Human Fall Detection Model with Lightweight Network and Tracking in Video

Abstract

In order to real time and accurately detect the action of human falling, combined with lightweight detection network, Kalman filter tracking, posture estimation network and spatiotemporal graph convolutional network, a joint algorithm for human fall detection in video is proposed. Firstly, the lightweight YOLOv3-Tiny algorithm is used to locate the frame of human in video, which can quickly detect the human-frame; among them, for the situation that the human body is likely to be missed in video, the Kalman filter tracking algorithm is integrated into the stage of target-detection and the accuracy of detecting is improved. Secondly, the human-frame detected or tracked in video is sent to the AlphaPose network to estimate the posture graph about human body. Finally, the spatiotemporal graph convolutional network is exploited to extract the spatiotemporal features of the human body, and eventually the result for classification is output. Experimental results show that the algorithm proposed in this paper, which is more appealing and successful than the other algorithm.