Research On Multi Person Human Pose Estimation Algorithm Based On Deep Learning

Multi-human posture estimation is an important task in computer vision and machine learning,which aims to locate and assemble key points in human bones for all human examples that appear in the image.Most research on the development of human posture estimation task model focuses on improving the accuracy of the task,which also increases the complexity of the model.In order to achieve the relative balance of computation,parameter number and accuracy,the network needs to be lightweight.In order to enhance the accuracy of the network and improve the detection capability,it is necessary to carry out a deep research on high-resolution information extraction and high-precision key point acquisition.Details are as follows:In this paper,the problem of large computation volume and large number of parameters in multiperson attitude estimation network model is discussed.In this paper,the light residual module is introduced based on HRNet backbone network.Depth separable convolution is used to reduce the computation amount and parameter number of the network model,and the global intentional force module is introduced to improve the accuracy of the network,so that the network model has a good balance in operation speed,parameter number and accuracy.Training and testing on the COCO dataset has shown that using lightweight residuals modules can speed up the network,reduce the size of the model,and maintain good accuracy while reducing storage size by 60%.In order to solve the problem of insufficient access to high resolution information in multiperson attitude estimation networks,a method based on subpixel convolution is proposed to obtain high resolution heat maps,which can effectively reduce the influence of artificial factors on networks in traditional models.In view of the fact that the heat map distribution of key points is multi-peak Gaussian distribution,a multi-Gaussian fitting method is proposed to optimize the heat map distribution so that the information of heat map distribution can be used more effectively.Finally,some quantization errors can be reduced by using Taylor expansion decoding method.Experiments and analysis under the COCO dataset show that this method is effective in improving the accuracy of detection.