Multi-camera non-overlapping visual object tracking system typically consists of two tasks: single camera object tracking and inter-camera object tracking. Since the state-of-the-art approaches are yet not perform perfectly in real scenes, the errors in single camera object tracking module would propagate into the module of inter-camera object tracking, resulting much lower overall performance. In order to address this problem, we develop an approach that jointly optimise the single camera object tracking and inter-camera object tracking in an equalised global graphical model. Such an approach has the advantage of guaranteeing a good overall tracking performance even when there are limited amount of false tracking in single camera object tracking. Besides, the similarity metrics used in our approach improve the compatibility of the metrics used in the two different tasks. Results show that our approach achieve the state-of-the-art results in multi-camera non-overlapping tracking datasets.

本文提出了一种全局图模型和改进相似度度量的方法来提高多摄像头对象跟踪的性能，并且成功地将单摄像头对象跟踪和跨摄像头对象追踪区别对待。实验结果表明，即使在单摄像头对象跟踪质量不佳的情况下，该方法也能更好地发挥作用。

一种基于全球均衡图模型的多摄像机物体跟踪方案