A Driver Abnormal Behavior Detection Method Based on Improved YOLOv7 and OpenPose

The current driver's abnormal behavior detection is interfered with by complex backgrounds, body self-obscuring, and other factors leading to low accuracy of small object detection and large error of human node detection. In this paper, we propose a driver abnormal behavior detection method based on improved YOLOv7 and OpenPose. Firstly, we add an attention mechanism to the backbone network of YOLOv7, increase the feature information contained in the shallow feature map, introduce Wise-IoU loss function to improve the detection accuracy of small and long objects (e.g., cigarettes, water glasses). Secondly, we calculate the IoU values between the driver's hand and the confidence frame of small objects to detect the abnormal behavior of driver-object interaction (e.g., smoking, drinking). Thirdly, we improve the OpenPose network model by replacing convolutional kernels and adjusting the connection of convolutional kernels to achieve two-dimensional node detection of the driver's driving posture. Finally, we use the FastDTW algorithm, and calculate the similarity of the driver's two-dimensional nodal information to detect the abnormal posture of the driver. The experimental results show that the method in this paper has high accuracy of small object detection and low nodal error, and can effectively detect abnormal behaviors such as smoking, drinking and abnormal posture of drivers during driving.