Abstract
Introduction
Traffic-related accidents, the primary cause of fatalities among individuals between the ages of 5 and 29 (Road traffic injuries, 2020), claim the lives of more than 30,000 individuals every year in the United States, with another 2.2 million accidents resulting in injuries (Fagnant & Kockelman, 2015). Despite the transportation sector’s focus on reducing the number of vehicular accidents, the United States experienced a 6.8% increase in fatalities from road-related accidents in 2020, with the number rising from 36,355 in 2019 to 38,824 (Stewart, 2022). In 2015, the National Highway Traffic Safety Administration (NHTSA) investigated the main factors contributing to traffic accidents and found that 94% of them were caused by driver error and approximately 30% of the fatalities involved alcohol (NHTSA, 2015; Traffic Safety Facts, 2015). Autonomous vehicles (AVs) have the potential to prevent such traffic incidents and mitigate road hazards associated with human drivers, thus reducing the number of fatalities resulting from traffic accidents in the future (Gkartzonikas & Gkritza, 2019; Patel et al., 2022).
AVs utilize groundbreaking technology that holds great promise in significantly reducing fatal accidents, minimizing emissions, alleviating traffic congestion, and enhancing fuel efficiency (Anderson et al., 2014; Greenblatt & Shaheen, 2015; Kockelman et al., 2017; Pakusch et al., 2018; Wang et al., 2020). They have the ability to analyze and navigate the road environment and monitor vehicle speed and steering control (Das et al., 2020; Governing, 2022; Rosique et al., 2019; Sivak, 2014) by employing a combination of artificial intelligence; mechanical systems; and hardware and software technologies such as cameras, RADAR sensor, GPS sensor, LiDAR sensors, digital maps, and a computational framework, The Society of Automotive Engineers identified six levels of AVs that correspond to the extent to which the technology impacts driving (SAE Levels of Driving Automation, 2021): L0 indicates no driving automation with the driver in full control, L1 indicates some form of driver assistance, L2 indicates partial automation, L3 indicates conditional automation, L4 indicates full automation in a specific environment, and L5 indicates full automation in every environment (Ithal, 2020; Leilabadi & Schmidt, 2019). It is estimated that the U.S. could see annual economic benefits of up to $25 billion with a mere 10% market penetration of AVs (Fagnant & Kockelman, 2015; Khan et al., 2022).
Beyond the well-documented advantages of enhanced safety, reduced traffic congestion, and increased mobility, AVs hold the promise of delivering a multitude of benefits to society. A growing body of research underscores the profound impact AVs can have on vulnerable road users, such as individuals with disabilities. A study by Harper et al. (2016) highlights how AVs can provide newfound independence and mobility to individuals with disabilities, removing the barriers that have long constrained their transportation options. Another study by Petrović et al. (2022) further elucidates how AVs can revolutionize accessibility, offering tailored solutions for those with mobility challenges. These studies underscore the significance of AVs in not only redefining urban mobility but also in promoting inclusivity and equity, ensuring that all members of society can reap the benefits of this transformative technology. The potential of AVs to enhance safety and revolutionize the driving experience is often met with skepticism caused primarily by two key factors: excessive reliance on vehicle automation and a disregard for the guidelines and warnings provided by AV manufacturers (Biever et al., 2019; Das et al., 2020). While AVs can easily handle basic driving behaviors and tasks such as maintaining lane position, ensuring the safety of an automated system in all circumstances, particularly in complex situations like crossroads with numerous potential conflict areas, poses significant challenges (Dixit et al., 2016; Khan et al., 2023). The California Department of Motor Vehicles (CA DMV) implemented safety regulations that mandate the presence of human operators during the testing of AVs on public roads and require that AV manufacturers report any collisions resulting in property damage or major/minor injuries within a 10-day timeframe and submit disengagement reports annually (Autonomous Vehicles Collision Reports, 2022).
Existing research examines the temporal patterns of AV collisions, using data up to 2021, but neglects the investigation of spatial patterns. The continuous advancements in AV technology make it crucial to examine the most up-to-date data to gain insights into the current state of the technology landscape. This study aims to address the gaps in the literature by comprehensively analyzing a diverse range of AV collisions and examining the temporal and spatial features to gain a more nuanced understanding of the contributing factors and identify common trends. To accomplish this, a thorough examination was conducted of the AV collision reports published by the CA DMV from October 2014 to November 2022. The outcomes of this study will offer valuable insights for transportation planners and policymakers, as well as support AV manufacturers in enhancing the safety and reliability of their vehicles.
Literature Review
The automotive industry has used advanced driving assistance systems such as antilock brake systems, an electronic stability program, autonomous emergency braking, and a lane-keeping assist system to improve traffic safety over a period of many years, (Chen et al., 2021), and statistics show that they have been successful in improving vehicle safety and decreasing the incidence of traffic accidents (Jallera et al., 2019; Stilgoe, 2018). Collisions involving AVs may be triggered by a variety of factors, including design and a lack of contextual understanding, as evidenced by the recent studies that have examined the data on AV disengagements and collisions provided by the CA DMV (Pamidimukkala et al., 2023; Dadvar & Ahmed, 2021; Wang & Li, 2019).
One study that examined the collision rate data from test programs run by Google, Delphi, and Audi for autonomous vehicles between 2012 and September 2015 found that AVs had twice the crash rate of regular cars; however, the small dataset that it employed rendered it statistically inconclusive (Schoettle & Sivak, 2015). Similarly, another study compared Google AV crashes to conventional vehicles, using simple statistical analysis, and concluded that Google cars performed better than human-driven vehicles; however, the small amount of data upon which the study was based made an accurate analysis impossible (Teoh & Kidd, 2017).
Favaro et al. (2018) used descriptive statistics such as disengagement breakdowns by manufacturers, comparisons of years, mileage, and other factors to investigate the disengagements reported by AV manufacturers while performing testing on California public roads from 2014 to 2017, but because the majority of the disengagements did not result in a crash, they concluded that disengagements add a layer of safety. Boggs et al. (2019) developed a database based on CA DMV collision reports that contained content-mining stories and answers to close-ended collision issues. Their analysis showed that the majority of AV collisions occurred while the AV was in the completely independent mode, which led them to conclude that AVs in that mode are more likely to be hit than conventional vehicles or AVs in which the driver takes control. Favaro et al. (2017) examined data from September 2014 to March 2017 and concluded that rear-end collisions are the most common type of collision and that in 60% of cases, cars are minimally impacted when the speed is less than 10 mph.
Chen et al. (2020) used 94 crash reports involving AVs in California in 2019 as the basis of their study and employed ArcGIS software to create a heat map of the crashes. They concluded that most of the accidents occurred in San Francisco and Palo Alto because they were the primary test sites for AVs. In 2021, Chen et al. used 131 crash reports from California between 2019 and October 2020 to investigate the mechanisms of AV-involved crashes and assess the impact of each feature on the severity of the crash. They concluded that vehicle damage, weather conditions, accident location, and driving mode are the most critical contributors to crashes and found by drawing a heat map of the crashes that the accident sites were concentrated in the northeastern section of San Francisco.
A study by Wang and Li (2019) examined collision reports from 2014 to 2018 and concluded that AVs are only responsible for a small portion (around 6.3%) of the accidents, while other parties, including pedestrians, cyclists, motorcycles, and conventional vehicles, are passively responsible for 93.7%. Petrović et al. (2020) studied accidents that occurred from 2015 to 2017 and concluded that drivers of conventional vehicles (CVs) following too closely and at unsafe speeds increase the number of rear-end accidents. They also vowed that AVs could reduce the number of broadside and pedestrian collisions and compensate for right-of-way violations made by CV drivers. Khattak et al. (2021) examined data on disengagements from 2014 to 2018 by calibrating a nested logit model with three different outcomes; namely, disengagement that occurred from a crash, disengagement that did not result in a crash, and no disengagement that occurred from a crash. Their study concluded that factors related to AV systems (such as software failures) and other road users increase the likelihood of a disengagement without a crash and that AVs have disengaged less frequently with the maturation of the technology. The study also implied that disengagements are a part of AVs' safe performance, but that disengagement alerts may be needed to avoid certain technological failures.
At present, a key concern in the advancement of autonomous vehicles and driver assistance systems lies in their subpar performance when faced with unfavorable weather conditions like rain, snow, fog, and hail (Zang et al., 2019). A study conducted a comprehensive overview to quantify the effects of various weather conditions on the performance of the selected AV sensors. Their analysis concluded that driving in less-than-ideal conditions can reduce visibility on the road and impair the functionality of AV sensors, making them susceptible to potential accidents. However, by utilizing satellite monitoring in conjunction with existing infrastructure like traffic lights and control centers accessed through mobile devices, we can enhance collision avoidance capabilities. Consequently, leading semiconductor companies are developing chips that will enable AVs to function as mobile data centers, enabling them to make critical real-time decisions (Vargas et al., 2021). Operating AVs in rural settings presents a unique safety challenge in contrast to urban environments, necessitating the recognition of these distinctions for a comprehensive grasp of AV safety considerations (Channamallu et al., 2022). Rural roads are often narrower, winding, and poorly maintained, lacking urban infrastructure and markings. Traffic is lighter but includes slower-moving vehicles like tractors. Rural areas may have limited connectivity, affecting AVs’ access to real-time data. In contrast, urban settings offer better connectivity and quicker access to emergency services (Walters et al., 2022). These distinctions underscore the importance of acknowledging the varying safety dynamics in different settings for effective AV deployment.
The advancement of AVs is intrinsically linked to effective policy formulation and collaborative efforts between state and federal agencies, as well as AV manufacturers. Emory et al. (2022) emphasizes the significance of AV-related policies with equity implications. The study underscores the gaps in addressing the needs of marginalized communities, considerations for people with low incomes and people of color, personal security within shared vehicles, and models for deploying AVs in rural communities. This highlights the pressing need for inclusive policies that ensure equitable outcomes in the AV revolution. Taeihagh and Lim (2019) emphasize the role of governance in maximizing the benefits of AVs while mitigating potential risks. Their analysis reveals that governments have generally avoided stringent measures to promote AV development. Instead, many responses focus on the creation of councils or working groups to explore AV implications. However, the United States has taken steps to address privacy and cybersecurity issues through legislation, while the UK and Germany have enacted laws to address liability concerns. Despite these efforts, some countries are yet to implement specific strategies.
The safety risks revealed during on-road testing by disengagements and actual accidents indicate that the majority of accidents are caused by other road users. This means that alerting others and avoiding safety risks caused by other parties is vital to making safe decisions that can prevent accidents. In the future, AVs will be in every corner of every city in the United States, necessitating further research into the effects on crashes of land use in close proximity to accident sites.
Methodology
The methodology employed for this study is illustrated in Figure 1. Initially, a comprehensive search was conducted across various online databases, including Google Scholar, IEEE Xplore, Springer, ProQuest, Web of Science, and Scopus, to extract relevant literature and pinpoint areas requiring further research. In the next step, collision data from September 2014 to November 2022 was obtained from the CA DMV’s website to gather details on the make and type of vehicles involved, the time and date of the incident, the address where the incident occurred, the type of road and whether it occurred at an intersection, the number of vehicles involved, the traffic congestion level at the time of the incident, the number and severity of injuries sustained, type of collision, extent of damage, weather conditions, and road conditions. Research methodology.
The information from each incident report was entered into an excel sheet to enable quantitative analysis, and the addresses of the accidents were converted to latitude and longitude coordinates to facilitate spatial analysis. After cleaning, validation, and spatial mapping, the dataset contained 508 incident points from 2014 to 2022 that were spread across several cities and within 292 census block groups. The spatial distribution of the incidents is presented in Figure 2. Spatial distribution of incidents in California (a) San Fransico (b) Los Angeles.
In the next step, descriptive statistical analysis was performed to understand the data’s patterns and trends. The analysis included information on how the incidents evolved over time and analyzed the distribution of the number of incidents, based on factors such as the yearly distribution, collision type, vehicle type, number of vehicles involved, weather conditions, lighting conditions, roadway conditions, road type, etc. Several spatial analysis techniques were utilized to investigate the spatial and temporal patterns of the incidents. The first analysis performed was a spatial autocorrelation analysis that employed the Global Moran’s I technique and ArcGIS software to determine whether the AV incidents in a specific area were random or spatially related. A spatial cluster analysis was conducted to identify clusters and concentrations of incident locations within the state, using Maptitude, a geographic information system software developed by Caliper Corporation. A hotspot analysis was then conducted within the largest cluster to unveil any spatial relationships among the incidents. This analysis aimed to identify any areas within the cluster that exhibited higher incident density or clustering patterns.
Analysis and Results
Descriptive analysis was conducted to investigate temporal patterns, trends, and distributions of incidents; spatial analysis was conducted to explore the spatial relationships and patterns within the dataset. Combining the two analyses provided a comprehensive perspective of the dataset and revealed patterns that contributed to a more nuanced understanding of the incidents under investigation.
Descriptive Analysis
A comprehensive descriptive analysis was conducted to gain insights into the dataset, based on the yearly distribution of accidents, collision types, vehicle types, number of vehicles involved, weather and lighting conditions, roadway conditions, road types, and other relevant variables.
Annual AV Collisions
In 2014, the California DMV established the Autonomous Vehicle Tester Program, enabling companies to test AVs on public roads. The annual distribution of the 508 AV collisions reported between October 2014 and November 2022 and shown in Figure 3 reflects a gradual increase in the number of accidents from 2014 to 2017. Yearly distribution of AV collisions.
Distinct trends and patterns became apparent upon analyzing the data on AV incidents from 2014 to 2022, one of which was a consistent upward trajectory from 2014 to 2019 that is indicative of the growing number of AVs on the roads during that period. The year 2018 stands out for its significant surge in incidents that were attributed to the increased deployment of AVs for testing purposes by various companies like Waymo, Zoox, and the trend continued in 2019, underscoring persistent challenges in the development and implementation of the technology. A discernible decrease in AV incidents in 2020 can be linked to the limitations imposed by the COVID-19 pandemic that resulted in reduced vehicle miles driven overall, but the data for 2021 demonstrates a substantial rise in AV incidents compared to previous years, possibly because of increased testing and deployment, as well as shifts in driving conditions. Encouragingly, there was a slight decrease in the number of AV incidents in 2022, possibly indicating improvements in safety measures and adjustments made in response to past incidents.
AV Testing Companies
Figure 4 provides insights into the distribution of collisions among various AV testing companies. Notably, Waymo and GM Cruise emerge as significant contributors to reported collisions, largely owing to the size of their fleets and the extensive number of miles they cover compared to other reporting companies. GM Cruise operates around 300 AVs, while Waymo maintains 285 vehicles, underlining their substantial presence on the roads. Other companies also contribute to the collision figures: Zoox actively participates and engages in AV testing, while Google is closely behind, both demonstrating their dedication to advancing AV technology. It is worth noting that Apple, Lyft, Pony.ai, and others also make contributions to the reported collisions, albeit with smaller percentages. Their involvement highlights the diverse landscape of AV testing and the collective effort being made to improve safety and performance. Collision distribution - AV testing companies.
Vehicular Factors
Figure 5 shows the distributions of collision type, vehicle type, vehicle mode, number of vehicles involved in the collision, vehicle fault, and AV status at the time the collision occurred. Upon examination, it becomes evident that rear-end collisions are the most commonly occurring incidents, comprising the majority at the time of analysis. Following closely, sideswipe collisions represent the second most prevalent type, with broadside collisions and head-on collisions constituting smaller proportions. It’s worth noting that collision types such as those involving pedestrians and overturned vehicles were rarely documented. Collision distribution based on vehicular factors.
An evaluation of the modes of operation in effect at the time of collisions indicates that a significant number of incidents involved vehicles operating in autonomous mode, while others were in manual mode. Additionally, the analysis uncovers that a majority of collisions involved two vehicles, primarily mid-size and compact cars. Furthermore, approximately half of the vehicles were in motion at the time of the collision, while the remaining were stationary. An interesting finding from the analysis is that AVs were not found at fault in substantial portion of the crashes, suggesting that the majority of incidents were attributed to factors outside the control of the autonomous vehicles.
Environmental Factors
Figure 6(a) illustrates weather conditions at time of collisions. A substantial number of incidents occurred during clear weather conditions, reflecting the dominant weather pattern in the California region where the data was collected. Cloudy weather was a factor in a portion of incidents, albeit less frequently reported. Notably, weather conditions were not documented in a significant number of cases, possibly due to the predominantly sunny climate in the data collection area. The analysis presented in Figure 6(b) showcases various lighting conditions: collisions occurred predominantly during broad daylight, with a notable portion taking place at night in the presence of streetlights. Additionally, a small percentage of incidents transpired during the transitional periods of dusk or dawn. These results need to be interpreted with caution since the testing was performed in California, where the weather is predominantly sunny and the analysis of AV collisions with respect to environmental factors may not accurately represent the broader range of weather conditions encountered in other regions. Therefore, relying solely on these findings may not provide a comprehensive understanding of the impact of weather conditions on AV collisions in diverse geographic areas. Further research conducted in various climates and locations would be necessary to obtain a more representative and robust assessment of the relationship between weather conditions and AV collisions. Collision distribution based on environmental factors.
Roadway Conditions
Roadway Conditions During Collisions.
Road Type
The breakdown of AV collisions based on the types of roads on which they occur, as illustrated in Figure 7, provides insightful information. The significantly higher occurrence of collisions on streets compared to other road types highlights the importance of addressing the unique challenges posed by urban environments. Streets are typically characterized by complex and dynamic traffic conditions, including the presence of pedestrians, cyclists, and various types of vehicles sharing limited space. The high frequency of collisions on streets may be attributed to the intricate interactions and unpredictable behaviors of road users in these settings, and enhancing the capabilities of AVs to navigate such environments effectively and safely is crucial to mitigating the risks associated with them. The substantial proportion of collisions occurring on avenues suggests that the challenges encountered on wider and more arterial roads are distinct from those on streets. Avenues often feature higher traffic volumes, higher speeds, and more intricate intersection configurations, and AV technologies must adapt to these specific conditions to ensure reliable and safe operation. The low percentage of collisions on freeways or highways suggests that AVs may have demonstrated a higher level of proficiency in handling high-speed, controlled-access roadways. Collision distribution based on road type.
Spatial Analysis
This section presents the various analyses performed to examine the spatial relationships between points or events. Three methods were utilized in this study: Global Moran’s I, Local Moran’s I, and cluster analysis.
Global Moran’s I Analysis
Global Moran’s I provided an overview of the spatial autocorrelation across the entire study area and helped determine whether AV incidents in specific areas exhibited spatial relationships or were merely random. By calculating the index, we determined whether there was a significant clustering or dispersion of points/events at a global scale, which helped us identify broad patterns and trends that may have existed across the entire dataset and informed us about the overall spatial structure. Figure 8 shows that with a z-score of .50 and a Spatial autocorrelation report.
Cluster Analysis
Cluster analysis was employed to identify specific clusters with unique characteristics and shed light on localized patterns. An analysis was conducted of 508 incidents in the point layer, using Maptitude, and the results are presented in Figure 9, which showcases the centroids of the five most significant spatial clusters. The analysis utilized a Euclidean shortest-distance approach, which calculated the centroid of each cluster, based on the shortest average distance from the centroid to all the data points in the cluster. It is important to note that the centroids do not represent the exact incident locations but instead signify the central points that are closest to the incident locations within their respective clusters. Each cluster is depicted by a circle, and the numbers within indicate the actual number of incidents within each cluster. Centroids of the five most significant spatial clusters (a) San Fransico (b) Los Angeles.
The map reveals that three clusters are situated in the San Francisco area, with one cluster each in the San Jose, Los Angeles, and San Diego regions. Notably, the largest cluster is concentrated in the San Francisco area, encompasses 313 incidents, and is accompanied by two smaller clusters nearby. The second-largest cluster, comprised of 112 incidents, is located in the San Jose area, while a smaller cluster with eight incidents represents the combined areas of Los Angeles and San Diego. These findings indicate that although the data used for this analysis encompasses the entire state of California, the incidents themselves are predominantly concentrated within a smaller geographic area.
Heat Map Analysis
Local Moran’s I, commonly referred to as a heat map analysis, was employed to further investigate individual incident points within the largest identified cluster and highlight potential hotspots. This analysis helped us pinpoint specific areas of interest or concern within the studied area and facilitated targeted interventions or further investigation by examining the relationship between each incident point and its neighboring points and providing insights into localized concentrations of incidents.
Given the likelihood of a relationship at a local scale, we conducted an analysis that specifically focused on areas with the largest share of incidents, as shown in Figure 10. The results from the Maptitude heatmaps highlight distinct hot spots in the San Francisco areas of the Mission District, Japantown, Union Square, and North Beach where the highest concentration of incidents occurred and indicate the presence of localized clusters within the city. Heat map analysis.
The Mission District stands out for its lively streets and dense population, making it a hotspot for AV collisions. The district’s renown for dining, entertainment, and cultural events attracts a constant flow of visitors, resulting in heavy traffic that escalates the risk of AV incidents. Japan town, on the other hand, draws both locals and tourists with its distinctive cultural attractions and events. The consequent surge in pedestrian activity, especially during cultural festivals and shopping sprees, poses a significant navigational challenge for AVs attempting to safely traverse congested thoroughfares. Moving to Union Square, this district serves as a thriving nexus for shopping, dining, and entertainment, bolstering its economic activity. This vibrancy, coupled with frequent deliveries and a surge in taxi traffic, compounds the collision risk, thrusting AVs into the midst of intricate vehicle and pedestrian interactions. Lastly, North Beach presents its own unique set of challenges, characterized by its intricate intersections and narrow streets. AVs navigating this neighborhood must contend with a complex and tightly woven road layout, demanding precise maneuvering and navigational expertise.
Collision Distribution in Observed Risk Zones
Figures 11 and 12 presents the distributions of collision type, vehicle type, vehicle mode, number of vehicles involved in the collision, vehicle fault, AV status and road type at the time the collision occurred in observed risk zones namely, Japantown, Mission District, North Beach and Union square. Collision type, vehicle type and vehicle mode during collisions in observed risk zones. Number of vehicles involved, fault, AV status and road type during collisions in observed risk zones.
Collision patterns in Japantown differ notably from the entire dataset, marked by a reduced rate of AV-related faults, and unique vehicle and road type preferences. Despite these disparities, collision types show a relative consistency across both datasets. In contrast, the Mission District shares similarities with the entire dataset in various aspects, such as vehicle mode distribution, AV fault rates, and collision types. However, it also exhibits distinct features, notably a higher incidence of two-vehicle collisions and a greater percentage of accidents occurring on streets.
North Beach stands out from the entire dataset in several keyways: it features a lower rate of AV-related faults, a unique blend of vehicle types, a distinct distribution of collision types, and a higher occurrence of street collisions. On the other hand, Union Square showcases a set of distinct characteristics in comparison to the entire dataset. These include a significantly higher involvement of AVs, lower AV fault rates, a unique mix of vehicle types, a divergent distribution of collision types, and a higher frequency of accidents on streets.
Conclusion
The market for AVs is growing rapidly, but it is accompanied by a significant increase in safety-related issues. Collisions, which can occur due to various reasons such as technical issues, safety concerns, or the need for human intervention in complex driving situations, require serious consideration. Analyzing collision data provides valuable insights by identifying patterns and assessing performance, and analyses play a crucial role in advancing technology and guiding the development of more robust and reliable autonomous driving systems, ensuring their readiness for real-world deployment.
This study utilized data from 508 collision reports submitted by multiple manufacturers in California between October 2014 and November 2022 to conduct a comprehensive analysis of factors that may significantly contribute to collisions involving AVs. The results revealed a sharp increase in the number of collisions from 2018 forward, although AVs were not at fault in most of them. Most of the incidents involved two mid-sized or compact vehicles and were rear-end collisions, followed by sideswipes, broadside, and head-on collisions; collisions involving pedestrians or vehicles that overturned were rarely reported. A review of the environmental factors indicated that most of the collisions occurred in clear weather conditions and broad daylight, and an analysis of the roadway conditions emphasized the significance of road maintenance and the importance of agencies and AV manufacturers collaborating to address road conditions. The distribution of collisions based on road type revealed a higher occurrence on streets, indicating the unique challenges posed by urban environments. These findings align with previous studies conducted on AV collisions, lending further support to the existing body of research. This information can be used to inform the design, programming, and testing of AV technology, with a focus on addressing and mitigating the challenges identified.
The results of the spatial autocorrelation analysis indicated no related patterns of incidents in the state of California, as the incidents were distributed all across the state. Cluster analysis revealed the presence of clusters in San Francisco, San Jose, Los Angeles, and San Diego regions, with the largest cluster concentrated in San Francisco. The results of the hotspot analysis revealed notable concentrations of incidents within specific areas of San Francisco, particularly the neighborhoods of the Mission District, Japantown, Union Square, and North Beach.
Overall, the combination of descriptive and spatial analysis provided a comprehensive perspective of the dataset, uncovering temporal and spatial patterns. These findings shed light on the factors that influence AV collisions in California and provide a roadmap for future interventions and strategies aimed at enhancing safety. Collaborations between stakeholders are paramount for effectively addressing the challenges identified and should involve AV manufacturers, regulatory agencies, policymakers, and road authorities. By working together, they can develop and implement measures to monitor and address road conditions that have been shown to impact collision occurrences. Tailored approaches should be considered for urban environments, such as the identified hotspot areas, as they present unique challenges for AVs and developing specific strategies to navigate these areas safely can help mitigate collision risks. Continued monitoring of road conditions, combined with ongoing analysis of incident data, will be crucial for identifying emerging patterns, assessing the effectiveness of interventions, and refining AV technologies accordingly. By continuously adapting and improving safety measures, the overall safety of AVs on the roads can be enhanced, ensuring a safer transportation ecosystem for everyone.
It is recommended that future research examine comprehensive datasets from other states to explore the impacts of AV testing and operations, as well as the built environment factors. California is currently the only state that monitors and records data on AV testing on public highways but given the potential significance of disengagements as AV accidents’ antecedents, disengagement and collision reporting should be mandated nationwide. The findings from this study will help transportation planners identify the built environmental factors impacting AV collisions and assist them in formulating policies and legislation that will enhance traffic safety in the future.
Footnotes
Acknowledgements
Declaration of Conflicting Interests
Funding
ORCID iD
