Research On Multiple Extended Target Tracking Algorithms Based On Labeled Random Finite Set

With the rapid development of high-resolution sensor technology such as modern radar,in the field of information fusion,multiple extended target tracking technology has become a research hotspot,and is widely used in military fields such as ship-borne early warning and aerial reconnaissance,as well as civilian fields such as intelligent driving and computer vision.In order to meet the practical requirements,multiple extended target tracking not only needs to estimate the number and kinematic states of targets,but also needs to estimate the extension states,measurement rates,tracks and classes of targets to improve the performance of weapon systems.However,with the increasing complexity of tracking scenarios in reality,the performance of multiple extended target tracking algorithms is faced with severe challenges due to unknown and time-varying target number,detection uncertainty and data association uncertainty.The labeled random finite set theory developed in recent years can effectively deal with the multiple extended target tracking problem in complex environments,and has become a hot topic of scholars at home and abroad.However,the existing labeled random finite set multiple extended target tracking algorithms still have some problems,such as inapplicability to nonlinear systems,high computational complexity,and failure to consider the relationship between tracking and classification,which restrict the further improvement of tracking system performance.Based on the labeled random finite set theory,this paper focuses on the multiple extended target tracking under nonlinear conditions,the construction of efficient extended target tracking filtering framework and multiple extended target joint tracking and classification.The main research results are as follows:1.Aiming at the problem that the nonlinear of tracking system makes it difficult to solve the multiple extended target posterior probability density function,an extended target generalized labeled multi-Bernoulli(ET-GLMB)filtering algorithm based on gamma box particles is proposed.Firstly,based on the comprehensive consideration of the target measurement rate state,kinematic state,and extension state,a measurement likelihood function under the interval analysis framework is constructed.Then,the multiple extended target posterior probability density function is expressed in GLMB form,and the weighted gamma box particle set is used to approximate the probability density of each target.Through interval operation and approximation,the implementation process of extended target gamma box particle GLMB filtering is derived.In addition,the algorithm preprocesses the measurements and eliminates some clutter measurements,effectively improving computing efficiency.The simulation experimental results show that,for multiple extended target tracking under nonlinear conditions,the proposed algorithm is suitable and has good tracking performance.2.To solve the problem that extended target labeled multi-Bernoulli(ET-LMB)filtering in nonlinear tracking systems is difficult to obtain a closed solution,an ET-LMB filtering algorithm based on gamma box particles is proposed.The weighted gamma box particle set is used to approximate the probability density function of each hypothesis track in the LMB parameter set,and the parameter prediction and updating process of the gamma box particle ET-LMB filtering is derived by recursively passing the LMB parameter set in the filtering process.The simulation experimental results show that,under nonlinear conditions,the proposed algorithm can stably track multiple extended targets and has smaller computational effort than gamma box particle ET-GLMB filtering with slight estimation performance loss.3.In view of the high computational complexity of ET-LMB filtering and the difficulty of tracking multiple extended targets with unknown detection probabilities and different shapes under nonlinear conditions,an extended target fast LMB(ET-FLMB)filtering algorithm based on beta gamma box particle Gaussian process is proposed.Firstly,the ET-FLMB filtering framework is constructed by integrating the prediction and update of ET-LMB filtering,and using Gibbs sampling to select the hypothesis components with greater weights,so as to reduce the computational complexity.Then,the beta gamma box particle Gaussian process implementation under the proposed filtering framework is derived.The algorithm takes the unknown detection probability as the target augmentation state and recursively estimates it during the filtering process,which can estimate the unknown detection probability while estimating the target kinematic state.At the same time,the algorithm uses Gaussian process to model the extension state of the target,and can estimate any star convex shape.Finally,an optimal sub-pattern assignment(OSPA)distance based on radial functions is proposed to improve the extended target tracking performance evaluation system.The simulation experimental results show that the algorithm has high computational efficiency and can effectively handle the tracking problem of multiple extended targets with unknown detection probabilities and different shapes under nonlinear conditions.4.An ET-LMB joint tracking and classification(JTC)algorithm based on Gaussian process is proposed.Firstly,the target class is extended as a part of the target state,and the prior class information is introduced into the filtering process to construct the ET-JTC-LMB filtering framework.Then,the Gaussian process is used to model the extension state of the target,and the relationship between the extension state and prior class information,as well as a new class probability update method,are constructed.The two are integrated under the proposed ET-JTC-LMB filtering framework to achieve target classification and improve tracking performance.Finally,an evaluation indicator called class recognition rate is proposed to evaluate classification performance.The simulation experimental results show that the proposed algorithm can simultaneously track and classify multiple extended targets with different shapes,and improve the extension state estimation performance.