Fall Behavior Recognition Based on Human Posture Estimation and Graph Convolutional Networks

A fall recognition method based on human posture estimation and graph convolutional networks is proposed to solve the problem of low efficiency and high false alarm rate in traditional video surveillance in fall detection. A diverse fall dataset covering 12 real scenes such as shopping malls, sports fields, and homes is constructed. Then, spatial domain enhancement (bilateral filtering, histogram equalization, and color enhancement) and frequency domain enhancement (Fourier transform and wavelet transform) are used to optimize video quality. After extracting human skeleton key points based on MMPose, the Multi-Stream Adaptive Graph Convolutional Networks (MS-AGCN) and Improved Spatio-Temporal Graph Convolutional Network++ (ST-GCN++) are used for recognition. Experiments show that wavelet transform significantly improves image quality, with an average PSNR of 40.41 dB and an average SSIM of 0.9741. Moreover, the recognition results show that the highest accuracy of wavelet transform in MS-AGCN is 64.23%, an increase of 5.69% over the baseline; the accuracy of color enhancement in ST-GCN++ reaches 62.60%, an increase of 8.94%. The improvement of recognition performance by video enhancement and spatiotemporal feature fusion is verified.