Research On Tracking Of Multiple Objects With Occlusion Handling In Video Surveillance

Intelligent video surveillance applies computer vision and image processing techniques to automatically analyze image sequences that are captured by surveillance cameras. The goal of intelligent video surveillance is to localize, recognize and track those moving targets in the scene, and further analyze and estimate the target behavior to complete the function of daily management and sounding alarms when abnormal events occur.Firstly, we study the single moving target tracking technology that uses online continuous tracking and sparse detection tracking framework to balance accuracy and computational cost, and design and implement a single moving target tracking system. Secondly, we study the problem of multiple moving targets tracking. The main work is to detect and deal with different kinds of occlusion. Finally, we integrate the single target tracking module, the automatic pedestrian detection module, the occlusion estimate module and the occlusion handling module into a multiple target tracking system. The contributions and results are as follows:1. Because of traditional Mean-shift algorithm moving target tracking loss problems when the color of the background and moving target tracking region is similar, we propose an improved Mean-shift algorithm in this thesis. The algorithm uses the back-projection of hue histogram in the tracking region’s neighborhood to find the connected components which are of smaller probability in the tracking region and track, resulting in improved tracking accuracy.2. A method for moving target tracking combined the proposed Mean-shift algorithm with Kalman filter is developed. The algorithm can predict moving targets states while tracking, and regard the predicted position as the initial position of Mean-shift iteration, thus reducing the number of iterations.3. We design and implement single moving target tracking system. In order to increase the accuracy of tracking, we further integrate corner tracking and object tracking algorithm based on online multiple instance learning into our system.4. Relative to the single moving target tracking, multiple moving target tracking focused on studying the method of occlusion detection and occlusion handling. According to the degree of occlusion, we effectively combine the holistic target tracking and partial feature tracking methods to deal with occlusion. An algorithm to estimate occlusion boundaries is also proposed.5. A multiple target tracking system that effectively combines the single target tracking system, the pedestrian detection module, occlusion detection module and occlusion handling module.6. We did sufficient experiments to evaluate the performance of the improved algorithm, and the experimental results show that the proposed algorithm has better tracking accuracy and real-time performance when monitoring multiple moving targets.