A Remote Scene Binocular Video Surveillance System with Small Object Tracking

This paper proposes a remote scene surveillance system without ignoring minutiae of object. The utilization of two cameras enables us to track small object as well as providing possibility of content analysis and object recognition. In our system, the position relationship of two cameras is computed automatically based on the images from two views. The object to be tracked is in remote scene so that it lacks enough samples (pixels) to represent itself in image, so no existing tracking method could be utilized directly. Camshift algorithm is improved adequately for our application in three aspects: color space, generation of virtual samples and computation of weight image. Also, these improvements can be employed in general cases to strengthen robustness of tracking. Examples are presented in two parts. Part one demonstrates promising results from the improved tracking approach and part two focuses on the proposed system.