Driver distraction recognition with pose-aware two-stream convolutional neural network

Abstract

Distracted driving is the main factor causing traffic accidents. The traditional distracted driving behavior detection method not only ignores the influence of the driver's posture on the recognition accuracy, but also only extracts the spatial position information in the image, and cannot capture the driver's motion information related to the time feature, which leads to the detection accuracy and robustness are not satisfactory. This study aims to build a more comprehensive model to expand the application scenarios, recognition objects, and behavior types of driver behavior recognition systems. This work produced a video dataset of distracted driving behavior and built a two-stream convolutional neural network (CNN), including spatial and temporal streams. The spatial stream introduces a Pose-Image Distracted Driving Detection (PIDDD) method, The input image is a pose-image that contains the driver's attitude information, uses the Hierarchical Bilinear Pooling (HBP) model to extract the driver's static attitude information, and improves it to enhance the model's receptive field and feature extraction capabilities, The input data of the time flow is the optical flow feature map extracted from adjacent image frames, using ResNet50 to extract driver motion information. The model combines the two stream results are weighted and fused to output the classification results, and the experimental results show that the method can better identify the driver's head and arm movements, and the recognition accuracy reaches 95.49% on the GADX901 dataset, which has important application value in distracted driving behavior recognition.