Tracking people in video sequences

Abstract

Tracking people in video sequences has found many applications such as video surveillance, computer animation, sports science, etc. In this project, multiple cameras are used to capture the human body. Thus, shortcomings of using only one camera such as Self-Occlusion and Human-To-Human Occlusion can be overcome. The camera self-calibration approach is applied by automatically tracking the laser spots captured in the video sequences. An efficiently computed background subtraction algorithm based on a computational color model separates the human subject from the background. Finally, the position of the human in the scene is determined by tracking in the synchronized video sequences. Two identical digital video cameras are used to capture one person in the video sequences. For the calibration result, the calibration coverage problem of traditional technique using pattern (e.g. co-planar) has been solved. In the image segmentation stage, the computed background subtraction algorithm yields good outcome and due to the use of the computational color model, side effects caused by shadows are eliminated as well. The result of the project shows that tracking people using ordinary digital video cameras is possible. However, since the detail intrinsic parameter of cameras used is not accessible, the accuracy of calibration result need to be further enhanced.