Research And Implementation Of Video 3D Human Pose Estimation Algorithm

Human pose estimation is a subtask of target detection in the field of computer vision.This task requires the extraction of local feature information and abstract structured information from image data,and detect the position of the human target in the image also with the coordinates of multiple keypoints of the human body.The human skeleton model connected by these keypoints can be used to describe the pose of the human body in the image.The problem studied in this subject is directed to the input of RGB video image sequences taken by a monocular camera,trying to predict the three-dimensional pose and motion trajectory of the target human body.This task not only has the complex background of the two-dimensional image and the occlusion of the human target,but also the problems of extracting timing information and the ill-posedness of the three-dimensional space.Aiming at this task,we designed and implemented a new multi-stage video 3D pose estimation framework.The framework processes the input video,and then performs human detection,human pose estimation,and finally three-dimensional pose estimation based on time series modeling in sequence.In each stage of the framework,this paper decomposes it into a sub-problem,and proposes a new algorithm based on the existing method,and finally realizes a pipelined 3D human pose estimation.In this paper,we also designs and implements a three-dimensional human pose estimation system based on the above algorithm framework.The system receives input of multiple video stream data formats,preprocesses it through frame compensation,target detection,etc.,and then performs 3D pose estimation,finally performs post-processing on the result,output the rendered 3D human motion animation.The algorithm proposed in this paper has an improvement of more than 5%compared with the baseline algorithm on some public data sets.