Research on Pedestrian Crossing Intention Prediction Based on Deep Learning

Abstract

Using computer vision techniques to detect, identify and track pedestrians in video sequences has important significance. For the pedestrian detection task, YOLOX is used as the pedestrian detection algorithm, and Ghost convolution module, asymmetric convolution, CBAM module, and DIOU loss function are introduced to improve YOLOX; the improved YOLOX is used as the detector of ByteTrack, and after classifying the detecting frames with high and low confidence, the Hungarian algorithm is used to match the Kalman filter predicted trajectory and the detection frame are matched using the Hungarian algorithm to realize the function of pedestrian tracking. The improved YOLOX model improves the precision by 0.82%, the detection speed by 17%, the recall by 0.9%, the model F1 by 1%, and the AP value by 0.27%; the improved ByteTrack improves by 5% compared to the original version.