Research On Moving Objects Detection And Tracking Algorithm In Video Surveillance

Intelligent video surveillance is a process that analyses automatically the image sequence from the camera by using the computer vision and video analysis method in the case of none human supervision,which can achieves the objects detection, identification and tracking to judge and dispose in the higher level. This thesis is committed to the research on the moving objects detection and tracking methods in video surveillance system.For objects detection, first of all, background subtraction algorithm based on Gaussian mixture model is verified by experiments. Based on the experimental analysis, an improved method is proposed to resolve the problem of Gaussian mixture model misclassing the background pixels as foreground pixels caused by the abrupt illuminance change. The improved method selects the new model parameters and improves the model update mechanism which uses a fixed learning rate and updates the variance using the adapting rate, so that it can adapt to the disturbance of local light. For global abrupt illuminance change, the improved method updates the mean selectively based on the rule of frames. Second, according to the disadvantage of background model using a single information such as color information and texture information, a background modeling method based on the color and texture information is proposed, which combines the color information based on Gaussian mixture model and the texture information based on Uniform LBP background model in the decision-making layer. This method not only preserves the advantages of Gaussian mixture model, but also solves the problem of missing target when the color of object does not constrast the color of background model enough, which can eliminates the shadows effectively.For objects tracking, objects tracking methods based on Mean Shift algorithm and Kalman filter are introduced and are verified by experiment about small targets. To solve the problem that the effect of tracking small objects with Mean Shift algorithm and Kalman filter are unsatisfactory, the traditional algorithm has improved. The improved algorithm combines the advantage of the two algorithms, which according to the information of the previous frame target center of mass and motion displacement predicts the possible locations of object in the current frame, and then uses Mean Shift algorithm looking for final destination in the neighborhood of this possible locations. The experiments show that this method can track the small objects effectively.