Human body posture recognition algorithm for still images

Abstract

Aiming at the low accuracy and poor robustness of the current algorithm based on manual features, this study proposed a posture recognition method combining joint point information with convolutional neural network. The deformable convolution is used in the proposed method to improve the stacked hourglass model, so that it can extract the position of the human joint point accurately. At the same time, the convolutional neural network structure is designed to analyse the position information and confidence of the joint point autonomously, and extract the intrinsic link of the joint point of the human body. Finally, the softmax classifier is used to determine the pose category. Experimental verification has been carried out on the Willow data set. Moreover, the recognition accuracy demonstrates the effectiveness and superiority of the improved method.