Research Of Human Pose Estimation Method On Deep Learning

Human pose estimation is an important part of computer vision and a key step of motion recognition.With the rapid development of deep learning,it is possible to achieve fast and accurate human pose estimation.The research on how to make the human posture estimation task better and apply it to the video based on the actual needs has become an important research content with practical significance.In this paper,the relative factors of the balance between processing speed and precision of human pose estimation algorithm are explored,and an effective method of human pose estimation is finally formed.At present,there are many kinds of human pose estimation methods,including single-person pose estimation method and multi-person pose estimation method.The research of single-person pose estimation method has been relatively mature,and the current research direction is multi-person posture estimation method.For the application of multi-person estimation in video,most methods only focus on accuracy and ignore the requirement of processing speed in practical application.In this paper,our main work is based on a variety of existing representative for both single-person and multi-person pose estimation methods,to explore and analyze the relevant indexes of these methods processing speed and accuracy by several groups of comparative experiments.Then same feasible improvement ideas are put forward to improved correlation method,and take the contrast experiment with the original method,thus it is concluded several factors which is influence of method's balance between processing speed and accuracy.Based on this,an improved Dense Net networks for human pose estimation is proposed.In order to solve the problem of the impact of speed due to the uncertain number of people in the video frame,and the relative size of different human bodies or body parts leads to the poor detection performance of the common keypoints detection method,an improved Dense Net network structure is proposed for human pose estimation.It is a single-stage and end-to-end network structure,which uses deep convolutional neural networks for feature extraction.At the end of the convolutional network,it can get 6 different scales of feature maps by using a specific scale transfer structure.Then the network can integrate different levels of features for multi-scale keypoints detection,which effectively improves the detection accuracy of keypoints.The bottom-up approach is adopted to ensure the processing speed of the multi-person pose estimation task in this paper.And experiments show that the proposed method is superior to other methods in the comparison of comprehensive performance.It provides a new method for balancing the speed and accuracy of attitude estimation.