Research On Background Modeling And Moving Target Detection Algorithm Based On Surveillance Video

With the development of image processing and computer vision technology,intelligent video surveillance system has been widely used in traffic,intelligent manufacturing,biomedicine and crime surveillance,etc.In intelligent surveillance,without any prior information,how to quickly and effectively construct the background model of surveillance video and extract the foreground moving target information accurately is the key to the success of target recognition,motion estimation and behavior understanding.Based on relavant theories,the Gaussian mixture model and Horn-Schunck optical flow method are improved.The main work of this paper is as follows:Firstly,nine kinds of background subtraction methods based on pixels are compared and analyzed in terms of the basic principle,parameter iteration updating,performance accuracy and complexity of the algorithm.At the same time,2623 papers on the topic of "background modeling" and "foreground extraction" collected by CNKI from 2003 to 2018 are analyzed from the perspective of bibliometrics in the way of visualization,which provides a good reference for the in-depth study.Secondly,for the initialization of background and the learning rate of parameter in traditional Gaussian mixture model,combining the Gaussian filter and mathematical morphology,an adaptive modeling method is proposed,where the background parameters are initialized based on the updating method of pixel region,at the same time,the change value of information entropy of image sequence is regarded as the feedback of parameter learning rate of adaptive updating Gaussian mixture model.Experimental results show that the method has high detection accuracy and fast processing speed for surveillance video with dynamic disturbance and large variation in the number of moving targets over time.Then,for the problems that the traditional EM algorithm needs a large amount of storage and computational memory,and updates parameters slowly in the application of Gaussian mixture model,an improved on-line EM algorithm is proposed to iteratively update the parameters of Guassian mixture model.According to Titterington on-line algorithm,the feasibility of the method is proved when the covariance matrix of Gaussian mixture model is diagonal matrix.At the same time,an adaptive adjustment method for the number of Gauss distribution is given.Experimental results show that the improved algorithm has better performance inbackground construction,target recognition performance and operation speed than the traditional one.Lastly,for the problem that the optical flow method with a large amount of calculation can not obtain the ideal prospect of surveillance video alone in complex situations,combining the Gaussian filter and mathematical morphology,a method of moving object region detection in video sequence based on normalized adaptive optical flow is proposed,where the Gaussian filter is introduced to smooth the surveillance video data,at the same time,Horn-Schunck optical flow method is adopted to calculate the optical flow,and an improved genetic algorithm is used to adaptively determine the two-dimensional OTSU threshold to process the noise of optical flow.Compared with the traditional algorithm,this method has the advantages of less computation,lower computational complexity,shorter computational time and higher performance evaluation index.