The rapid development of autonomous vehicle (AV) technology promises safer transportation systems. Human error contributes to up to 90% of accidents, and AVs are anticipated to significantly reduce this figure. While AVs offer potential benefits, safety risks remain a concern, evidenced by a yearly increase in AV accidents on public roads (California data 2015-2022, excluding 2020 due to the pandemic). The RAND corporation suggests that even a small improvement in AV safety (10% higher than human drivers) could prevent numerous fatalities in the US. However, limited real-world accident data makes it challenging to fully understand the differences in accident characteristics between AVs and human-driven vehicles (HDVs). This study aims to address this gap by comparing a large dataset of AV and HDV accidents to identify specific accident scenarios where AVs may be more or less prone to accidents compared to HDVs.

Existing research on AV safety presents conflicting viewpoints. Some studies indicate AVs are safer than HDVs, citing lower accident rates per million vehicle miles traveled (VMT) and a lack of fatalities in certain AV testing programs. For example, Dixit et al. (2016) compared Google self-driving cars to HDVs, finding lower fatal accident rates for the AVs. Other research suggests AVs may not always be safer than HDVs, pointing to higher accident rates per million miles in certain limited conditions (e.g., Schoettle and Sivak, 2015) and a higher frequency of rear-end accidents (Favarò et al., 2017). These studies often suffer from limitations like insufficient data, particularly matched AV and HDV data, leading to an incomplete understanding of accident characteristics and how they differ between AVs and HDVs. Previous work has also analyzed the number of accidents and injury severity for different AV levels (ADAS and ADS), with variations in sample size and data sources across studies. This study aims to address these limitations through a comprehensive analysis of a larger and more balanced dataset of both AV and HDV accidents.

This study utilizes a matched case-control design to compare AV and HDV accidents. The AV dataset included 2100 Advanced Driving Systems (ADS) and Advanced Driver Assistance Systems (ADAS) accidents (1099 level 4 ADS, 1001 level 2 ADAS) from multiple sources including the AVOID dataset (California Department of Motor Vehicles and NHTSA database). The HDV dataset comprised 35,113 accidents from the Statewide Integrated Traffic Records System (SWITRS). The analysis started with a general descriptive comparison between the full datasets of AV and HDV accidents looking at accident type, weather conditions, pre-accident vehicle movements and accident outcomes. Subsequently, a matched case-control logistic regression model was used to compare ADS (SAE Level 4) accidents with HDV accidents. Strata were created based on accident location (matched location for AV and HDV cases), day of the week and time of day. To find the optimal number of control samples for each AV case, a sensitivity analysis was performed, resulting in a 1:5 ratio (548 ADS accidents and 2740 matched HDV controls) being chosen for the final analysis. The model estimated the odds ratios for various variables, which represent the relative likelihood of an accident in AVs versus HDVs. The variables included road and environmental conditions, accident type, accident outcomes and pre-accident conditions. The statistical analysis was performed using the survival package in R.

The analysis revealed several key findings. Overall, vehicles equipped with ADS showed a generally lower risk of accidents compared to HDVs in most scenarios. The most frequent pre-crash movements for both AVs and HDVs were proceeding straight. However, significant differences emerged under specific conditions. The odds of an ADS accident were considerably higher (5.25 times) during dawn/dusk conditions compared to HDVs, suggesting potential challenges for AV sensors in adapting to changing light conditions. Similarly, the odds of an ADS accident were 1.988 times higher during turning maneuvers. Conversely, ADS accidents showed a lower risk during rainy conditions (0.335 times the odds of an HDV accident) indicating superior performance in adverse weather. The risk of rear-end and broadside accidents were significantly lower for ADS compared to HDVs (0.457 and 0.171 times respectively). Pre-accident movements such as proceeding straight, running off the road and entering a traffic lane were less likely to result in accidents for ADS compared to HDVs (0.299, 0.021, and 0.267 times respectively). The risk of moderate and fatal injuries were also lower for ADS accidents compared to HDVs (0.634 and 0.097 times respectively). The analysis also looked at the differences between ADS and ADAS accidents, revealing different patterns across weather, road conditions, pre-accident movements, and accident types. For example, ADAS accidents were more frequent during clear skies but less frequent in rainy conditions than ADS accidents. ADAS accidents also showed higher average pre-accident speeds compared to ADS accidents.

This study's findings offer valuable insights into the relative safety of AVs versus HDVs. While ADS generally showed a lower accident risk, this was not consistent across all conditions. The higher risk during dawn/dusk and turning maneuvers highlights areas requiring further technological improvement. The superior performance of ADS in rain underscores the potential of advanced sensor technologies to overcome challenges associated with adverse weather. The significantly lower risk of rear-end and broadside collisions with ADS indicates the effectiveness of advanced sensing and collision avoidance systems. The findings suggest that AV technology can improve road safety, but further research and development are needed to address specific limitations identified in challenging driving scenarios. The large sample size and matched case-control design enhance the reliability of the findings.

This study's comprehensive analysis of AV and HDV accidents using a matched case-control design reveals that while AVs with ADS generally exhibit lower accident risk than HDVs, specific conditions like dawn/dusk and turning maneuvers require further improvement. Advanced sensor technologies show promise in mitigating the effects of adverse weather. Future research should focus on improving situational awareness in complex driving scenarios and addressing limitations of current AV systems to enhance overall safety. Further analysis using more granular data on specific AV systems and ADAS features could provide even more detailed insights into specific safety aspects.

This study primarily focuses on California data, limiting the generalizability of the findings to other regions with different road conditions, traffic patterns, and driving behaviors. The analysis only included reported accidents, potentially underrepresenting less severe incidents. Detailed information about the specific functionalities activated in the AV systems during accidents was not fully analyzed. While a 1:5 matching ratio was chosen to address data imbalances, inherent limitations of observational studies remain, including the potential for unmeasured confounding factors to influence the results. Future work could explore the impact of driver behavior in the case of ADAS accidents, as manual override and conventional modes of ADAS closely resemble HDV.