Human climbing detection algorithm based on multi-level HRNET and LSTM network

A multi-level HRNet-based algorithm for height-climbing behavior detection is proposed to address the problem of low recognition accuracy of existing height-climbing behavior detection algorithms. Firstly, the multi-level HRNet network is used as the human pose estimation module to obtain the 2D joint point data of the human body. Then the LSTM-based classification network is constructed, and the initial 2D joint point information is fed into the classification network for height-climbing action recognition of human pose. Finally, the target keypoint similarity OKS is used to evaluate the accuracy of keypoint recognition. Experiments were conducted on the COCO2017 dataset, and the accuracy of the proposed multi-level HRNet network for pose estimation in this paper reached 87.6%. Experiments on the human pose climbing action on the own dataset show that the accuracy of human pose climbing action detection of the multi-level HRNet and LSMT network proposed in this paper reaches 96.4%.