Recovering Human Body Configurations And 3D Model Simulation

Human Motion capture now is widely used in the fields of simulation technology, robot control, medical analysis, sports analysis, and some other applications. However, human motion capture simulation technology and computer image is one of the most difficult issues addressed. This is a very difficult problem, partly because human bodies are versatile, presenting a wide range of pose and aspects, many including self-occlusion, and partly because variations in clothing and background clutter deny a simple appearance model.Given the seemingly insurmountable difficulties, many existing approaches to this problem make simplifications of one sort or another, either assuming knowledge of scale and appearance/color, or using motion information from video sequences for background subtraction, or limiting evaluation to restricted domains such as walking figures. In these cases, a canonical tree-based model is typically used to model body parts, where dynamic programming can be applied.We tackle the problem in a more general setting. We use the video sequences from 2007 East Asian Games women's gymnastics. Without restrictions in pose, appearance, or background clutter, a tree-based model no longer suffices. Additional sources of information, not provided by tree based models, are required to succeed. For example, the symmetry of clothing is a powerful cue to constrain limb appearance. What reveals the body position to us are the connection between the two upper legs and the relative geometric relationship between arms and legs, both of which are not in the traditional tree-based model.It is an open question what models can express sufficient constraints and are computationally feasible. In this work, we develop a strategy that exploits a rich set of cues, defined on arbitrary pairs of parts, to constrain body configurations. We learn these constraints from empirical data and use Binary Quadratic Programming to find the most probable configurations. Our program is a well-studied computational framework, where efficient approximations exist. Many cues for estimating human body configuration can be expressed as pair-wise constraints. In our experiments we have found that this program works well for this problem.We have tested our algorithm on a variety of images. With recovered body configurations and the associated segmentation masks. As compared to a brute-force search approach, we are able to handle a much larger set of candidate parts and do not rely on the availability of a few being very salient. We have found that a two-step strategy using the linear approximation of our works well for our assignment problem, produces satisfactory results on a variety of images without relying on extensive low-level processing, and is computationally efficient. We believe that the program formulation will find more and more use in detecting articulated objects.