Research For Human3D Pose Reconstruction And Action Analysis

In the field of computer vision, research on human motion analysis andthree-dimensional reconstruction are always highly focused. Researchers tryto use image sequences to complete detection, tracking, identification andinterpretation of human behavior. So far, three-dimensional reconstructionbased on human visual analysis and action can not only be used for visualmonitoring, user interface sensing and image retrieval, can also be used forvideo conferencing, virtual reality and motion image analysis, etc. Because ofits wide range of applications and great potential value, it still attracts wideattention from scholars.Focusing on this area, this paper aims to build a comprehensive systemfor3-D reconstruction and human motion analysis based on the idea that thesolution schedules for different systems are various.First we study technical characteristics of motion analysis based onmulti-camera theory. In experiments, we use more popular Kinect depthcamera as image acquisition devices rather than multiple cameras to captureimages from different angles. Kinect makes efforts to simplify imageacquisition, and also help us gain rich human motion data including depthinformation. Then, we conduct a3-D human body model according to thedepth images, which is a human skeleton model with16joints. In themodeling process, we propose a different approach from the past one to mapthe point cloud data to the skeleton model based on the traditional quantumevolutionary algorithm. But we break the limitation of a quantum bit binaryexpansion. The experiment results show a substantial improvement to theprevious ones. In addition, to use action classification, SVM, which is less costly, we propose a new action feature used to describe the classification ofhuman motion. These features not only contain location information andangle information of a skeleton model, but also contain a wealth of motioninformation.Secondly, we have investigated the reconstruction of3-D human bodyand movement analysis under monocular camera condition. Taking intoaccount the requirements in the practical application, we focus on addressinghuman upper body movement during tracking and reconstruction. To achievethis, we search joints firstly and then project them from two-dimensional tothree-dimensional model, and get a rough three-dimensional model of thehuman body. And then we use the simulated annealing method to optimizethe model. In this session we introduce creative9likelihood functions, toimprove the optimization results.