Classification of Crash and Near-Crash Events from Dashcam Videos and Telematics

The identification of dangerous events from sensor data is a fundamental sub-task in domains such as autonomous vehicles and intelligent transportation systems. In this work, we tackle the problem of classifying crash and near-crash events from dashcam videos and telematics data. We propose a method that uses a combination of state-of-the-art approaches in computer vision and machine learning. We use an object detector based on convolutional neural networks to extract semantic information about the road scene, and generate video and telematics features that are fed to a random forest classifier. Computational experiments on the SHRP2 dataset show that our approach reaches more than 0.87 of accuracy on the binary problem of distinguishing dangerous from safe events, and 0.85 on the 3-class problem of discriminating between crash, near-crash, and safe events.