Design And Implementation Of Moving Object Detection And Tracking Algorithm In Video Surveillance

In the field of machine vision, the system tends to analyze the behavior of moving targets or make meaningful judgment for moving target scene , and these analysis and determine need to make the detection, recognition and tracking on a reasonable target.In this thesis, we make the relevant research in this context, and then get the design of the target detection and tracking algorithm, and compare with the current epidemic algorithm to judge the performance of the algorithm.Target detection is the basis for automated video analysis in many visual applications in terms of moving target detection. Target detection in video is generally achieved by a target detector or background separation technique. Normally, the target detector needs to manually mark the sample to train the classifier, and the background separation needs to be trained without a foreground-containing video frame. In order to achieve real-time detection, not through a separate training to achieve the target detection is critical. Due to the existence of complex environments (such as non-rigid movement, dynamic background, etc.), the use of motion information to solve the target detection ability is very limited. In this thesis, a moving object is detected by detecting a continuous anomaly frame in a low rank matrix, which combines target detection and background learning into a single optimization process, which can be achieved by the relevant iterative algorithm. The comparison between experiments and related algorithms shows that the algorithm can be effectively applied in complex video surveillance environment.For target tracking, this thesis proposes a robust algorithm for target tracking in complex environments such as motion blur, illumination change, pose change and occlusion. Through the different characteristics of the convergence to the kernel particle filter framework to overcome the complex changes in the environment, in which each feature of the study are involved in a feature of the target representation, in order to construct a different tracker. Using the interaction of multiple trackers and the ability to robustly track the performance of the interaction, the interaction is done using the transition probability matrix to complete the shared sample between the trackers, and the selection process is to determine the most reliable tracker as the final target tracking of the frame. The tracker probability is determined by the likelihood function of the tracker, and the transition probability matrix and tracker probability of the tracker are updated in the Bayesian framework. The experimental results show that the proposed algorithm has strong robustness and adaptability in target tracking.