Research On Multi-target Tracking Algorithms Based On Generalized Label Multi-Bernoulli Filter

In recent years,multi-target tracking technology based on random finite set(RFS)has brought new vitality to the multi-target tracking field because it does not need to deal with complex data associations and has relatively small computational complexity.However,the filtering algorithms in the random set model framework are not multi-target trackers in nature because they rest on the premise that targets are indistinguishable.By introducing label into random set,the problem of track identification is solved well,and the developed Generalized Labeled Multi-Bernoulli filter(GLMB filter)is also widely used in multi-target tracking.This paper focuses on the multi-target tracking method based on GLMB filter algorithm.The specific research content is as follows:Firstly,the core idea and implementation of GLMB filter and δ-GLMB filter are introduced.By introducing label idea into the framework of random finite set,the theoretical foundations of the GLMB algorithm and an important modification,called δ-GLMB filter,introduced in detail,and their specific predict/updating steps are given.Secondly,aiming at the problem that it's difficult for conditional joint decision and estimation(CJDE)algorithms to calculate the cost of decision and estimate with no prior density information,this paper presented a new multi-target joint tracking algorithm based on GLMB filter.We derive the optimal solution that minimizes the CJDE risk,which is a union cost of cardinality estimation,state estimation and decision estimation,and present an optimal CJDE algorithm.Besides,in order to solve the problem that the traditional δ-GLMB algorithm has a large estimation error for new-born targets,a new adaptive target-born algorithm is proposed.Using the new constructed adaptive model,the measure information received at the last moment is used to predict the survival probability and state distribution of the birth target at the current moment,and then the predicted birth target density is filtered and iterated.Simulation experiments verify the temporality and accuracy of the proposed algorithm.Thirdly,in order to deal with the problem that it's difficult for multi-target tracking filters available to track multiple targets accurately in spatial position overlapping scenarios,a more complete exposition of the derivation and implementation of the merged measurement GLMB filter is proposed.Furthermore,we present a computationally cheaper variant of the algorithm using relaxed measurement assignment to provide accurate mappings from measures to target groups.At the same time,in order to further improve the accuracy of the target state estimation,a smoothing algorithm is proposed.This algorithm performs forward and backward smooth recursion on multi-target states,and uses the measurement data at future times for backward smoothing,which improves the target tracking accuracy.Experimental simulations show that this algorithm has a good performance in multi-target tracking.