The Extended/Group Target Tracking Under The Complex Conditions

Multi-targets Recursive Bayesian filter is capable of fusing and processing the totally different style data from information source. Based on the theory of Finite Set Random(RFS), Probability Hypothesis Density(PHD) is a kind of approximate technique of Multi-targets Recursive Bayesian filter, which can track multi-targets as a whole. PHD can estimate the multi-targets state and the targets number simultaneously. Unlike the traditional target tracking technology, PHD avoids complex data association problems,so it is suitable for extended/group target tracking perfectly. In recent years, the PHD of extended/group target tracking algorithm has received extensive attention from international and domestic scholars. The existed filtering methods can track extended/group target accurately, such as Extended/Group Target Gaussian Mixture Probability Hypothesis Density(Extended/Group Target-GM-PHD, ET-GM-PHD) and Gaussian Inverse Wishart Probability Hypothesis Density(Gaussian inverse Wishart-PHD, GIW-PHD) etc.However, the original filtering method will not be able to track the extended/group target effectively under the complex conditions, such as multi-sensor environment, crossing targets case, as well as the monitoring area overlap. This thesis is supported by National Natural Science Foundation project “Some Problems Research of Multiple Targets Tracking Method Based on Random Finite Set Theory”(NO.61201118). Aiming at the problems of extended/group target tracking under the complex conditions, the thesis researches the filtering method for extended/group target deeply. On the basis of existing methods of extended/group target tracking, thesis analysis the practical problems and improves the existing algorithms, so that it can track extended/group target under complex conditions effectively. The specific work of this thesis includes:(1) Aiming at the problem of extended/group target tracking with shape under multi-sensor environment, two algorithms are proposed, i.e., Gaussian Inverse Wishart Parallel PHD(GIW-PPHD) and Gaussian Inverse Wishart Sequential PHD(GIW-SPHD). New algorithms combine the ideas of parallel filter and sequential filter respectively. They are both effective to estimate the centroid and the shape of extended/group target. In the GIW-PPHD, the measurement sets generated by all sensors at the same time are combined into one measurement set, and then this measurement set is partitioned. In update stage, the partitioned measurement sets are used to augment the measurement vector, thereby the multi-sensor tracking problem is translated into a single sensor tracking problem. In the GIW-SPHD, the measurements generated by all sensors are partitioned respectively, and then are used to update in sequence. In this manner, the multiple sensors’ measurements are fused all together.(2) If the multi-sensor monitoring region is overlapped and the traditional sequential filtering algorithm is used to track the extended target, the estimated target number will miss. In order to solve this problem, firstly, the measurements are updated by using current predicted Gaussian component, rather than updated Gaussian components that belongs any sensor, so that the updated results using each sensor keep an independent. Then, when all sensor measurements are updated, the algorithm will judge the Gaussian component, if the target positions of components locate in the overlapped region, this part of the Gaussian weight will be adjusted. Finally, all Gaussian components are pruned and merged.(3) In the process of extended target tracking, if there are some crossing targets, target missing problem in target estimation will appear when the extended target Gaussian mixture probability hypothesis density filter is used directly. Aiming at this problem, an improved algorithm is presented. Firstly, Euclidean distances among targets estimated are calculated at each time, and then these targets are distinguished if they are in an adjacent region or not. After that, in the next time step, Gaussian components weights of the targets in the same adjacent region are compensated if the target number estimated become small; otherwise, there is nothing to do. Finally, these Gaussian components are applied to estimate and track the targets. The improved algorithm solves the estimated target missing problem when the extended targets are so close with partitioning into the same subset.