Research On Video-based Human Body Motion Tracking

Video-based motion capture system uses cheap equipments like digital cameras and personal computer to track a human's motion without any sensors or markers attached to the body. This topic has a wide spectrum of promising applications in areas such as smart surveillance, human computer interaction, athletic performance analysis etc., and becomes a hot topic of computer vision in recent years. Because of the complexity of the problem and lack of comprehension of human's vision system essence, visual tracking is still hard in computer vision.The background of this research is the project of"Science and Technology for the Olympic Games". Athletic motion analysis requires a motion measure and simulation system without any markers or obstructive. According to this requirement, we use techniques in computer graphics, image processing, computer vision to explore new methods of tracking and simulating human's 3D motion from multi-camera videos. The main work and contribution of this thesis is as following:(1) Propose a human body tracking framework based non-linear optimization. Three cues including gray value, edge and silhouette are combined in this tracking framework to construct the tracking object function. By defining the body model, projection process and the similarity function, we define our tracking object function as a form of sum of squared residuals. We use Gaussian-Newton algorithm to optimize this object function.(2) Design a multi-camera body tracking environment. It includes a parameterized skeleton model, a body shape model and a practicably multi-camera calibration algorithm. Based on these models, we implemented an experimental platform for 3D human body tracking in multi-camera environments.(3) Propose a method based on shape matching to solve the occlusion and error accumulation which are two difficult problems in the human body tracking. We also propose to use priors such as"self-intersection limitation","skin color region constrain","symmetry constrain"to improve the algorithm's adaptability to occlusion ,noise and change of environments.(4) Implement a primitive weight lifting tracking system. This system can track the athlete's motion and the barbell's motion parameter automatically.(5) Design and implement a real-time upper-body tracking system. This system can initialize the tracking and recover from failure automatically, and is robust to illumination change and body self-occlusion. In this system, several techniques, including face-detection, skin color segmentation, and probabilistic tracking are used to track a person's upper body's...