Extended Target Tracking Algorithms Based On Random Matrix

With the development of high-resolution sensors,such as radar,infrared,laser,etc.the more abundant information of targets can be obtained.Due to the difficulty of utilizing the obtained target information of the point target model,the more complicated extended target model is needed to accurately estimate the target state,which also poses a great challenge to target tracking algorithms.The traditional point-target tracking assumes that a target generates no more than one measurement per time step and only gives the kinematic state estimation of the target.For the situation that multiple measurements can be generated by a target per time step,to fully dig target information contained in measurements,extended target tracking algorithms need estimating the extended,measurement rate and classification states,in addition to the kinematic estimation.Those state estimations provide an accurate and multi-angle views of targets for the later procedures of a tracking system,such as recognition and classification.However,due to the complex tracking environment,multiple extended target tracking and extended state estimation of an extended target with complicated shape have many problems,such as high computational complexity,deteriorated extended state estimation,which has become the focal and difficult point in the field of the target tracking.Recently,the advent of the random matrix and random finite set(RFS)provides a new path to the research on this subject.Utilizing the random matrix method,the research of this dissertation puts its emphasis on multiple extended targets tracking based on RFS theory and complicated shape extended state estimation.The main contributions of the dissertation are listed as follows:1.The biased cardinality estimation problem of the multi-target multi-Bernoulli(Me MBer)tracking framework has been studied.To obtain the close estimation form,the inappropriate approximations are adopted by the Me MBer filter during its derivation,which leads to a biased cardinality estimation problem.The cardinalized balanced Me MBer filter amends the biased cardinality estimation by introducing the high detection probability limitation.To deal this problem,a new way to amends the biased cardinality estimation of the Me MBer filter is given,and the framework with the unbiased cardinality estimation based on the Me MBer filter is obtained without introducing extra assumption.To expand this framework into extended target tracking situation,an improved Me MBer extended target filter is proposed.The proposed algorithm can well accommodate the tracking scenarios with relatively low detection probability and achieves a superior tracking performance at cost of lower computational complexity.2.The framework construction of join tracking and classification of multiple extended target problem has been studied.An extended tracking algorithm is capable of estimating the target kinematic state,as well as classifying the extended target based on the size and structure of extended state of the extended target.However,for the sake of convenience,the existing algorithms are developed under the single target tracking and non-maneuvering assumptions,and more practical issues,such as multi-target tracking and maneuvering problems,are not considered.To deal this problem,a probability hypothesis density join tracking and classification of extended targets filter is proposed based on RFS theory.The proposed algorithm is capable of estimating target measurement rate,kinematic,extended and classification states,as well as the unknown and time-varying number of multiple targets.3.The varying number of sub-ellipses used by a non-ellipsoidal extended target tracking algorithm for reasonably describing the complicated extended state with changing attitude problem has been studied.The non-ellipsoidal extended target tracking algorithm utilizes multiple sub-ellipses to estimate the complicated extended state of the extended target and each sub-ellipse describes a part of the extended target,which gives an accurate extended state estimation by dividing the target into parts.The number of sub-ellipses used by existing non-ellipsoidal extended target tracking algorithms is assumed to be a constant.However,the attitude of the target with respect to the sensor is changing due to maneuvering,such as spinning and rolling of the target.The number of sub-ellipses used by the algorithm for describing the complicated extended state is changing.To deal this problem,a varying number of sub-ellipses non-ellipsoidal extended target filter is proposed.The proposed algorithm designs the spawning and combination criterion between sub-ellipses,and can well accommodate the scenario with changing attitude of a non-ellipsoidal extended target and gives an accurate estimation of the complicated extended state.4.The extended state estimation of a non-ellipsoidal extended target plunging into a local optimization problem has been studied.Due to maneuvering of a target,not only the number of sub-ellipses used by a non-ellipsoidal extended target tracking algorithm is changing,but also the target is illogically divided into parts by sub-ellipses,which deteriorates the extended state estimation and plunges the algorithm into a local optimization.In addition,its computation complexity increases exponentially with the increasing number of sub-ellipses,which brings a heavy computation burden to the algorithm.To deal this problem,an efficient global optimization non-ellipsoidal extended target filter is proposed.The proposed algorithm combinatorially use the measurement clustering,adjacency matrix and probability look ahead methods,not only guarantees the avoidance of local optimization in extended estimation,but also greatly reduces the computation complexity,which has a wide range of applications.