模式识别国家重点试验室.ppt

Motivations: Motivations: The field of view of a single camera is limited. Multiple cameras are used to achieve wide area surveillance. How to automatically analyze and fuse the information gathered from multiple cameras??? 模式识别国家重点实验室 中国科学院自动化研究所 National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences 模式识别国家重点实验室 中国科学院自动化研究所 National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences Matching Tracking Sequences Across Widely Separated Cameras Yinghao Cai Kaiqi Huang, Tieniu Tan Presented by Zhuoshi Wei National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences 10/13/2008 Outline Experimental Results Motivations Hierarchical appearance matching Problem Description: Images taken at Camera 1 Images taken at Camera 2 establish correspondences between observations across cameras Problem description: Objective: To establish correspondences between observations across cameras. Difficulties: Low resolution Illumination variations Pose variations Different camera parameters Segmentation errors and partial occlusions. Motivations Traditional methods: snapshot-based appearance matching: Match objects by a single image Sequence matching Camera 1: Camera 2: To alleviate the uncertainties due to segmentation errors and partial occlusions Hierarchical Appearance Matching Focus on sequence-based appearance matching to alleviate the uncertainties due to segmentation errors and partial occlusions. Training stage: Obtain the appearance model for the moving object of the whole sequence. Testing stage: Sequentially accumulating the posterior likelihood by Bayesian inference. Training stage: Partition the blob into regular patches for localization of color components. Dominant color representation Each patch is represented by its dominant colors similar to [Piccardi, CVPR05]. Colors within a distance threshold are regarded as a sing