Comparison between Partial Least Squares Regression and Support Vector Machine for Freeway Incident Detection

This paper presents the development of automatic incident detection (AID) models based on the partial least squares regression (PLSR), and compare it with support vector machine classifier which has exhibited good performance for freeway incident detection. The performance of AID algorithms is evaluated using the common criteria of detection rate, false alarm rate, and mean time to detection. Moreover, the curve of receiver operating characteristic (ROC) is also used to compare the detection performance. Simulated traffic data and real data collected at the 1-880 freeway in California were used in these experiments. Traffic flow parameters, such as volume, speed, occupancy and time headway both at upstream and downstream, and derived data generated from basic traffic flow parameters are used to build the PLSR model and SVM models. Several experiments using the original data or derived data have been performed to make comparisons between PLSR and SVM. The problem resulted from imbalance data and its influence on detection performance is also discussed. The test results have demonstrated that the PLSR has great potential to detect incident.