Muiti-target Detection And Tracking Based On Particle Filter And Background Difference

Detection and tracking of moving objects is a branch of computer vision and it has always been one of the research issue.With the development of computer technology,image processing technology improving.The application of video surveillance has penetrated into the fields of national education,life and entertainment,medical security and other fields from the traditional security field.At present,most of the target tracking methods in the field of video surveillance are based on the prior information of moving objects.Without prior information,these methods are not sufficient to track multiple targets simultaneously.This paper proposes an automatic detection and tracking algorithm,which is a new real-time algorithm based on particle filtering and background difference.It can automatically detect and track multiple moving targets without learning the stage or knowing in advance the relevant information,such as the size,nature and initial position of the object.The traditional Gaussian Mixed Model(GMM)algorithm is improved,and under the premise of ensuring accuracy,the model running speed is accelerated.Then combined with the results of foreground detection,a particle filter tracking algorithm without prior information is proposed.Focusing on the above algorithm,the main research work in this paper is as follows:(1)Research several classical algorithms for moving target detection and tracking.According to the analysis of the characteristics of various algorithms,combining with the context of this article,we choose the target detection and tracking algorithm that is suitable for this article.(2)Improve the traditional GMM algorithm.Aiming at the characteristics of background stable regions in surveillance video,a model clearing mechanism is proposed to reduce the number of Gaussian distributions in the region,thereby reducing the amount of calculation.In addition,a Temporary Gaussian Model(TGM)is established.It can reduce the frequent update operation of non-matching pixels and improve the detection rate.Then analyze and discuss TGM cleaning methods and related parameters of TGM distribution.the algorithm of this paper compares with the traditional GMM algorithm.Experiments show that the accuracy of the algorithm is guaranteed in different scenes,the overall speed of the algorithm is improved by about45%.(3)Combining the results of foreground detection,a particle filter tracking algorithm based on non-prior information is proposed.The color distribution model is integrated into the particle filter.And a new calculation is used to increase the efficiency of dealing with the target rotation and scale change.Four challenging sequences are selected from the CAVIAR dataset and the VTB dataset.Under the complicated environment of illumination change,small object tracking,scale scaling,deformation and background clutter,this paper compares with the classical seven tracking algorithms and evaluates the performance of the method from both qualitative and quantitative perspectives.Experiments show that this algorithm can achieve multi-target automatic tracking in a variety of complex scenes.