Video-based Human Motion Analysis

Human motion analysis is an important problem in many research areas. It has vital applications in biomedical engineering, virtual reality, robot, security surveillance. The object researched in this thesis is video-based human motion analysis that track, capture and analysis human motion in video. Human motion is non-rigid object motion and is nonlinear. It's worth to research and has wide application.Although video-based human motion analysis is a challenging problem, many researchers present some method to solve this problem partially. In this thesis we introduce this problem in detail. And review the existing solutions by categorizing these solutions to feature-based method and model-based method. We research the human motion in two different type videos. In the first kind video, the object is the walking human who is dressing straitjacket and whose joints are attached by MLDs(Moving light displays). We track this kind motion using the feature-based method. We take correlation method to detect and extract feature and it is proved effective. In contrast to traditional tracking method which using single patter, we track different body features using different strategies combined with human body structure constraint. The experiments show that this method can track well in non-occlusion condition. In the second video, we track the walking human who addressing in general clothes using model-based method. We discuss the preprocessing of video and whole body tracking problem firstly. Then we build a 2~D lower limb model. After this, we track the lower limber motion using the method that finds best parameters that minimize the difference between synthesis data and real data. How to search parameters in parameter space is a problem. Traditional methods solve it using optimization techniques. Here we get the parameters by analysis respectively. This method reduces the space dimension and2makes computation more efficient. We also define an appropriate similarity function that makes the tracking result more accurate and robust in occlusion situation. The experiments on real video demonstrate that the methods proposed in this thesis can track lower limb motion of walking human motion very well and overcome occlusion in some extent.