A Heterogeneous Network-Oriented Strategy to Highway Incident Estimation for People Driving

To explore the relationship between driver-related factors and the occurrence of traffic violations, a traffic violation dataset was constructed by cleaning and filtering real-world data from driving studies. The driver factors were analyzed using an indicator significance method, leading to the development of a comprehensive multidimensional indicator set for predicting traffic violations. Building upon this, a hierarchical network-based prediction model is proposed. The approach begins with preprocessing time-series data from the multidimensional inputs, followed by the use of a hybrid framework that incorporates both Convolutional Neural Networks (CNN) and Long Short-Term Memory networks (LSTM). CNNs are utilized to extract time-sequential features related to traffic violations, while LSTM networks capture the temporal dynamics of these sequences, allowing for an initial mapping between driver behavior and the likelihood of violations. To enhance prediction capabilities, an improved attention-based network is introduced, which integrates two plug-and-play modules: spatio-temporal interaction and deep convolutional feature extraction. These modules enable the network to autonomously adjust and learn the importance of various indicators, thereby calculating the probability of future traffic violations more effectively. The model was tested on the traffic violation dataset and demonstrated superior prediction accuracy compared to traditional non-hierarchical and joint methods. This approach supports the advancement of intelligent connected vehicle systems by improving traffic violation prediction through enhanced data analysis and modeling techniques.