Lightweight Multi Person Attitude Estimation Based On Symmetric Space Transformation

Human pose estimation task is one of the hotspots of computer vision task,which mainly includes the task of accurately locating the human body and the key points of human body in the image or video,and then estimating and recovering the human pose.In the fields of motion modeling,monitoring and security,behavior analysis and human-computer interaction,human posture estimation has a wide range of application value.In real life,due to the influence of crowd occlusion,light changes and various complex scenes,there are some common and typical problems and challenges in human pose estimation.Based on these problems and challenges,on the basis of learning and summarizing the knowledge methods of related fields,this thesis proposes the corresponding solutions by using deep learning neural network.(1)Aiming at the problem that the application effect of target detection results in attitude estimation network is not good,a multi person attitude estimation method based on symmetric space transformation is proposed.In this method,a spatial transformation module is added in the output phase of target detection to extract high-quality single person pose regions from inaccurate human candidate frames,so as to improve the performance of the target detection network,and then the single person pose estimation network can get highquality candidate frames,so that the network can get high-quality candidate frames Better human key point results.Using the above method,the problem of poor application effect of target detection algorithm in attitude estimation network can be solved.At the same time,the experimental results on mpii data set and MS COCO data set verify the effectiveness and competitiveness of the proposed algorithm in improving the quality of candidate area frame and improving the overall performance of the network.(2)In order to solve the problem of human detection in the crowded state,a multi person pose estimation method based on KM algorithm is proposed.After the network output of target detection and single person pose estimation,the method generates the human body key point graph by optimizing the loss function.Using the KM matching algorithm,the key point prediction problem is transformed into the optimal matching problem of solving the human body key point graph.Finally,the algorithm matching is used to solve the problems of wrong judgment and missing detection of key points in the crowded state.Finally,the experimental results show that the proposed method can improve the performance of the multi person attitude estimation network in the crowded state.(3)Aiming at the problem of complex network model and long operation time of multi person attitude estimation,a lightweight multi person attitude estimation method based on mobilenet is proposed.Based on the network structure of method one,this method uses deep separation convolution and inverse residual model to reduce the complexity of the network,so as to balance the accuracy of the algorithm and improve the speed of the algorithm.Finally,the effectiveness and competitiveness of the algorithm are verified by comparative experiments.