An improved Convolutional Neural Network model for human posture detection
Xin Guo, Jian Liu, Yaoxu Lei · 2023
The classic OpenPose convolutional neural network model has excellent performance in the field of human posture detection. But it also has some shortcomings, such as poor accuracy and low accuracy of human posture detection partly. An improved OpenPose network model is proposed to solve these problems. First of all, the feature extraction network VGG -19 of OpenPose model is replaced by the residual network (ResN et) with residual learning structure to improve the training accuracy of the network. Secondly, at the same stage, the dual-branch parallel structure is changed to the single-branch serial structure to predict the part confidence maps (PCMs) and the part affinity fields (PAFs) of human joint points. Finally, the structure of the convolution kernel is optimized and the residual network is introduced. The accuracy of the algorithm is improved and the computational speed of the model is guaranteed as much as possible. Experimental verification on MPII data sets shows that the detection accuracy of the improved algorithm model in this paper can reach 79.5%, which is 3.9% higher than the original model. At the same time, it has a higher accuracy of detection compared with other human posture detection algorithms. It can be applied to human posture detection with a high pursuit of the accuracy of detection of the model, such as the field of intelligent driving and rehabilitation training and so on.