Research And Application Of Human Pose Estimation Based On Deep Learning

Human pose estimation is an important problem that has attracted much attention in the field of computer vision in recent decades.Its purpose is to locate the keypoints of human body in images and connect them to output the human pose skeleton.It is a key step in understanding the behavior of people in images and videos.Human pose estimation has important applications in security,automatic driving,sports and rehabilitation training and other fields.In practical applications,due to problems such as large data scale,limited network bandwidth and privacy security risks,human pose estimation often needs to run on edge devices instead of transmitting data to the cloud for calculation.However,the memory and computing resources of edge devices are limited and cannot support complex and huge network models.Therefore,the lightweight research of network models is of great significance.In this paper,Lightweight Cascaded Pyramid Model(LCPN)is proposed for the first time.The main work can be summarized in the following two parts:1.The lightweight design of the model network was carried out from the convolution structure level.The traditional convolution method was replaced by depth wise separable convolution.The comparative experiments prove that the model proposed in this paper reduces the amount of network weight parameters and the amount of calculation under the premise of achieving the same operation effect,which means that the complexity of the model was reduced.This work greatly improves the efficiency of the model.2.The method of predicting keypoints in the human pose estimation is improved by a new method.The heatmap prediction and keypoint coordinates directly regression parts both work during training.The heatmap prediction part is used to supervise the coordinates directly regression part in order to improve the accuracy of the model,while only the latter part is kept during testing.The experiment results prove that the network is simple,lightweight and fast,which can effectively improve the prediction efficiency during testing.This paper designs and implements a fitness action scoring applet based on actual application requirements.The applet runs the LCPN model on the mobile terminal,which can estimate and display the pose of user's fitness actions in real time,and it will give users their score of fitness actions by comparing with the standard posture.The running result of the applet verifies the correctness and effectiveness of the research and design work in this paper.