Targets Association across Multiple Cameras by Learning Transfer Models

다국어 초록 (Multilingual Abstract)

In this paper, we propose a novel method to solve the problem of targets association and tracking across multiple cameras of non-overlapping views. The method is divided into two parts. One is an improvement on appearance transfer model, another is an improvement on spatio-temporal transfer model. To learn inter-camera appearance transfer models, lαβ color space is used to calibrate images. By this way, the overall and local information can be used, which has advantage to color transform correction. To learn spatio-temporal transfer model, entry/exit zones of a non-overlapping topology can be effectively estimated by defining valid link and using clustering method. Then a kind of time constrain is set between two nodes to judge whether there is correlation of observations. Experiments show the effectiveness of the proposed method.

목차 (Table of Contents)

• Abstract
• 1. Introduction
• 2. Improved on Appearance Transfer Model
• 3. Improved on Spatio-temporal Transfer Model
• 4. Experimental Results
• 5. Conclusion
• References