Real-Time Crash Likelihood Prediction Using Temporal Attention–Based Deep Learning and Trajectory Fusion
Pei Li, Mohamed Abdel‐Aty
Abstract
A crucial component of the proactive traffic safety management system is the real-time crash likelihood prediction model, which takes real-time traffic data as input and predicts the crash likelihood for the next 5+ min. This study aims to investigate the application of trajectory fusion to crash likelihood prediction and improve the predictive accuracy of the deep learning crash likelihood prediction model using the temporal attention mechanism. Two trajectory data were integrated using data fusion techniques. Specifically, trajectory data from Lynx buses and the Lytx fleet were collected using the automatic vehicle locator (AVL) and Lytx DriveCam, respectively. A deep learning model was developed for predicting real-time crash likelihood using features extracted from trajectory data. The proposed model contained a temporal attention–based long short-term memory (TA-LSTM) and a convolutional neural network (CNN). Temporal attention was introduced to capture temporal correlations between time-series data. Experimental results suggested that temporal attention could significantly improve the model’s performance on crash likelihood prediction. The proposed model outperformed other benchmark models in terms of sensitivity and false alarm rate. Moreover, trajectory fusion improved the predictive accuracy of the proposed model, which indicated the importance of having data from different types of vehicles for developing real-time crash likelihood prediction models.