Incident detection algorithm based on partial least squares regression

摘要

We present the development of freeway incident detection models based on the partial least squares regression (PLSR), which has become a standard tool for modeling relations between multivariate measurements with flexibility, simplicity and strength. The PLSR models are built with the components extracted from the training dataset, and it distinguish incidents state from normal traffic state according to the output whether exceeding the threshold predefined. The performance of detection is evaluated using the common criteria of detection rate, false alarm rate, mean time to detection. Moreover, classification rate (CR), receiver operating characteristic (ROC) analysis and the area under the ROC (AUC) are also used to evaluate the model performance. Several experiments are performed to investigate the potential application of partial least squares regression to automatic incident detection. Simulated traffic data of Ayer Rajah Expressway (AYE) in Singapore and a real data collected at the 1-880 Freeway in California were used in these experiments. The available traffic measurements, including speed, volume and occupancy collected at both upstream and downstream, are used to develop the PLSR model. The experiments conducted on the simulated traffic data studied the influence that the proportion of incident instances in training set and different length of time series of measured data have on the detection performance. In addition, empirical results are presented comparing with neural networks for freeway incident detection. The experiments conducted on the real traffic data discussed the problem resulted from imbalance data (incident instance is rare class in real world), and compares its detection performance with support vector machine (SVM). The experimental results have demonstrated that the PLSR model is comparative to a MLF neural networks and SVM implementation for AID applications, and PLSR has the potential for the application of automatic incident detection in the real world.