Fully automatic vision-based system for vehicle crash prediction and recognition

Just as they were half a century ago, automobile accidents are, unfortunately, one of the leading causes of death today. Therefore, it is no surprise that automated traffic analysis systems generate a supreme amount of interest. Despite the rapid advances in technology today, many traffic monitoring systems require substantial amounts of careful annotation. As such, a fully automated traffic analysis system that can perform accident prediction would be highly beneficial to multiple parties. Such systems would make a traffic analyst's workload more manageable and would provide a more sophisticated tool for determining the root causes of traffic accidents. In this thesis, we present an automatic vision-based system for both accident prediction and recognition. Our method first detects and tracks vehicles using Robust Principal Component Analysis (Robust PCA) and Kalman Filters in order to extract trajectories. Pairs of vehicles trajectories are then segmented and classified by a Support Vector Machine (SVM) in order to determine the likelihood of a collision. We also tackle the problem of accident recognition by classifying crashing trajectory pairs into distinct categories. An ontology is used to define the relationships between the accident types and to train a tree-based classifier for recognition. We demonstrate the effectiveness of each algorithm by evaluating them on a crash dataset provided by the Toyota Motor Corp