¹¦·ò: 2014Äê12ÔÂ18ÈÕ£¨ÖÜËÄ£©ÉÏÎç10:00

µØÖ·: У±¾²¿¶«ÇøÏèÓ¢´óÂ¥516ÊÒ

½²×ù: Human Pose Estimation and Action Recognition £¨ÈËÌåÐÐΪ¹À²â¼°×÷Ϊ¼ø±ð£©

Ñݽ²Õß: ÃÀ¹ú»ªÊ¢¶Ù´óѧ Jenq-Neng Hwang ½ÌÊÚ

Ñݽ²Õß¼ò½é£º

»ÆÕýÄܽÌÊÚ±ðÀëÓÚ1981ÄêºÍ1983Äê´Ó¹úÁ¢Ì¨Íå´óѧ»ñµÃµç×Ó¹¤³ÌרҵµÄѧʿºÍ˶ʿѧλ£¬Ö®ºó¸°ÃÀÁôѧ£¬ÔÚÄϼÓÀû¸£ÄáÑÇ´óѧ»ñµÃ²©Ê¿Ñ§Î»¡£±ÏÒµºóÓÚÃÀ¹ú»ªÊ¢¶Ù´óѧ£¨Î÷ÑÅͼ£©ÈνÌ£¬1999ÄêÌáÉýΪÕý½ÌÊÚ£¬Ä¿Ç°ÊÇ»ªÊ¢¶Ù´óѧ£¨Î÷ÑÅͼ£©µç×Ó¹¤³ÌϵƽÉú½ÌÊÚ£¬Ö÷¹Ü¿ÆÑеÄϵÖ÷ÈΡ£

»Æ½ÌÊÚ»ñµÃ¹ý¶àÏî¸÷ÀàÈ˲żν±ºÍѧÊõ¼Î½±£¬ÓµÓзá˶µÄ¹ú¼ÊºÏ×÷¾­Àú£¬Óëб¦GG·¢Õ¹Á˽ϳ¤¹¦·òµÄºÏ×÷Ó뻥»»£¬ÊÇб¦GGµÄº£±íÃûʦ¡£»Æ½ÌÊÚÔÚѧÊõÉÏÓµÓкÜÉîµÄÔìÒ裬ÊÇÓйØÁìÓòÓµÓнϸ߹ú¼ÊÓ°ÏìµÄ³ÛÃû¿Æѧ¼Ò¡£ÏȺóÔÚ¶àýÌåÐźŴ¦Ö㬶àýÌåϵͳ¼¯³ÉºÍÍøÂçÁìÓò°ä·¢ÁË300¶àƪÂÛÎÄ¡¢ÖøÊöµÈ¡£Í¬Ê±»¹µ£Èζà¼ÒѧÊõίԱ»áµÄÖ÷ϯ¼°¶à±¾³ÛÃûµç×Ó¹¤³ÌÀàÔÓÖ¾µÄÖ÷±à£¬´Ó2001ÄêÆð¼´³ÉΪIEEE»áÔ±¡£

½²×ùÌáÒª£º

This talk will mainly focus a human pose estimation and action recognition based on monocular video sequences. First, using shape, color and time continuity information, the 2D body parts are tracked in each frame, more specifically, the shape information is explored by a skeletonization process, the color information by the mean shift algorithm, and the time continuity information by Kalman filter. Then the downhill simplex algorithm is applied on the tracked 2D body parts to search the best 3D pose of the 3D body model. For each frame, the 3D coordinates of 13 joints are converted into 3D geometric relational features (GRFs). Finally, with the sequences of 3D GRFs, each action is trained as a cyclic hidden Markov model (CHMM), and the trained HMMs are used to classify different human actions.