| Summary: | 碩士 === 華梵大學 === 資訊管理學系碩士班 === 96 === In this thesis, a system for tracking hands without markers in complex scenes has been developed. Our developed hand tracking system used the hand templates generated by the contours of hands. Currently, our hand tracker can track a hand with different scales (depth) and different orientations.
In generating the hand templates, 36 hand templates are generated for a hand pose (with rotation of 10 degrees for a hand template). According to the edge orientation, each hand template is divided into several channels. In our hand tracking system, each image has been processed using the following procedures: skin color detection, edge orientation and distance transform (DT). Given the DT maps, the hand can be tracked by using template matching.
If there are lots of hand poses, unfortunately, the number of hand templates for hand tracking is very large. To overcome the drawback of the large number of hand templates, in this thesis, the Bayesian filter and tree-based filter are used for hand tracking in order to reduce the computation time. The Bayesian filter is used to track a hand robustly and efficiently. The tree-based filter is used to reduce the computation time. Some experimental results are shown in the paper. In this thesis, the main reference is from the B. Stenger’s work [1]. In this thesis, some modifications of Stenger’s work are proposed: there are two parts of our modifications, such as coarse template matching and fine template matching. In the tree-based filter, for the coarse template matching, the better results for root nodes can be obtained by reducing the resolution of an image. For the fine template matching, the better results for nodes (except root nodes) can be obtained by using the characteristic properties of channels of hand templates. In this thesis, we focus on the Bayesian filter, tree-based filter and template matching.
|
|---|
