A Machine Learning Approach for Detecting Traffic Incidents from Video Cameras

In the area of vehicular traffic analysis, many cities have surveillance devices (cameras) installed in different junctions as well as along roads, to obtain information on how vehicles behave in a certain area. To help in the process of preventing accidents, this article proposes a computer vision pipeline for detecting different types of dangerous/risky driving behavior. The pipeline includes object detection, tracking, speed normalization, and the application of computational intelligence, among other relevant features for pattern detection and traffic behavior analysis. The developed models are applied on videos from traffic cameras from eight different sites in the metropolitan area of the city of Montevideo. Three different algorithms were developed for detecting dangerous incidents in traffic in real time. Accurate results are reported for the case studies addressed in the experimental validation, reaching precision values up to 0.82 and recall values up to 0.91.