Research Of Motion Capture Key Technology Based On Monocular Video

Video-based human motion analysis is one of the important subjects in computer vision research areas, and has been also a front direction which is received much concern during the past few years. This thesis focuses on human motion detection, 2D motion tracking and human 3D reconstruction.Firstly, Single Gaussian and Mixture Gaussian Background Model are studied according to the feature of each pixel in the background fits Gaussian distribution, and a comparison is made between them, and analyzes their advantages and disadvantages. A new shadow detecting algorithm is put forward, and it can remove the shadow and keep the integrity of human field.Secondly, skeletonization algorithm based on Mathematical Morphology and Thinning are studied. On the basis of analyzing their heir advantages and disadvantages, this thesis adopts the second. A burr removing algorithm is proposed. The results show that the algorithm can remove burr and keep the integrity and connectivity of the skeleton. Because joint points of neck and shoulder can't be determined only depending on human skeleton, this thesis presents a torso quadrangle fitting algorithm, and accurately calibrating the joint points of human by combination of skeleton with quadrangle fitting.Lastly, because the depth of human itself is small compared with the distance between human and vidicon, a brief vidicon model is presented. This thesis adopt a approximate perspective model——scaled orthographic projection model, and having reconstructed the 3D data of human based on the proportion relation of human skeleton, manual work was only made at the first frame, and the post treatment can be conducted automatically.The results show that the algorithm of this thesis can resolve the key problem of motion capture, and laying a theoretical foundation for further study.