Incident detection using support vector machines

Abstract

This paper presents the applications of a recently developed pattern classifier called support vector machine (SVM) in incident detection. Support vector machine is constructed from a unique learning algorithm that extracts training vectors that lie closest to the class boundary, and makes use of them to construct a decision boundary that optimally separates the different classes of data. Two SVMs, each with a different non-linear kernel function, were trained and tested with simulated incident data from an arterial network. Test results have shown that SVM offers a lower misclassification rate, higher correct detection rate, lower false alarm rate and slightly faster detection time than the multi-layer feed forward neural network (MLF) and probabilistic neural network models in arterial incident detection. Three different SVMs have also been developed and tested with real I-880 Freeway data in California. The freeway SVMs have exhibited incident detection performance as good as the MLF, one of the most promising incident detection model developed to date. © 2003 Elsevier Ltd. All rights reserved.