Research On Multi-target Tracking Technologies Using Random Finite Set

Multi-target tracking(MTT)refers to the problem of jointly estimating the number of targets and their states or trajectories from noisy sensor measurements.Stable and efficient multi-target tracking algorithm is the core of the multi-target tracking technology and system,also the focus of this article.The current trend in multiple target tracking is to deal with complicated scenes,in which target number is unknown and varying with uncertain detection,measurement source and data association.Mahler's Random Finite Set(RFS)based multiple target tracking technology naturally suits such complicated scene.The most prominent advantage of these algorithms is that joint detection and state estimation for targets whose number is unknown and varying over time could be completed without complicated data association.RFS based multi-target tracking algorithms are successfully applied in the field of target monitoring and defense,autonomous vehicles and robotics,remote sensing,computer vision,biomedicine,modern communication and so on,represents a new direction for the development of multi-target tracking technology.This paper focuses on joint multi-sensor bias estimation and multi-target tracking using RFS,track extraction for RFS based filters,multiple group target tracking using RFS,multiple image target track-before-detect(TBD)using RFS,posterior cramer-rao lower bound(PCRLB)for RFS based filters.The main results obtained are as follows:The second chapter proposes a new joint multi-sensor bias and multi-target state estimation algorithm based on the multi-group multi-target probability hypothesis density(MGMT-PHD)filter and hierarchical point processes modeling.The parent processes is the set of the multi-sensor biases,the daughter processes is the set of multi-target state spaces corresponding to the multiple sensors.By separating the two interacting point processes,a large computational complexity arising from jointly estimating multi-sensor biases and multi-target states in an augmented high dimension state space for the existing methods can be avoided.When the MGMT-PHD filter is used to jointly estimating multi-sensor biases and multi-target states,the number of the sensors is known,that is to say it is not required to estimate the number of elements in the parent processes.In addition,each sensor independently obtains the corresponding measurement set,which means that the partition of the measurement set is determinate and unique.The implementation of the MGMT-PHD filter can be greatly simplified.Then,the particle implementation of the MGMT-PHD filter is proposed to jointly estimating the multi-sensor biases and multi-target states under nonlinear conditions.Simulation results demonstrate that the proposed algorithms behave in a manner consistent with our expectations.The third chapter is to address the problem of how to handle joint rapid detection,identification and stable track of decoy and aircraft within radar beam in the interception endgame with active radar seeker when aircraft and decoy are vertically resolved.In initial part,an improved labeled particle probability hypothesis density(IL-P-PHD)filter is proposed through improving traditional L-P-PHD filter.Then,using multiple-model(MM)method,a MM-IL-P-PHD filter for interception of highly maneuvering target is developed.Finally,based on the proposed MM-IL-P-PHD filter and echo amplitude fluctuation characteristic based interference detection method,a comprehensive frame for joint rapid detection,stable track and reliable identification of aircraft and decoy which are resolved on range dimension within radar beam is constructed.Simulation results are presented to show the effectiveness of the proposed frame.The fourth chapter proposes a new RFS based algorithm,which is able to jointly estimate the motion states and number of the group centers,the motion states and number of the group components,and obtain the tracks of the group centers,by modeling multi-group targets as a hierarchical point processes.The basic ideas and the main contributions of the proposed algorithm are as follows.1)The measurement update process of the UT-PHD filter is described in detail,and the specific calculation method of this process is given.The particle implementation of the UT-PHD(P-UT-PHD)filter is proposed.By adding identity label to each state particle,labeled particle UT-PHD(L-P-UT-PHD)filter is proposed,which can jointly estimate the number of group targets,states of the group target centers,and extract the tracks of multi-group target centers.Joint detection and tracking of the multi-group target centers is realized.2)Using the estimated group center states,a more accurate measurement set partition algorithm is proposed to partition the measurement set.Each cell in the obtained partition is assigned to the corresponding single-group particle PHD filter for each group target.The single-group particle PHD filter estimates the component states and the number of components.The estimated results are then fed back to the L-P-UT-PHD filter.The experimental results show that theproposed algorithm can well detect the appearance and disappearance of group targets,estimate the motion states of the group centers,obtain the tracks of the group centers and estimate the number and motion states of components.The fifth chapter is to contribute to an extremely challenging task,joint detection and tracking of multiple image dim targets when target influencing areas continuously non-overlapping or overlapping.The classic PHD-TBD algorithm firstly introduced the classic PHD filter into infrared image dim target tracking before detection.While,its response for detecting new targets is slow and the precision for its target number estimation is unsatisfactory.To obtain better detection and tracking performance,this chapter proposes general PHD-TBD(GPHD-TBD)algorithm as well as its particle implementation.As to multiple image dim target tracking when target influencing areas continuously overlapping,the contributions and innovations mainly include two points.Firstly,we construct a superpositional measurement model for multiple overlapping image targets and derive corresponding multiple target measurement likelihood function.Secondly,on the basis of this model,we propose labeled AS-PHD filter with introduction of the AS-PHD filter into image target tracking and labeling the state space of AS-PHD filter,and further propose particle implementation of labeled AS-PHD filter with application of sequential Monte Carlo(SMC)method,which finishes tracking before detection of multiple overlapping image targets on low signal-to-noise ratio.Numerical examples demonstrate that the proposed algorithm behaves in a manner consistent with our expectations.The sixth chapter researchs the performance lower bounds of multi-target tracking algorithms in dealing with complicated multi-target tracking scenes.Contributions from this chapter mainly include three points.Firstly,under the random set frame,this chapter derives a multi-target PCRLB(MT-PCRLB)for complicated multiple target tracking problems,and its iterative calculation expression for lower bound of such algorithms.Secondly,based on obtained multi-target tracks using IL-P-PHD filter,we propose a new data association method for precisely obtaining association relations between multiple target tracks and measurement set in this work.Thirdly,based on this new data association method,this chaper derives the detailed expression of MT-PCRLB for evaluating lower bounds of typical radar multiple target tracking problems.Besides,such lower bound is compatible with current labeled FISST based filters.Then,based on the association relations between multiple target tracks and measurement set,iteration of MT-PCRLB is accessible.Simulation experiment shows that MT-PCRLB proposed is capable of qualitatively evaluating the lower bounds of multiple target tracking algorithm's performance in dealing with complicated multiple target tracking scenes.The seventh chapter makes a summary of the thesis,while several open problems are proposed.