Human abnormal behavior recognition based on two-stream convolutional neural network

Abstract

With the gradual improvement of people’s security awareness, the need for intelligent video surveillance is becoming more and more urgent. As a hot spot in the field of computer vision, the recognition of abnormal human behavior has received widespread attention from the whole society. This article proposes a human action recognition algorithm by learning joints information using two-stream convolutional networks. The spatial convolutional neural network is used to describe the spatial position and appearance information of the human body in the video frame. The temporal convolutional neural network constructs a motion feature vector based on the changes in the position of the human skeleton joint points between frames to describe the human body's movement. Finally, the spatio-temporal two-stream is fused to recognize the abnormal behavior.