Towards a Learning Traffic Incident Detection System

The state of the art in traffic incident detection is dominated by approaches that require significant manual tuning. Our hypothesis is that these time-consuming solutions can be sucessfuly eliminated with the help of machine learning methods and past traffic data collected nowadays on major highways. We show that combining the output of a set of simple, imperfectly tuned, “offthe-shelf” detectors via classification methods is a promising way to obtain a detector with an acceptably low false-positive rate and high and fast recall. We evaluate the performance of a number of simple and combined detectors on real-traffic data and incidents recorded for a section of highway in the Pittsburgh metropolitan area. We show that a relatively simple support vector machine classifier solution outperforms the widely used baseline, the California 2 algorithm. Finally, we discuss the possibilities of improving detector performance by accounting for certain untimeliness of accident recording.