Human Action Recognition Using Weighted 3-Viewpoints Motion History Histogram

• Main Authors: Mei-HsuanChao, 趙美琁
• Other Authors: Chu-Sing Yang
• Format: Others
• Language: zh-TW
• Published: 2012
• Online Access: http://ndltd.ncl.edu.tw/handle/87315699583018142312

碩士 === 國立成功大學 === 電腦與通信工程研究所 === 100 === A human action recognition system based on depth image is proposed in this paper. Extracting foreground human object by depth data is more robust with environment affect. Besides of the motion with main direction parallel to the camera, depth motion trajectory can also be clearly presented by projecting depth data to three-dimensional volume object. First, the system can solve the self-occlusion and different speed problem in Motion History Image (MHI) method by detecting the start, continued and end time of a simple motion automatically. Then, project three-dimensional volume object to three orthogonal planes. Three-dimensional motion history trajectory can be described by different viewpoint of MHIs. The three-viewpoint MHIs are weighted to increase importance to the planes with greatest detail. For the efficiency purpose, Motion History Histogram (MHH) is extracted as motion feature. From the proposed method, any actions with different action speed and different main directions in 3D space can be efficiently recognized. The experimental results demonstrate the accuracy and effectiveness of proposed weighted three-viewpoint Motion History Histogram in different situations.