The prediction of red-light running (RLR) at a signalized intersection is a crucial component of dynamic all-red extension (DARE) that avoids potential collisions caused by RLR behaviors. Previous studies on RLR prediction are usually based on discrete measurements of fixed sensors, e.g. inductive loop detection (ILD). This paper formulates the RLR prediction as a binary classification problem that can be solved by the least square support vector machines (LS-SVM) models based on continuous trajectories detected by radar sensors. We adopt non-weighted and weighted LS-SVM models for the RLR prediction. Tuning parameters are conducted by the cross-validation in the discrete parameter space and the Bayesian inference based on the maximum a posteriori probability (MAP) estimate in the continuous parameter space. Based on continuous trajectories, this study selects the velocity, acceleration and distance to the stop line at the red light onset time as attributes plotted in a 3-D scatter diagram. The validation results show that the performance indicated by the receiver operation characteristics (ROC) curves of the LS-SVM RLR prediction performs much better than that of the ILD model. Compared with the non-weighted LS-SVM classifier, the weighted LS-SVM classifier for RLR prediction is more flexible. In order to further reduce the RLR missing rates of based on the previous LS-SVM solution, the weighted LS-SVM is developed with a tolerated small false alarm rate by allowing a variety of unbalanced ratios to identify the relative significance of each class.
↧