A Real-time Motion Detection with Differential Images and Tracking with Mean-Shift and Kalman Filter

Abstract

The main techniques of computer vision that can be helpful to surveillance system are detection and tracking. This thesis proposes a real-time motion detection and tracking system based on a single camera as a cost-effective solution for reducing human labor on surveillance. The detection algorithm deals with the change in pixels between sequential frames. Arithmetic operations on these pixel values provide position information of motion. Tracking process is more complicated. In this project, the tracking system requires selection of ROI (region of interest) as preprocessor. Then, mean-shift algorithm examines the distinct pattern of ROI and track the pattern every frame. To prevent a failure of mean-shift tracking, the tracking system is equipped with mathematical tool, Kalman filter. Kalman filter estimates and predicts the desirable route of mean-shift tracking, using its position and velocity information. The filter corrects unacceptable deviations from the route and helps a tracking window keep functional. This project separately developed detection algorithm and tracking algorithm and combined them at the final stage. The redundant imaging techniques are excluded in the proposed system in order to minimize the computation time, which ultimately shorten the delay for a real-time implementation. This system will promote low delay but high performance real-time surveillance system.