A Hierarchical Network-Based Method for Predicting Driver Traffic Violations

Through the cleaning and filtering of real data from driving studies, a traffic violation dataset was constructed in order to determine the relationship between driver factors and the occurrence of traffic violations. Driver factors were analyzed using the indicator significance method and a multidimensional indicator set for predicting driver traffic violations was created. On this basis, we propose a hierarchical network-based method for predicting driver traffic violations. First, time sequences data preprocessing is performed to the multidimensional input data and a framework is presented for analyzing traffic violation data using the convolutional neural network(CNN) and long short term memory network (LSTM). CNN is used to obtain the time sequences of factors related to traffic violations and LSTM is adopted to acquire the temporal characteristics of these sequences, thus completing the initial calibration of the relationship between driver factors and the occurrence of traffic violations. Finally, the types of driver traffic violations are predicted using an improved attention network that adds two plug-and-play modules, including spatio-temporal interaction and deep convolutional feature extraction, which accomplishes the self-learning recalibration function of the indicator weights for calculating the probability of the occurrence of a certain traffic violation in the future period. On the traffic violation dataset, the proposed method was evaluated, and it increases prediction accuracy when compared to non-hierarchical methods and existing joint methods, thus contributing to the synergy of smart connected vehicle systems.