FM2GI-CNN Network-Based Rear-End Traffic Conflict Prediction at Signalized Intersections

This study applied a novel approach to predict rear-end conflicts at signalized intersections at the signal cycle level based on the FM2GI-CNN network. Firstly, the random forest (RF) is used to select appropriate input variables. Secondly, a generalized image transformation technique named FM2GI is employed to convert data to images. Lastly, combined with 2D-convolution for image feature extraction, a CNN network is constructed for traffic conflict prediction. The architecture was applied to a 7-day data from one approach at a signalized intersection collected by a LiDAR sensor. Based on the indicator MTTC, the serious rear-end conflicts were selected by peak over threshold. Eight important features were chosen by an RF model, and the FM2GI-CNN model was developed. Seven types of data from 1-Cycle to 7-Cycle were constructed to explore the impact on frequency prediction. The results show that 4-Cycle data holds the best performance, and the model has low sensitivity to the image size.