Investigating Autonomous Vehicle Crashes Using Data Mining and Artificial Intelligence Tools

Abstract: Autonomous vehicles (AVs) have the potential to enhance road safety. However, integrating them into a human-dominated environment poses safety challenges, as AVs may still be prone to errors until all vehicles become autonomous. Hence, continuous monitoring and review of AV crashes are necessary to address safety issues. Current AV crash datasets, while informative, suffer from inconsistencies and missing information, impeding the safety assessment of AVs. This paper aims to improve AV crash data quality to investigate AV crashes in San Francisco. Using data mining and web scraping techniques data was collected from AVcrashes.net, which is an open-access online data platform. The collected data was enhanced by extracting variables through image processing and natural language processing (NLP).
For image processing, Google Street View images were captured from 70 random San Francisco street locations, with eight images per location obtained and saved by rotating the view every 45 degrees using a Google API script. The images were annotated for key built environment features, including intersections, traffic lights, stop signs, crosswalks, bus lanes, and bike lanes. After filtering unsuitable images, the Roboflow platform was used for training, validation, and testing, with a model based on Microsoft's COCO dataset. Confidence thresholds were fine-tuned for each detected element, such as intersections, crosswalks, and bus-bike lanes. In the NLP application, a text mining script using OpenAI’s GPT-4o-mini was developed to extract crash details, such as severity and involved parties, from crash report descriptions. The script also filtered out non-crash incidents, such as vandalism.
The extracted data was then used for descriptive analysis and to model injury severity using multinomial logistic regression, with three injury categories identified: no injury, minor injury, and major injury. The results revealed that rear and front vehicle damage, driverless operation, and sideswipe crashes are associated with a decreased probability of minor injuries, while pedestrian, scooter, bike, and motorcycle involvement in crashes and crashes located near traffic lights are associated with a higher probability of minor injuries. Moreover, adverse weather, nighttime crashes, and the involvement of pedestrians, scooters, bikes, and motorcycles are linked to a higher probability of major injuries. This study demonstrates the potential of AI tools in safety analysis, enhances the understanding of AV safety, and provides valuable insights for policymakers, manufacturers, and researchers by offering a framework to improve AV crash databases. The proposed methods, while focused on AV crashes, can be applied to any crash setting.