Visual Multi-target Tracking Algorithm Based On Multi-bernoulli Filter

Visual target tracking is one of the hot issue and most difficult problems in the field of computer vision.Especially,the research on multi-target tracking technology is related to the problem of information fusion.Whether in the military or in civilian filed,the visual multi-target tracking has important research significance and has shown a wide range of applications.This paper is supported by the National Natural Science Foundation of China(No.60130157),and makes a thorough and systematic research on the application of multi-Bernoulli filter in the field of visual multi-target tracking algorithm.The research contents and main results are as follows:1.To solve the problem that it is difficult to obtain the accurate estimations of the multiple target states with complex environment,this paper proposes an adaptive multi-target tracking algorithm based on kernel-density and multi-Bernoulli filter.First,the kernel density background subtraction technique is introduced in this paper,which can effectively restrain the background interference.Then,continuously adaptive mean shift(CAMShift)method is integrated into the framework of multi-Bernoulli filter,and adaptive mechanisms are proposed to handle problems of closely spaced target tracking and the variation of target size.In addition,particle labeling technique is introduced to identify the path of each target in the video.Experimental results show that the proposed algorithm with strong robustness can effectively achieve the visual multi-target tracking in complex circumstances.2.Aiming at the problem that it is possible to make errors under the framework of multi-Bernoulli filter because of the features of multi-target are too simple,this paper proposes an adaptive multi-target tracking algorithm based on characteristic covariance and multi-Bernoulli filter.First,the feature covariance is introduced in this paper,which combined with the integral graph method.Then,and adaptive mechanisms are proposed to handle problems of closely spaced target tracking and the variation of target size,and particle labeling technique is introduced to identify the path of each target in the video.Finally,the Monte Carlo particle filter is integrated into the framework of multi-Bernoulli filter,Experimental results show that the proposed algorithm with strong robustness can effectively achieve the visual multi-target tracking in complex circumstances.3.In order to solve the problem that particle filter multi-Bernoulli(PF-MeMBer)requires many particles to approximately the posterior probability density distribution of target states,which is more time-consuming and prone to particle degradation and sample dilution.This paper proposes an adaptive multi-target tracking algorithm based on QMC-GPF and multi-Bernoulli filter.First,the Quasi-Monte Carlo method is used to replace the traditional Monte Carlo framework to solve the degeneration and impoverishment of the particles.Then,the Gaussian mixture Particles are used to approximate the posterior probability distribution of the target state,and it avoids the resampling and particle degradation.The proposed algorithm greatly improves the utilization rate of the sampled particles in order to achieve better tracking effect by using fewer particles and improve the tracking performance.