Multi-discriminant feature fall detection algorithm based on joints

Traditional fall detection algorithm is difficult to use accurately extract and recognise human posture features, and easily loses feature joints in the process of falling, resulting in low detection accuracy. Therefore, this paper proposes a multi-discriminant feature fall detection algorithm based on joints for nursing homes, medical rehabilitation centres and other places. Firstly, the initial features of human posture are obtained by the improved VGG-19 feature extraction model, and the initial positions of the joints are obtained and coded by adding a residual module. Secondly, the decoder network is used to complete deconvolution and upsampling operations to achieve greater fine-grained resolution. Finally, the image pose refinement module is designed to analyse the relationship between different adjacent feature nodes, so as to realise the accurate identification of the node position when the fall occurs. On this basis, the corresponding fall discriminant characteristics are proposed to achieve the detection of the elderly fall action. The results show that the proposed algorithm is more accurate and effective than other traditional algorithms on some datasets.