Human Tracking in a video sequence is an important task in civilian surveillance. Successful human tracking provides data for security-purposes and personal monitoring systems. The trajectories of humans are one of the indications of potential criminal activity....[ Read more ]

Human Tracking in a video sequence is an important task in civilian surveillance. Successful human tracking provides data for security-purposes and personal monitoring systems. The trajectories of humans are one of the indications of potential criminal activity.

However, human tracking in video sequences is always a challenging problem. Due to rapid changes in shape with irregular motion, typical methods may not have satisfactory results, especially under frequent occlusion and interaction. Occupancy which makes use of the foreground pixels is one of the possible solutions. With a single camera, the lack of 3D information on humans makes the problem more challenging under frequent occlusion and interaction. Recently, methods based on multiple cameras have been proposed. These make use of views shot from different locations. The tracking is based on a 2D occupancy map built from the different views. Occlusion can be handled well, but it requires a high computation cost and suffers from the need for synchronization.

In this thesis, an approach with stereo cameras is proposed. The setting required is easier to implement than the one in the multi view approach, and the run time required is much shorter than in the multi view approach, which makes the stereo cameras approach suitable for real time tracking. A similar 2D occupancy map to the one in the multi view approach can be built. Without views from different locations, an occlusion and an interaction model are required to obtain a similar performance to the multi view approach.

The proposed algorithm combines the occlusion and interaction model and Markov Chain Monte Carlo (MCMC) such that humans can be tracked under frequent interaction and occlusion. The number of failures has been reduced successfully by 78%. This thesis presents an efficient and effective algorithm under frequent interaction and occlusion.