Autonomous Vehicle Safety: A Machine Learning Analysis of Crash Data

In the evolving landscape of the automotive sector, Autonomous Vehicles (AVs), popularly termed as self-driving cars, have gained significant traction. Their promise encompasses an array of benefits: from the potential to dramatically reduce traffic congestion, to promises of heightened mobility, and overarching safety improvements for all road users. Despite the advancements, recent crash data from California's Department of Motor Vehicles (DMV) rings alarm bells regarding AV safety. This research dives deep into understanding the predominant factors behind injury during AV accidents, assessing 441 incidents that occurred between 2014 and 2023 sourced from the California DMV. Given the skewness and inherent imbalance in the dataset, a commonplace issue in crash data where severe accidents is fortunately less common, our methodological approach encompassed a thorough evaluation of multiple statistical and machine learning models. The models in our investigative arsenal included Logistic Regression, Decision Tree, Gradient Boosting, Random Forest, Bagging Classifier, and an Ensemble Classifier. The merit and efficacy of each model were gauged against standard metrics such as balanced accuracy, precision, recall, and a specially tailored F1-score. The findings conclude that the Random Forest model was outstanding in predicting injury outcomes from crashes, highlighting pivotal elements such as the vehicle's manufacture, road conditions, prior vehicle movements, lighting conditions, and intersection layout as major contributors to accident severity. These findings present a blueprint for traffic planners and AV manufacturers to devise strategies or policies geared towards curtailing the severity of AV-related mishaps.