Research And Application Of Multi-person Pose Estimation Algorithm Based On Deep Learning

With the rapid development of modern science and technology,computer vision’s automatic analysis and processing capabilities of images and videos reflect increasingly important application value in the field of artificial intelligence.With the continuous improvement of computer hardware and the development of deep learning algorithms,the accuracy and speed of deep learning-based human pose estimation have been significantly improved,and have been widely used in the field of humancomputer interaction,motion analysis,augmented reality and virtual reality,etc.However,in complex scenarios with multiple people and severe occlusions,the existing 2D and 3D multi-person pose estimation methods often adopt a two-stage paradigm,that is,bottom-up or top-down method.This paper aims to propose a better grouping method for the 2D multiperson pose estimation of the bottom-up and a new single-stage paradigm for 3D multi-person absolute pose estimation to achieve a more accurate and concise model of human pose estimation.In addition,this paper will also use the human pose estimation model based on the edge device to achieve specific application.The main work of this paper is as follows:1.A 2D multi-person pose estimation algorithm based on center point for grouping is proposed.Based on the bottom-up paradigm,the algorithm proposes to use the center point as a grouping clue to predict the position of the keypoints and the offset to the center point,and then match the keypoints with the aligned results.In order to solve the variance problem of centripetal offset,a Multi-scale Transform Layer is designed.In addition,the algorithm proposes a greedy grouping strategy with a dynamic threshold,which groups the keypoints under the adaptive threshold.The algorithm finally shows excellent performance for medium-size instances,and achieves excellent performance in bottom-up methods.2.A single-stage 3D multi-person absolute pose estimation algorithm is proposed.The algorithm proposes an effective single-stage decoupling regression model.The algorithm uses parallel branches to simultaneously regression 2D pose and human depth,thus achieving a more compact pipeline.Secondly,the algorithm proposes a 2D Pose-guided Depth Query module,which links the features from the 2D pose regression branch and the features from the depth regression branch to help the model obtain better performance in 3D pose.In addition,the algorithm also proposes a Decoupled Absolute Pose Loss,which maps the predicted depth into the camera coordinate system using the ground truth of the 2D position of the instance,so as to realize the direct monitoring of the pose in the 3D space and improve the depth prediction performance.The algorithm finally achieved great performance among 3D multi-person absolute pose estimation methods.3.A motion analysis system based on edge devices is designed,and a post-processing algorithm of motion scoring based on DTW is proposed as the system data processing layer.In order to solve the problem of slow inference speed of human pose estimation model on edge devices,TensorRT is used to speed up,and a smaller backbone is used to improve the inference speed of the model.The whole system design also includes the terminal APP to provide user interaction,and finally realizes the practical motion analysis system with high intelligence,high precision and low delay,which can be used for the scenes of teenagers’ physical testing and fitness.