Advancing Autonomous Vehicle Safety: Machine Learning to Predict Sensor-Related Accident Severity

Abstract

Autonomous vehicles (AVs) represent an exciting frontier in transportation, promising increased safety and efficiency on the roads. However, like any technological advancement, they are not immune to accidents. Understanding the severity of accidents involving AVs is crucial for enhancing their reliability and ensuring public trust in this transformative technology. To address this challenge, our study has employed cutting-edge natural language processing techniques combined with machine learning to predict the severity of accidents involving AVs. Our study has contributed significantly by creating a novel dataset derived from post-disengagement accident reports, covering the years 2019-2022. This dataset comprises detailed descriptions of accidents, sensor information, and other critical parameters. Moreover, we have introduced a novel approach called Multi-Distance Synthetic Technique (MDST) to balance the imbalanced nature of our dataset, which included only 334 samples due to the rarity of such accident data. Utilizing MDST for data balancing, we aimed to enhance the robustness of our analysis. Additionally, we employed Recursive Feature Selection (RFS) to extract a valuable feature set that was crucial in predicting accident severity. Leveraging this selected feature set, we trained an ensemble model, which remarkably outperformed expectations, achieving an impressive accuracy score of 0.92.

Keywords

• Autonomous vehicles
• feature selection
• machine learning
• accident severity prediction
• data balancing