AV Sensor Architectures for V2V Crash Reconstruction and Prediction

Abstract

The major challenge in Autonomous Vehicle (AV) crash reconstruction and prediction is the need for appropriate AV sensor data to accurately reconstruct a range of crash events and predict potential crashes across diverse driving scenarios. This thesis aims to develop methods for collecting such data for effective evidence-based crash reconstruction and prediction to address these limitations. This research introduces a novel Simulation Method for Performance Evaluation (SMTPE) to select a sensor architecture for collecting relevant AV sensor data. The crash data collected using the chosen architecture are utilised for crash reconstruction and, subsequently, for prediction with a proposed Vehicle Crash Reconstruction and Prediction model (VCRPM). By delving into pre-crash and crash scenarios through simulated Vehicle-to-Vehicle (V2V) crash events, this research offers a solution to time and resource constraints, pushing the boundaries of what is possible in the pursuit of safer, more intelligent AVs.

The findings indicate that AVs can effectively reconstruct crashes and predict potential accidents in real time by integrating data from Radio Detection and Ranging (RADAR), cameras, and light detection and ranging (LIDAR) sensors, utilising data fusion and machine learning techniques. This research presents a functional sensor architecture for AV crash studies, a new approach for crash reconstruction, and three ensemble-based machine learning models for real-time crash prediction.