Research On Distributed Multi-sensor Fusion Algorithm Based On State-Extended Multi-Bernoulli Filter

Based on local sensor filtering,distributed sensors communicates with its neighbors to exchange multitarget posterior densities,and fuses these posterior densities well to improve the tracking performance.Affected by the detection capability,field of view of sensors,midsection of local sensor and clutter,posterior densities between sensors are quite different,then fused sensors may occur midsection.In this case,generalized covariance intersection method can’t be implemented for distributed fusion to improve the tracking performance.It is still worthy of in-depth research to explore effective multi-sensor fusion methods.Due to its low computational complexity and tracking performance,multi-Bernoulli filter that based on random finite sets has received extensive attention since proposed.On the basis of multiBernoulli filter,Generalized Labeled Multi-Bernoulli filter and Labeled Multi-Bernoulli filter have been proposed successively,and can provide the function of target track management.This paper will study a multi-sensor multi-target tracking method based on distributed fusion under the labeled multi-Bernoulli framework.The main contents are as follows:1.Aiming at problems that label inconsistencies among different sensors and high computational complexity with cardinality underestimation by using the Generalized Covariance Intersection(GCI)method for the distributed multi-sensor multi-target tracking based on the labeled multi-Bernoulli(LMB).This paper introduce a distributed fusion algorithm of state-extended label matching under the labeled multi-Bernoulli filter framework.First,in order to overcome label inconsistencies and reduce label matching computation,a variable is extended in state vector for recording the matching history when the first label is matched.Then only the target-like LMB components are transmitted for fusion among the sensors and the "divide and conquer" strategy is introduced for fusing surviving targets,newborn targets,and misdetection targets.Moreover,the misdetection targets and false alarm targets are recorded on a table and the misdetection target posteriors fed back for compensation,which can effectively alleviate the degradation of GCI fusion accuracy caused by target misdetection.Finally,the experimental results prove the effectiveness and robustness of the proposed method.2.Aiming at the problem that GCI fusion method will lead to the misdetection in nonoverlapping areas on account of the limited fields of view of sensors.A label multi-Bernoulli multi-target tracking method based on limited field of view sensors is proposed to solve the problem.First,sensors exchange its posterior with each other by flooding,which will expands the fields of view of a single sensor.Sensors may get the posterior of all targets in the whole sensor network’s fields of view.Secondly,according to the characteristics of fusion targets,a partition-based clustering method is proposed which restricts the division of targets with constraint condition and different fusion strategies are adopted for different types of targets.Aiming at the problem of target track fragmentation caused by sensors directly extracting the cross-sensor moving target filtering posterior through neighbor sensors,a track maintenance table is proposed to correct the label of the cross-sensor moving target to maintain the track continuity.The experimental results show that the proposed algorithm enables each sensor to output the state of each target in the global fields of view of the sensor network,and ensures that sensors can keep track continuity of the cross-sensor moving targets.3.Aiming at GCI fusion that it is vulnerable to misdetection and complex to apply to,also with high computational complexity.Besides,it is not suitable for large-scale sensor network.This paper explores the application of arithmetic average fusion method to fuse LMB posterior.Firstly,sensors exchange its posterior with each other by flooding,then the partition-based clustering method presented in chapter 4 will adopted to solve match the targets from different sensor which is difficult to do by optimization algorithm.Secondly,according to the classification results,the labels of the targets in the cluster are corrected by define a benchmark sensor node,so the sensors will have the same label for same target.Finally,the arithmetic average fusion is performed on the targets posterior.Implementation details are given based on the Gaussian mixture,which only needs to calculate the average of the existence probability of the targets in each cluster,and re-weight the Gaussian component representing the targets’posterior densities of the LMB component.Experiments show that the proposed algorithm in this chapter can effectively improve the performance of multi-target tracking in two experimental scenarios.