Kui Yang, Rongjie Yu, Xuesong Wang
Due to the rapid increase of traffic flow and frequent crash occurrence, traffic safety has become a severe issue for urban expressway. Real-time crash prediction is an important approach to identify traffic condition causing crash, which could further be used in the Active Traffic Management control strategies to reduce crash risk. Crash risk evaluation and threshold selection (short for crash risk prediction threshold selection) are two essential steps of real-time crash prediction: Crash risk evaluation could output the posterior probability of the crash of interest, based on real-time traffic data; threshold selection could provide the cut-off point for posterior probability, which was used for separating potential crash warnings versus normal traffic conditions. However, previous studies mostly focused on crash risk evaluation, and few on threshold selection. For the purpose of finding a threshold selection method, six methods (Ordinary method, Bimodal histogram threshold method, P-tile method, Otsu’s method, Maximum entropy method, and Minimum cross entropy method) were proposed and their predictive performance was compared by several evaluation criterions, based on the results of crash risk evaluation. Traffic data and historical crash data were used, and randomly divided into training data and testing data: the training data were used to develop crash risk evaluation model and further select the thresholds by six methods; the testing data were used to calculate crash risk and further compare predictive performance of threshold selection methods. The fact that Otsu’s method could provide the most promising threshold for real-time crash prediction was proved.