Detection And Motion Estimation Of Moving Object In Static Scene

Due to the computer technology, image processing, artificial intelligence, computer vision and other scientific research, intelligent monitoring more and more attention by researchers. Moving object detection and tracking is the base of intelligent monitoring, and it is the purpose of this study to extract the moving objects quickly and accurately from the background. Target detection algorithm for the stability of the complex background and target tracking algorithm to maintain the accuracy and timeliness of the research topic.In this thesis, the detection and tracking of moving objects in fixed scenes is studied, the main work and research results are as follows:First, several kinds of video moving target detection methods and mathematical morphological processing methods are introduced. The selection of the size of structure element in the process of mathematical morphology is discussed. The real-time performance and accuracy of the algorithm are compared with several moving object detection methods. Based on the comparison results, the Gaussian mixture model background modeling method and frame difference method has the advantage of, avoid their disadvantages, is given a feature descriptor algorithm, Gaussian mixture model based on fading. The algorithm on Gaussian mixture model is improved. Through the single frame image feature description, image to determine the complexity, after the non maxima suppression, introduce mask idea of the establishment of the spatial correlation between pixels, with fading factor feature descriptors to solve the correlation problem, makes the improved Gaussian mixture model can according to the scene complexity adaptive regulatory model number. And combined with the frame difference method, according to the stability of the scene, the stability of the model is reduced, the number of the model is reduced, and the computational redundancy is reduced.Secondly, the optimal Bayesian estimation method under the framework of. Computer vision tracking can be understood as the posterior probability of solution, the solution to the Bayesian framework, the KF, UKF and IMM based on posterior probability, based on motion estimation of target tracking, the theory of optimal estimation is achieved. KF, UKF and an IMM tracking performance compared, for moving object non-linear and model error, try to using UKF to solve the target problem of nonlinear equation, using the IMM to solve the problem of the model motion error, guarantee of video target tracking accuracy. And the selection of IMM model set is discussed.Finally, the introduction of two commonly used adaptive sampling period algorithm, based on adaptive sampling period algorithm, the approximate location of the target video process too much give useful information of a use of moving target area change rate and covariance characterization of maneuvering characteristics method. Using the change rate of the area of the moving target to characterize the motion posture change of the non rigid object, and solving the problem of the lag of the sampling period according to the filter residual error. Through a comparative analysis of the residuals, using prediction residuals of the model error is smaller, more sensitive as the target of non regulation basis of the sampling period under dynamic conditions. Simulation results show that the desired tracking accuracy in that condition, the average sampling period is reduced, saving the system resources.