Research On The Algorithm Of Motion Target Detection In Complex Scene

As an important base task in the field of video analysis,the motion target detection technology is widely used in the intelligent video monitoring system,the intelligent traffic system,the video coding and so on.Reliable moving target detection technology is the foundation of the successful implementation of the subsequent high-level tasks.The paper mainly focuses on the motion target detection technology based on the background subtraction framework,and solves the difficult problems of the background subtraction algorithm in the practical application.In this paper,the frame difference method,optical flow method and background subtraction method in moving target detection are described in detail.Frame difference method has the characteristics of simple principle and high real-time performance.It is easy to use frame difference method to make the detection results incomplete,so it is more suitable for other algorithms to simplify the time complexity of the algorithm.The optical flow method has a rigorous theoretical basis and can adapt to the changes of the background scene,but it costs a great deal of calculation.Researchers usually combine optical flow and feature point matching to propose a solution to camera motion,especially to deal with PTZ cameras.The tradeoff between time complexity and processing accuracy is considered,and the background subtraction algorithm is concerned by researchers.All kinds of algorithms based on GMM model focus on adaptively adjusting the number of Gaussian distributions and reducing the sensitivity to light environment.All kinds of algorithms based on codebook model mainly consider multi-color model,multi-feature and multi-level modeling.All kinds of algorithms based on sample consistency model are modeled in pixels.Considering each pixel independently,it can make it have adaptive segmentation threshold,and try to control different update rate to restrain neighborhood diffusion.According to the processing framework of sample consistency model,a background subtraction algorithm based on sample dynamic estimation and contour similarity is pro-posed in this paper.From a simple and convenient point of view,the algorithm selects a single color feature.In the initialization phase,the algorithm not only establishes the conventional background sample,but also estimates the standby initial reference background.The thinning process of the segmentation results at the region level is realized by reference to the background.When neighborhood updating is used to detect intermittent object motion,the mechanism has a congenital defect.Instead of neighborhood updating,the algorithm uses contour similarity to identify static or low-speed foreground,ghost,and counter to achieve different updating methods,and finally achieve segmentation results refinement.In addition,a pixel-level adaptive feedback mechanism is implemented by using the recently observed minimum distance standard deviation as a background dynamic indicator.The algorithm selects the post-processing version of the binary result image as the update template,because canceling the neighborhood update means that the spatial position of the pixels is no longer considered,which further suppresses the noise interference.The change detection dataset evaluation results show that the change detection data set evaluation results show that: The proposed algorithm can adapt to video scenes with dynamic background and intermittent moving objects,and the overall evaluation performance is comparable to that of most advanced algorithms.The above algorithms are re-evaluated and the following problems are found: the modeling method based on a single feature is unable to balance the contradiction between the tolerance to noise and the sensitivity to illumination;The adaptive feedback mechanism may over-adjust the threshold of a certain pixel,and it is not reasonable to solve the camera motion only by accelerating the rate of update.A background subtraction algorithm with local adaptive sensitivity and contour similarity is proposed.The combination of texture features and color features makes the detection of camouflage foreground more sensitive.The improved pixel-level adaptive feedback mechanism enables the algorithm to self-adjust the parameters.When constructing the background sample,the reference background of initialization is estimated.The ghost suppression is accelerated by determining whether the foreground object exists in the initial reference background.The elimination of neighborhood diffusion overcomes the defect that the foreground object is absorbed to the background sample.In addition,the paper pays special attention to the solution of camera motion problem,adopts the same idea as the video steady-state technology,uses the fea-ture point matching to estimate the global offset distance between consecutive frames.To counteract the effect of camera motion on foreground segmentation,the evaluation results of change detection data set show that the algorithm shows superior performance.