A Rapid Recognition Method for Pedestrian Abnormal Behavior

In recent years, with the development of deep learning technology, human behavior recognition has gradually become a very challenging and important topic in the field of computer vision. Human behavior recognition algorithms generally exploit temporal domain information such as optical flow and LSTM, or utilize 3D convolution, leading to large calculation and slow recognition speed, which is difficult to meet the requirements of speed and cost in practical applications. Therefore, this paper proposes a rapid recognition method that can quickly recognize the abnormal behavior of pedestrians. We extract the human skeleton information and remove the background to obtain a feature image containing only the human skeleton information, and then the Multi-scale information Fusion recognition Network (MFN) designed in this paper is used to recognize the abnormal behavior of pedestrians. We have achieved an accuracy rate of 96.3% in the self-built skeleton-based pedestrian abnormal behavior data set and the amount of calculation has dropped significantly.