| dc.description.abstract | During the past decade, the technique of computer vision has been widely applied in several fields. Typical applications include virtual reality, intelligent surveillance system, human-interface, etc. There are two categories of human motion recognition approaches including model based and non-model based. Model based approach usually fits the given image or blob to a shape model, which represents joint parts and human body parts. One has to segment images into different parts, such as head, torso, arms, and legs. The drawback of this approach is that it needs more stable foreground segmentation. As to non-model based approach, it extracts features from the image, and the correspondence between consecutive frames is obtained based on estimation or prediction of features relating to shape, texture, and colors. The drawback of this kind of approach is that it is difficult to define the activity because of the lacking of pre-defined model.
In this thesis, the two approaches are combined. First, we use a pre-defied model, and features are extracted from different regions in this model. In this way, the complexity of features can be reduced due to the utilization of segmented images and the system can still perform well even if the foreground image is not stable. Human motions, like walking and crawling, usually transfer smoothly in each state. Hence, a transition diagram is designed to describe the transition between different motions. Experiments were conducted and results reveal the validity of our proposed approach. | en_US |