Matching of Objects Moving Across Disjoint Cameras Eric D. Cheng and Massimo Piccardi IEEE International Conference on Image Processing Outline Introduction Moving Objects Image Pre-processing Major Color Spectrum Histogram Track Matching and IMCSHR Experimental Results and Analysis Conclusions 2 Introduction Computer vision-based tracking of moving objects can based on [1]: shape motion appearance Appearance features are the main cue to track a same person from separate camera views. 3 Introduction Illumination effects must be eliminated to make the appearance of same object comparable for color constancy algorithms. The intrinsic images [2] and camera transfer functions [3] are effective methods. A simple method: Incremental major color spectrum histogram representation (IMCSHR). 4 Moving Objects Image Pre-processing To reduce illumination effects for disjoint camera views, we need pre-process for images. 5 6 Major Color Spectrum Histogram In RGB color space, there are over 16.7 million colors. How to reduce the number of colors? A possible approach to limit the size of this space is that of dealing with the three color components separately. This dismisses their spatial co-occurrence in the object and inadequate for accurate representation and comparison. 7 Major Color Spectrum Histogram Like PCR [4], we can scale the number of colors without losing accuracy by major colors in representing an object. Colors within a given mutual distance threshold are dealt with as a single color. 8 Major Color Spectrum Histogram 9 Track Matching and IMCSHR How to match two moving objects? similar or not? 10 Track Matching and IMCSHR object AMCSHR(A) object B MCSHR(B) 11 Track Matching and IMCSHR The similarity between two objects: 12 color distance ? For each major color of object A, choose the color distance less than for all major colors of object B. Track Matching and IMCSHR The color of B closest to A is defined as: For each major color of A, we could find the closest color of B to it. 13 Track Matching and IMCSHR The portion of in object A can calculated as: Similarly, the portion of in object B is: 14 Track Matching and IMCSHR The similarity of color in object A with its corresponding color in object B is defined as: The similarity of the whole objects A and B in the direction from A to B is defined as: 15 Track Matching and IMCSHR In the same way, we could find: Because its an asymmetric similarity measurement, we define: 16 Track Matching and IMCSHR The similarity between object A and B: if, a discrimination threshold, else 17 The bigger the difference between maximum and minimum similarity, the less similar are considered the two objects. Track Matching and IMCSHR Matching is assessed if Similarity(A, B) is above the chosen similarity threshold. Using multi-frame and IMCSHR to match better. 18 Experimental Results and Analysis The Matching of the same Moving Person in disjoint Camera Views 19 Experimental Results and Analysis 20 Matching Result: Experimental Results and Analysis The Matching of Two Different People from Two Disjoint Camera Views 21 Experimental Results and Analysis Matching Result: 22 Conclusions The overall object matching procedure can provide video surveillance applications with the ability of tracking single objects across disjoint camera views. 23 References [1] T. H. Chang and S. Gong, "Tracking Multiple People with a Multi-Camera System, Proceedings of the 2001 IEEE Workshop on Multi-Object Tracking, 19-26, [2] Y. Weiss, "Deriving intrinsic images from image sequences," Proc. of Int. Conf. on Computer Vision, [3] 0. Javed, K. Shafique, M. Shah, "Appearance modeling for tracking in multiple non- overlapping cameras," in Proc. of IEEE CVPR [4] Liyuan Li, Weimin Huang, I. Y. H. Gu, K. Leman, Qi Tian, "Principal Color Representation for Tracking Persons," in Proc. of IEEE Int. Conf. on Systems, Man and Cybernetics 2003, vol. 1, pp