Huang et al., 2023b). Lane changes by crash partners were also found to be associated with these crashes which may be due to complications in the detection and response for ADS-equipped vehicles.
Another key finding regarding this cluster is that ADS-engaged AV crashes leading to minor injuries also involve non-motorists, like cyclists or scooter riders (but not pedestrians). These non-motorists, who may be distracted, are often at fault in crashes (Tamakloe et al., 2024b) or may not have been properly detected by the ADS sensors. Due to their smaller size and lower speed, rear-end crashes involving non-motorists usually result in minor rather than severe injuries.
For manufacturers, greater emphasis is needed to enhance ADS technologies for detecting and responding to vulnerable road users, such as cyclists, through improved sensors, robust algorithms, and targeted testing. Policymakers should also establish regulations and roadway design standards that prioritize non-motorist safety.
4.2.2. Cluster 2 (32.8 %)
The second cluster (Fig. 5(b)) predominantly consists of crashes involving remote-operated vehicles used for commercial or testing purposes. A key finding is the absence of passengers in the ADS-engaged vehicle, reinforcing the remote nature of its operation. Notably, these crashes frequently occur on roads illuminated by streetlights at night, particularly during the late-night and early-morning hours (12 AM–5 AM) and evening hours (6 PM–11 PM). While artificial lighting is present, it does not fully mitigate the challenges of nighttime driving for remote operators. Limited visibility remains a critical issue, reducing both human and sensor-based hazard detection capabilities. This suggests that even with ADS, nighttime conditions may still pose significant
risks, as lower visibility can delay reaction times and increase the likelihood of crashes. Previous research highlights that while images collected under adverse conditions can be pre-processed to improve localization, addressing the challenges posed by unpredictable weather conditions remains difficult (Yurtsever et al., 2020). These findings highlight the need for further improvements in remote vehicle operation, including enhanced sensor performance in low-light conditions and more robust hazard detection systems to mitigate nighttime crash risks (Amador et al., 2022).
Key crash dynamics indicate a high prevalence of left-side crashes and involvement with heavy trucks in this cluster. This pattern highlights specific challenges in maneuvering around larger vehicles, likely due to the blind spots associated with such vehicles. Many crashes involved crash partners that were backing up prior to the crash, indicating challenges in detecting and responding to reversing vehicles in remote-operated scenarios. Additionally, the frequent deployment of airbags and the towing of ADS-engaged vehicles after these crashes highlight the severity of the impacts involved. From these findings, a policy implication is that stricter regulations and technological enhancements are needed for remote-operated vehicles, especially during nighttime operations. This includes improving sensor accuracy, developing better algorithms for detecting and predicting the movement of heavy vehicles, and implementing robust safety measures to mitigate crashes in low-visibility conditions.
4.2.3. Cluster 3 (8.7 %)
The final cluster (Fig. 5(c)) represents high-speed crashes occurring under unusual roadway conditions. These incidents are primarily characterized by bottom-side and top-side impacts, which strongly suggest the occurrence of rollovers. High-speed crashes often lead to rollovers due to the dynamics of rapid deceleration or sharp turns (Alrejjal et al., 2022; Czarnecki, 2018). This cluster is predominantly associated with non-intersection road segments, where speed limits range from 51 to 70 MPH, and in some cases, even higher. These conditions create an environment in which maintaining control of the vehicle becomes more challenging.
Additionally, the cluster shows a significant correlation with lane departures and straight-line driving. This suggests that vehicles may struggle to maintain lane discipline and vehicle control, even in relatively simple driving scenarios. The presence of encounters with animals and lane changes further underscores the safety risks posed by unexpected obstacles and high-speed conditions. The underlying reason for these challenges lies in the interaction between poor road surfaces, sudden obstacles, and high-speed travel, all of which significantly increase the likelihood of losing control.
At elevated speeds, maintaining AV control becomes increasingly difficult and exacerbates the risk of rollovers and severe crashes – as identified in the literature (Czarnecki, 2018). These findings emphasize the need for further development and improvement of ADS, particularly in handling high-speed situations and managing unforeseen obstacles, such as animals, in non-intersection areas. Enhanced system responsiveness could reduce the likelihood of rollovers and mitigate overall crash severity. Policymakers should consider implementing stricter testing guidelines for ADS, focusing on complex road conditions and unexpected events, to ensure these systems are adequately prepared to handle real-world challenges. This approach would promote greater road safety and more reliable system performance in critical situations.
5. Conclusions
5.1. Summary
Previous studies have investigated the factors contributing to AV crashes. Others have applied clustering tools to categorize AV crashes and crash sequences into distinct groups for analysis. Since ADAS and ADS technologies function differently, crashes involving these systems likely fall into unique groups shaped by distinct influencing factors. However, the specific factors associated with distinct clusters of crashes occurring under the control of both ADAS and ADS technologies remain unclear. This study advances the understanding of AV safety by analyzing crash data involving vehicles with ADAS and ADS engaged at the time of the crash. As unique patterns of factors are likely to influence different types of crashes in varying ways, this study aimed to identify distinct clusters within ADAS and ADS-engaged AV crash data and extract the critical patterns of factors associated with each cluster. The NHTSA data on AV crashes in the USA was used for this study. The study employed the unsupervised machine learning tool, CCA. This tool clustered data into homogeneous groups while identifying clusterspecific key factor patterns without the need for assumptions as in the case of regression analyses. This study contributes to improving AV safety in several ways. Specifically, manufacturers can use the findings to refine algorithms to address specific crash patterns, while policymakers can develop regulations to address identified safety gaps. Additionally, related government agencies can educate drivers on AVspecific requirements, while roadway engineers can benefit by identifying issues that can be addressed by enhancing infrastructure for AVs and other road users.
5.2. Key findings
What are the homogeneous clusters in the ADS & ADAS-engaged crash data? (RQ1)
Key findings emerged from applying the CCA algorithm to analyze the data. The algorithm identified three distinct clusters, each of crashes involving AVs with their ADAS and ADS engaged at the time of the crash (answering RQ1), along with the associated patterns of factors for each cluster. The analysis identified three clusters each for the ADAS-engaged and ADS-engaged AV crash datasets.
What are the patterns of critical factors associated with each cluster? (RQ2)
Regarding the patterns of critical factors in the identified clusters (answering RQ2), several observations were made. For the ADASengaged clusters, the largest cluster reveals a pattern of fatal left-turn crashes at intersections, often involving passenger cars or SUVs that may have stopped abruptly or experienced lane departures. These crashes typically occur at moderate speeds (31–50 MPH) and under dark or cloudy weather conditions. The second-largest ADAS-engaged cluster reveals a pattern of minor injury crashes occurring on high-speed roads (51–70 MPH or > 70 MPH) in dark or unlit conditions, often involving fixed objects or wildlife. This highlights the need for improved visibility solutions and enhanced object detection capabilities. The third ADASengaged cluster highlights property-damage rear-end crashes involving heavy trucks, either proceeding straight or changing lanes during daylight. Interestingly, ADAS-engaged vehicles are likely to both rear-end crash partners and be rear-ended by them. This suggests challenges in navigating dense traffic and complex environments.
Concerning the ADS-engaged AV crashes, the largest cluster reveals a notable pattern: crashes mostly involve seat-belted occupants in ADSengaged AVs, often resulting in minor injuries. These crashes typically occur in daylight conditions on moderate-speed roads and involve nonmotorists, such as cyclists, highlighting the system's challenges in safely interacting with vulnerable road users. Another pattern from the secondlargest ADS-engaged cluster highlights issues with remote-operated vehicles colliding with heavy trucks backing up at night, even with artificial lighting present. The AV had no occupants, and the airbags were deployed as a result of the crash. This suggests that while nighttime illumination helps visibility, it may not fully address the limitations of remote operation. Finally, the third ADS-engaged cluster revealed patterns of high-speed crashes involving ADS-engaged vehicles and animals on unusual roadways, particularly in non-intersection areas. These crashes often resulted in impacts to both the top and bottom of the AV and its crash partners, suggesting rollovers. This highlights the
challenges of maintaining control and effectively responding to unexpected obstacles in such scenarios.
What are the similarities and differences between the critical factor patterns associated with the various AV technologies? (RQ3)
Concerning the similarities of the patterns identified for both AV technologies (RQ3), the study found that crashes involving animals (as crash partners) are highly likely on high-speed roads with posted speed limits of 71 mph or higher, for both ADS and ADAS-engaged AVs. To mitigate these crashes, it is crucial to develop sensors and train algorithms capable of promptly detecting and responding to wildlife on the roadway. A few dissimilarities were identified. Crashes involving AVs and heavy trucks were common in both ADAS and ADS-engaged scenarios. However, with ADAS engaged, AVs are more prone to collide with a heavy truck moving straight or changing lanes during the day. In contrast, when ADS is engaged, AVs are more likely to crash with a heavy truck backing up on a well-lit road at night. Further, rear-end crashes are more likely to result in property damage only when ADAS is engaged but tend to lead to minor injuries when ADS is engaged. More focused attention should be given to reducing the occurrence and severity of rear-end crashes particularly for ADS AVs. Strengthening adaptive cruise control and crash avoidance systems is essential. Additionally, increasing drivers' awareness of the need to keep adequate distance from ADS-engaged AVs in the lead can be helpful in terms of preparing adequate time for sudden braking of AVs. Additionally, advocating for legislation that requires the installation of external human–machine interfaces to communicate with pedestrians, cyclists, and other drivers could help reduce these rear-end crashes.
5.3. Limitations and future work
The study has several limitations. It relies solely on NHTSA data, which is U.S.-based, and may not capture all variables influencing AV crashes or reflect the circumstances of AV crashes in other jurisdictions. Consequently, caution should be exercised in generalizing the results of this study to other international contexts. To address these limitations, future research can employ cross-national datasets and multi-regional analyses to validate our findings and explore whether the same critical factors emerge consistently across various driving environments. Additionally, while data imputation is useful for replacing missing category information and has been used in prior research, it may affect the quality of the results. Further, due to the low occurrence of fatal, serious, and moderate crashes, combined with the overwhelmingly high proportion of minor and no injury (reported) crashes, particularly in ADS-engaged scenarios, this study did not employ a regression model to evaluate the individual factors influencing injury severity outcomes. While the CCA tool provided valuable insights into factor patterns in each cluster, we acknowledge that more advanced techniques, such as deep learning and advanced statistical models, could better capture complex, non-linear relationships that may not be fully represented in our analysis. Exploring such methods falls outside the current scope of this study, but future research could incorporate these techniques for more robust identification and comparison of crash factor patterns associated with each cluster across AV technologies. Although the relatively small sample size in our AV crash dataset may limit generalizability, the current design provides insights by demonstrating how MissForest can handle non-normal, and multi-variate data. Through careful crossvalidation and performance assessment, it underscores the algorithm's utility under data-scarce conditions. We agree that future studies should extend these findings by using larger, more diverse datasets, conducting comprehensive simulation analyses, and comparing MissForest to other imputation methods to ensure its reliability and scalability in broader AV crash contexts.
CRediT authorship contribution statement
Reuben Tamakloe: Writing – review & editing, Writing – original
draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Mahdi Khorasani: Writing – review & editing, Writing – original draft, Investigation, Formal analysis. Subasish Das: Writing – review & editing, Validation, Investigation, Data curation. Inhi Kim: Writing – review & editing, Writing – original draft, Investigation, Formal analysis.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This research was supported by a grant from the Korea Agency for Infrastructure Technology Advancement (KAIA), funded by the Korean government (MOLIT) (RS-2022-00156089), and by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Education (RS-2024-00462359).
Data availability
Data will be made available on request.
References
- Al-Bdairi, N.S.S., Behnood, A., Hernandez, S., 2020. Temporal stability of driver injury severities in animal-vehicle collisions: A random parameters with heterogeneity in means (and variances) approach. Anal Methods Accid Res 26, 100120. https://doi. org/10.1016/j.amar.2020.100120.
- Alrejjal, A., Farid, A., Ksaibati, K., 2022. Investigating factors influencing rollover crash risk on mountainous interstates. J Safety Res 80, 391–398. https://doi.org/10.1016/ j.jsr.2021.12.020.
- Amador, O., Aramrattana, M., Vinel, A., 2022. A Survey on Remote Operation of Road Vehicles. IEEE Access 10, 130135–130154. https://doi.org/10.1109/ ACCESS.2022.3229168.
- Antony, M.M., Whenish, R., 2021. Advanced Driver Assistance Systems (ADAS). Eai/ springer Innovations in Communication and Computing 165–181. https://doi.org/ 10.1007/978-3-030-59897-6\_9/TABLES/2.
- Ashifur Rahman, M., Das, S., Sun, X., 2022. Using Cluster Correspondence Analysis to Explore Rainy Weather Crashes in Louisiana. Transportation Research Record: Journal of the Transportation Research Board 2676 (8), 159–173. https://doi.org/ 10.1177/03611981221082582.
- Ashraf, M.T., Dey, K., Mishra, S., Rahman, M.T., 2021. Extracting Rules from Autonomous-Vehicle-Involved Crashes by Applying Decision Tree and Association Rule Methods. Transportation Research Record: Journal of the Transportation Research Board 2675 (11), 522–533. https://doi.org/10.1177/ 03611981211018461.
- Bellis, E., Page, J., 2008. National Motor Vehicle Crash Causation Survey (NMVCCS) SAS Analytical Users Manual. Washington, DC.
- Biever, W., Angell, L., Seaman, S., 2020. Automated Driving System Collisions: Early Lessons. Human Factors: the Journal of the Human Factors and Ergonomics Society 62 (2), 249–259. https://doi.org/10.1177/0018720819872034.
- Boggs, A.M., Wali, B., Khattak, A.J., 2020. Exploratory analysis of automated vehicle crashes in California: A text analytics & hierarchical Bayesian heterogeneity-based approach. Accid Anal Prev 135, 105354. https://doi.org/10.1016/j. aap.2019.105354.
- Brijs, T., Mauriello, F., Montella, A., Galante, F., Brijs, K., Ross, V., 2022. Studying the effects of an advanced driver-assistance system to improve safety of cyclists overtaking. Accid Anal Prev 174, 106763. https://doi.org/10.1016/J. AAP.2022.106763.
- Canalys, 2021. Huge Opportunity as Only 10% of the 1 Billion Cars in Use Have ADAS Features [WWW Document]. URL https://canalys.com/newsroom/hugeopportunity-as-only-10-of-the-1-billion-cars-in-use-have-adas-features (accessed 4.1.24).
- Chen, H., Chen, H., Liu, Z., Sun, X., Zhou, R., 2020. Analysis of Factors Affecting the Severity of Automated Vehicle Crashes Using XGBoost Model Combining POI Data. J Adv Transp 2020 (1), 8881545. https://doi.org/10.1155/2020/8881545.
- Chen, S., Wang, H., Meng, Q., 2019. Designing autonomous vehicle incentive program with uncertain vehicle purchase price. Transp Res Part C Emerg Technol 103, 226–245. https://doi.org/10.1016/j.trc.2019.04.013.
- Chen, Y., Zou, Y., Kong, X., Wu, L., 2024. Investigating the impact of influential factors on crash types for autonomous vehicles at intersections. Journal of Transportation Safety & Security. https://doi.org/10.1080/19439962.2023.2289403.
Chen, Y., Zou, Y., Kong, X., Wu, L., 2023. Investigating the impact of influential factors on crash types for autonomous vehicles at intersections. Journal of Transportation Safety & Security 1–28. https://doi.org/10.1080/19439962.2023.2289403.
Czarnecki, K., 2018. Automated Driving System (ADS) High-Level Quality Requirements Analysis [WWW Document]. URL https://www.researchgate.net/profile/Krzysztof-Czarnecki-2/publication/326546953_Automated_Driving_System_ADS_High-Level_ Quality_Requirements_Analysis_-_Driving_Behavior_Safety/links/
5b549a59a6/Automated-Driving-System-ADS-High-Level-Quality-Requirements-Analysis-Driving-Behavior-Safety.pdf (accessed 3.19.25).
Das, S., Dutta, A., Tsapakis, I., 2020. Automated vehicle collisions in California: Applying Bayesian latent class model. IATSS Res. 44 (4), 300–308. https://doi.org/10.1016/j. iatssr.2020.03.001.
Das, S., Hossain, M.M., Ashifur Rahman, M., Kong, X., Sun, X., Al Mamun, G.M., 2022a. Understanding patterns of moped and seated motor scooter (50 cc or less) involved fatal crashes using cluster correspondence analysis. Transportmetrica a: Transport Science 1–21. https://doi.org/10.1080/23249935.2022.2029613.
Das, S., Kong, X., "Jack", Hossain, M.M., 2023. Exploration on prior driving modes for automated vehicle collisions. Int. J. Urban Sci. 27 (4), 622–644. https://doi.org/ 10.1080/12265934.2022.2142650.
Das, S., Tamakloe, R., Kutela, B., Hossain, A., 2022b. Pattern recognition from injury severity types of frontage roadway crashes. Journal of Transportation Safety & Security 1–22. https://doi.org/10.1080/19439962.2022.2123581.
Ding, S., Abdel-Aty, M., Barbour, N., Wang, D., Wang, Z., Zheng, O., 2024. Exploratory analysis of injury severity under different levels of driving automation (SAE Levels 2 and 4) using multi-source data. Accid Anal Prev 206, 107692. https://doi.org/ 10.1016/J.AAP.2024.107692.
Dong, S., Cui, Z., Sun, D., Ye, L., Ding, S., 2025. How do environmental and road factors impact automated vehicle crashes? An evidence from ADAS and ADS. Journal of Transportation Safety & Security. https://doi.org/10.1080/ 19439962.2024.2415957.
Dzinyela, R., Dadashova, B., Westfall, G., Das, S., Silvestri-Dobrovolny, C., Adanu, E.K., Lord, D., 2025. Analysis of motorcyclists crash severity using cluster correspondence and hierarchical binary logit models. Multimodal Transportation 4 (1), 100197. https://doi.org/10.1016/j.multra.2025.100197.
Goodall, N.J., 2021. Comparison of automated vehicle struck-from-behind crash rates with national rates using naturalistic data. Accid Anal Prev 154, 106056. https://doi. org/10.1016/j.aap.2021.106056.
Hadley, W., 2016. ggplot2: Elegant Graphics for Data Analysis. Springer-Verlag, New York.
Houseal, L.A., Gaweesh, S.M., Dadvar, S., Ahmed, M.M., 2022. Causes and Effects of Autonomous Vehicle Field Test Crashes and Disengagements Using Exploratory Factor Analysis, Binary Logistic Regression, and Decision Trees. Transportation Research Record: Journal of the Transportation Research Board 2676 (8), 571–586. https://doi.org/10.1177/03611981221084677.
Huang, C., He, D., Wen, X., Yan, S., 2023a. Beyond adaptive cruise control and lane centering control: drivers' mental model of and trust in emerging ADAS technologies. Front Psychol 14, 1236062. https://doi.org/10.3389/
Huang, C., Wang, A., Yan, S., He, D., 2024. Investigating the Interrelationships among Factors Associated with Automated Vehicle Crashes Using Additive Bayesian Network. Transportation Research Record: Journal of the Transportation Research Board. https://doi.org/10.1177/03611981241274152.
Huang, C., Wen, X., He, D., 2023b. Characteristics of Rear-End Collisions: A Comparison between Automated Driving System-Involved Crashes and Advanced Driving Assistance System-Involved Crashes. Transportation Research Record: Journal of the Transportation Research Board. https://doi.org/10.1177/03611981231209319.
Hungund, A.P., Pai, G., Pradhan, A.K., 2021. Systematic review of research on driver distraction in the context of advanced driver assistance systems. Transp Res Rec 2675 (9), 756–765. https://doi.org/10.1177/03611981211004129/ASSET/ IMAGES/LARGE/10.1177\_03611981211004129-FIG1.JPEG.
Islam, M.T., 2016. Investigating the Speed and Rear-End Collision Relationship at Urban Signalized Intersections. Transportation Research Record: Journal of the Transportation Research Board 2601 (1), 10–16. https://doi.org/10.3141/2601-02.
Jim´enez, F., Naranjo, J.E., Anaya, J.J., García, F., Ponz, A., Armingol, J.M., 2016. Advanced Driver Assistance System for Road Environments to Improve Safety and Efficiency. Transp. Res. Procedia 14, 2245–2254. https://doi.org/10.1016/j. trpro.2016.05.240.
Katrakazas, C., Quddus, M., Chen, W.-H., 2019. A new integrated collision risk assessment methodology for autonomous vehicles. Accid Anal Prev 127, 61–79. https://doi.org/10.1016/j.aap.2019.01.029.
Khan, R.R., Hanif, A., Ahmed, Q., 2022. Cooperative Navigation Strategy for Connected Autonomous Vehicle Operating at Smart Intersection. IFAC-PapersOnLine 55 (24), 279–284. https://doi.org/10.1016/j.ifacol.2022.10.297.
Kim, K., Nitz, L., Richardson, J., Li, L., 1995. Personal and behavioral predictors of automobile crash and injury severity. Accid Anal Prev 27 (4), 469–481. https://doi. org/10.1016/0001-4575(95)00001-G.
Kohanpour, E., Davoodi, S.R., Shaaban, K., 2024. Analyzing Autonomous Vehicle Collision Types to Support Sustainable Transportation Systems: A Machine Learning and Association Rules Approach. Sustainability 2024, Vol. 16, Page 9893 16 22 , 9893. doi: 10.3390/SU16229893.
Kuo, P.F., Hsu, W.T., Lord, D., Putra, I.G.B., 2024. Classification of autonomous vehicle crash severity: Solving the problems of imbalanced datasets and small sample size. Accid Anal Prev 205, 107666. https://doi.org/10.1016/J.AAP.2024.107666.
Lee, H., Kang, M., Hwang, K., Yoon, Y., 2024. The typical AV accident scenarios in the urban area obtained by clustering and association rule mining of real-world accident reports. Heliyon 10, 3. https://doi.org/10.1016/j.heliyon.2024.e25000.
Lee, Y., Kwon, J.H., 2019. Performance analysis of the GNSS / MEMS-IMU / on-Board vehicle sensor / magnetometer-based positioning system during GNSS signal
blockage. Int. J. Urban Sci. 23 (3), 434–443. https://doi.org/10.1080/ 12265934.2018.1473043.
Li, P., Chen, S., Yue, L., Xu, Y., Noyce, D.A., 2024. Analyzing relationships between latent topics in autonomous vehicle crash narratives and crash severity using natural language processing techniques and explainable XGBoost. Accid Anal Prev 203, 107605. https://doi.org/10.1016/J.AAP.2024.107605.
Liu, P., Guo, Y., Liu, P., Ding, H., Cao, J., Zhou, J., Feng, Z., 2024a. What can we learn from the AV crashes? – An association rule analysis for identifying the contributing risky factors. Accid Anal Prev 199, 107492. https://doi.org/10.1016/j.
Liu, Q., Wang, X., Liu, S., Yu, C., Glaser, Y., 2024. Analysis of pre-crash scenarios and contributing factors for autonomous vehicle crashes at intersections. Accid Anal Prev 195, 107383. https://doi.org/10.1016/J.AAP.2023.107383.
Liu, Q., Wang, X., Wu, X., Glaser, Y., He, L., 2021. Crash comparison of autonomous and conventional vehicles using pre-crash scenario typology. Accid Anal Prev 159, 106281. https://doi.org/10.1016/j.aap.2021.106281.
Mahdinia, I., Mohammadnazar, A., Arvin, R., Khattak, A.J., 2021. Integration of automated vehicles in mixed traffic: Evaluating changes in performance of following human-driven vehicles. Accid Anal Prev 152, 106006. https://doi.org/10.1016/j. aap.2021.106006.
Markos, A., D'Enza, A.I., van de Velden, M., 2019. Beyond Tandem Analysis: Joint Dimension Reduction and Clustering in R. J Stat Softw 91, 10. https://doi.org/ 10.18637/jss.v091.i10.
Mih´ aly, A., Farkas, Z., Zsuzsanna, B., G´ asp´ ar, P., 2021. Performance Analysis of Model Predictive Intersection Control for Autonomous Vehicles. IFAC-PapersOnLine 54 (2), 240–245. https://doi.org/10.1016/j.ifacol.2021.06.029.
NHTSA, 2023. Standing General Order on Crash Reporting [WWW Document]. URL https://www.nhtsa.gov/laws-regulations/standing-general-order-crashreporting#overview (accessed 7.4.23).
NHTSA, 2018. Critical Reasons for Crashes Investigated in the National Motor Vehicle Crash Causation Survey [WWW Document]. URL https://crashstats.nhtsa.dot.gov/ Api/Public/ViewPublication/812506 (accessed 3.6.24).
NHTSA, 2015. Critical Reasons for Crashes Investigated in the National Motor Vehicle Crash Causation Survey. Washington.
Novat, N., Kidando, E., Kutela, B., Kitali, A.E., 2023. A comparative study of collision types between automated and conventional vehicles using Bayesian probabilistic inferences. J Safety Res 84, 251–260.https://doi.org/10.1016/j.jsr.2022.11.001.
Papadoulis, A., Quddus, M., Imprialou, M., 2019. Evaluating the safety impact of connected and autonomous vehicles on motorways. Accid Anal Prev 124, 12–22. https://doi.org/10.1016/j.aap.2018.12.019.
Ren, Q., Xu, M., 2025. Heterogeneity in crash patterns of autonomous vehicles: The latent class analysis coupled with multinomial logit model. Accid Anal Prev 209, 107827. https://doi.org/10.1016/J.AAP.2024.107827.
SAE International, 2021. Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles [WWW Document]. accessed 4.1.24. https://www.sae.org/standards/content/j3016\_202104/.
Schoeppe, S., Duncan, M.J., Badland, H.M., Oliver, M., Browne, M., 2015. Associations between children׳s active travel and levels of physical activity and sedentary behavior. J Transp Health 2 (3), 336–342. https://doi.org/10.1016/J.
JTH.2015.05.001. Schoner, J., Sanders, R., Goddard, T., 2024. Effects of Advanced Driver Assistance Systems on Impact Velocity and Injury Severity: An Exploration of Data from the Crash Investigation Sampling System. Transp Res Rec 2678 (5), 451–462. https:// doi.org/10.1177/03611981231189740/ASSET/IMAGES/LARGE/10.1177\_
03611981231189740-FIG3.JPEG. Sohrabi, S., Khodadadi, A., Mousavi, S.M., Dadashova, B., Lord, D., 2021. Quantifying the automated vehicle safety performance: A scoping review of the literature, evaluation of methods, and directions for future research. Accid Anal Prev 152, 106003. https://doi.org/10.1016/j.aap.2021.106003.
Sohrabi, S., Lord, D., Dadashova, B., Mannering, F., 2024. Assessing the collective safety of automated vehicle groups: A duration modeling approach of accumulated distances between crashes. Accid Anal Prev 198, 107454. https://doi.org/10.1016/j. aap.2023.107454.
Stekhoven, D.J., Bühlmann, P., 2012. MissForest—non-parametric missing value imputation for mixed-type data. Bioinformatics 28 (1), 112–118. https://doi.org/ 10.1093/BIOINFORMATICS/BTR597.
Talebpour, A., Mahmassani, H.S., 2016. Influence of connected and autonomous vehicles on traffic flow stability and throughput. Transp Res Part C Emerg Technol 71, 143–163. https://doi.org/10.1016/j.trc.2016.07.007.
Tamakloe, R., Das, S., Adanu, E.K., Park, D., 2023. Key factors affecting motorcyclebarrier crash severity: an innovative cluster-regression technique. Transportmetrica a: Transport Science. https://doi.org/10.1080/23249935.2023.2230310.
Tamakloe, R., Park, D., 2023. Discovering latent topics and trends in autonomous vehicle-related research: A structural topic modelling approach. Transp Policy (oxf) 139, 1–20. https://doi.org/10.1016/J.TRANPOL.2023.06.001.
Tamakloe, R., Zhang, K., Atandzi, J., Park, D., 2024a. Examining urban delivery service user profiles and determinants of drone delivery adoption in Ghana considering usage before and after the COVID-19 pandemic. Transp Policy (oxf) 146, 279–294. https://doi.org/10.1016/j.tranpol.2023.12.004.
Tamakloe, R., Zhang, K., Hossain, A., Kim, I., Park, S.H., 2024b. Critical risk factors associated with fatal/severe crash outcomes in personal mobility device rider atfault crashes: A two-step inter-cluster rule mining technique. Accid Anal Prev 199, 107527. https://doi.org/10.1016/j.aap.2024.107527.
U.S. Department of Transportation, 2021. Automated Vehicles Comprehensive Plan. Washington, DC.
van de Velden, M., D'Enza, A.I., Palumbo, F., 2017. Cluster Correspondence Analysis. Psychometrika 82 (1), 158–185. https://doi.org/10.1007/s11336-016-9514-0.
Ventura, L., Shrestha, R., Venkataraman, N., Shankar, V., Feknssa, N., 2024. A nonlinear mixed logit model of occupant severity in autonomous vehicle crashes. Anal Methods Accid Res 44, 100348. https://doi.org/10.1016/J.AMAR.2024.100348.
Weaver, A.A., Talton, J.W., Barnard, R.T., Schoell, S.L., Swett, K.R., Stitzel, J.D., 2015. Estimated Injury Risk for Specific Injuries and Body Regions in Frontal Motor Vehicle Crashes. Traffic Inj Prev 16, S108–S116. doi: 10.1080/15389588.2015.1012664.
Wen, X., Cui, Z., Jian, S., 2022. Characterizing car-following behaviors of human drivers when following automated vehicles using the real-world dataset. Accid Anal Prev 172, 106689. https://doi.org/10.1016/j.aap.2022.106689.
Wicki, M., 2021. How do familiarity and fatal accidents affect acceptance of self-driving vehicles? Transp Res Part F Traffic Psychol Behav 83, 401–423. https://doi.org/ 10.1016/j.trf.2021.11.004.
Wiedemann, K., Naujoks, F., Worle, ¨ J., Kenntner-Mabiala, R., Kaussner, Y., Neukum, A., 2018. Effect of different alcohol levels on take-over performance in conditionally automated driving. Accid Anal Prev 115, 89–97. https://doi.org/10.1016/j. aap.2018.03.001.
Wu, Y., Kihara, K., Takeda, Y., Sato, T., Akamatsu, M., Kitazaki, S., 2019. Effects of scheduled manual driving on drowsiness and response to take over request: A simulator study towards understanding drivers in automated driving. Accid Anal Prev 124, 202–209. https://doi.org/10.1016/j.aap.2019.01.013.
Xu, C., Ding, Z., Wang, C., Li, Z., 2019. Statistical analysis of the patterns and characteristics of connected and autonomous vehicle involved crashes. J Safety Res 71, 41–47. https://doi.org/10.1016/j.jsr.2019.09.001.
Xu, Z., Jiang, Z., Wang, G., Wang, R., Li, T., Liu, J., Zhang, Y., Liu, P., 2021. When the automated driving system fails: Dynamics of public responses to automated vehicles. Transp Res Part C Emerg Technol 129, 103271. https://doi.org/10.1016/j.
Yan, S., Huang, C., He, D., 2023. A comparison of patterns and contributing factors of ADAS and ADS involved crashes. Journal of Transportation Safety & Security 1–28. https://doi.org/10.1080/19439962.2023.2284175.
Yurtsever, E., Lambert, J., Carballo, A., Takeda, K., 2020. A Survey of Autonomous Driving: Common Practices and Emerging Technologies. IEEE Access 8, 58443–58469. https://doi.org/10.1109/ACCESS.2020.2983149.
Zhang, Y., Carballo, A., Yang, H., Takeda, K., 2023. Perception and sensing for autonomous vehicles under adverse weather conditions: A survey. ISPRS J. Photogramm. Remote Sens. 196, 146–177. https://doi.org/10.1016/j. isprsjprs.2022.12.021.
Zhao, X., Wang, Z., Xu, Z., Wang, Y., Li, X., Qu, X., 2020. Field experiments on longitudinal characteristics of human driver behavior following an autonomous vehicle. Transp Res Part C Emerg Technol 114, 205–224. https://doi.org/10.1016/j. trc.2020.02.018.
Zhou, R., Huang, H., Lee, J., Huang, X., Chen, J., Zhou, H., 2023. Identifying typical precrash scenarios based on in-depth crash data with deep embedded clustering for autonomous vehicle safety testing. Accid Anal Prev 191, 107218. https://doi.org/ 10.1016/j.aap.2023.107218.
Zhu, M., Wang, Y., Pu, Z., Hu, J., Wang, X., Ke, R., 2020. Safe, efficient, and comfortable velocity control based on reinforcement learning for autonomous driving. Transp Res Part C Emerg Technol 117, 102662. https://doi.org/10.1016/j.trc.2020.102662.
Zhu, S., Meng, Q., 2022. What can we learn from autonomous vehicle collision data on crash severity? A cost-sensitive CART approach. Accid. Anal. Prev. 174, 106769. https://doi.org/10.1016/J.AAP.2022.106769.
