Research On Moving Multi-Object Detection System Based On Image

Multiple target tracking is one of the hot issues in computer vision and has been widely used in the areas of video surveillance, video compression, and imaging guidance. So far, many target tracking algorithms have been proposed in the field of computer vision. Mixture Gaussian background method could detect moving foreground from the multi-modal scene accurately. Mean-Shift tracking algorithm is fast and shows good real-time performance. Particle filter tracking algorithm is able to deal with the nonlinear non-Gaussian system effectively and shows good accuracy performance. These algorithms are widely used in target detection and tracking. But so far, no one method could be applied to the entire tracking scene. The main task of this dissertation is to make up for disadvantage of these algorithms, and then integrate these optimized algorithms to design a new detection and tracking system.This dissertation introduces basic principles of frame difference, optical flow and background subtraction. Mixture Gaussian background method is not sensitive to the shadow. For this defect, this uses the difference between the target and shadow in HSV space to find and weaken shadows. This background method is difficult to overcome the light mutation. This calculates brightness changes of all the pixels to detect and remove serious light noise of the frame to solve this problem.This dissertation then analyzes the Mean-Shift Algorithm theories and Tracking.This algorithm is easy to fail when the target is blocked with each other or the target shape or gesture change. For this defect, the dissertation adds covariance parameters to the algorithm to capture the target shape and adjust the tracking path immediately. The simulation results proved that the new algorithm is accurate more than 40% of the traditional algorithm in the premise of guaranteed real-time.Finally this dissertation analyzes the theory and tracking algorithms of particle filtering, and proves particle filter with a higher reliability when it deals with multiple objective blocked. Based on the research front, the dissertation proposes a multiple target tracking system composed of optimized mixture Gaussian background, Mean-Shift algorithm and particle filter. The system uses optimized mixture Gaussian background to capture observations. Then it matches the observation and target trajectory based on logic, combines conflicting objectives with similar acts. According to blocked case, it selects the appropriate tracking algorithm. When blocked doesn't happen, the optimal Mean-Shift algorithm is selected, which could capture the shape change of moving objective. When blocked occurs, the particle filter with particles integration would be selected.Proposed tracking system combines accuracy of Gaussian mixture background, real-time of Mean-Shift algorithm and stability of particle filter reasonably. The simulation results prove that the tracking system presented above can handle object blocked, merge and split, etc. Besides, its accuracy performance is close to particle filter. Real-time performance has been improved significantly.