Learning Traffic Incident Detectors

Automated traffic monitoring and traffic incident detection is becoming a reality thanks to expanding networks of traffic sensors placed on US highways. However, the current state-of-the-art incident detection systems employ algorithms that require significant manual tuning. In this work, we study machine learning traffic detection solutions that are intended to reduce the need for the time consuming setup. We show that combining a set of simple data traffic sensor data streams via classification methods is a promising way to obtain a detector with an acceptably low false-positive rate and high and fast recall. We build and test a number of SVM-detector solutions and their refinements and show that they can outperform the widely used baseline, the California 2 algorithm. The algorithms are tested on incident data obtained for a section of monitored highway in a metropolitan area.