Tracking partially occluded objects by region-wise linear subspace representations

• Main Authors: Ying-Yu Su, 蘇穎俞
• Other Authors: Cheng-Chin Chiang
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/v79b88

碩士 === 國立東華大學 === 資訊工程學系 === 95 === Moving object tracking is a critical issue in researches of computer vision technologies. This thesis proposes a vision-based tracking algorithm. This algorithm builds an appearance model of the tracked objects based on a linear subspace transform which reduces the dimension of representations for lowering the computation costs. The built appearance model and the subspace are continuously updated for adapting to the dynamic appearance change of the object. This simultaneous updates of both appearance model and the subspace distinguish the proposed tracking algorithm from others. In addition, we also invent a region-wise linear subspace representation to partition the objects into several regions, with each region represented by a linear subspace. During the tracking, the algorithm detects those occluded regions and uses non-occluded regions for reliable tracking. The experimental results do demonstrates that the proposed algorithm can effectively handle various object appearance changes such as partial occlusion, illumination change, zoom in/out, pose change, and local deformation.