Research On Extended Target Tracking Algorithm Based On Labeled Random Finite Sets

Random finite set theory provides a relatively straightforward Bayesian recursive form for multiple target tracking and avoids high complexity which exists in conventional approaches due to data association.On the basis of the labeled random finite set framework,the generalized labeled multi-Bernoulli(GLMB)filter realizes the distinction of target trajectory in a real sense.Since the accuracy of sensor has made a great improvement,the extended shape of target becomes non-negligible.Traditional point target theory cannot meet various needs in actual implementations,thus,extended target tracking has drawn the attention of domestic and foreign researchers and has been widely employed in both military and civil systems.This dissertation mainly focuses on the application of GLMB filter in extended target tracking.The main achievements are as follows: First,aiming at the problem that GLMB filter requires background environment parameters in the observation region while these parameters cannot be obtained in advance in real-world scenario,a GLMB extended target tracker under circumstance of unknown clutter rate and detection profile is proposed.This approach treats clutter as another kind of target with no dynamics and establishes joint target-clutter state space model to jointly propagate the hybrid multi-target density.In the meantime,the extended target part of the joint state space model is augmented with an additional detection probability space and a Beta distribution is applied to directly model the unknown detection probability.Besides,to solve the problem of a wide range of fluctuation of Beta distribution estimation when target birth or death happen throughout the observation progress,an inverse gamma distribution is adopted to characterize the unknown amplitude feature of target and then the current detection probability in the surveillance area can be acquired with the amplitude feature.Simulation results indicate that both proposed algorithms are able to accommodate high clutter rate or low detection probability scenario and are capable of continuous tracking multiple targets and estimating the unknown parameters and have almost comparable performance compared with the conventional filter with a priori knowledge about those parameters.Compared with Beta distribution,inverse gamma distribution has a steadier estimation result of the detection probability in the scenes where target birth and death appear.Then,in order to cope with the degradation of target tracking filter in estimating the cardinality,centroid and shape parameters of extended target caused by the inaccurate measurement partition,a likelihood-based single-step data association GLMB filter is given.After bypassing the progress of clustering measurements and assigning them with targets,the proposed filter search for the optimal association hypothesis directly through the calculation of the likelihood function based on possible assumptions of the association of measurements and targets.The posterior probability density of multi-target is then updated.In addition,multiplicative noise model is employed to represent the extended shape and the recursion of this tracker is also provided.The results from the simulation demonstrate that the proposed single-step data association approach has the ability of accurately estimating the number of targets in challenging scenarios with close targets.Improvement in the estimation of target central position and shape can be seen when compared with filters utilizing measurement partition algorithm.At last,for the problem of shape estimation under the theory of labeled random finite set,a GLMB extended target filter based on control points is proposed.This algorithm takes advantage of the extension-deformation approach to model the shape of extended objects and the real extension is deformed by moving some control points being on the boundary of a reference extension.Adaptive computation of the parameters of reference extension is used to improve the similarity after deformation.And the iteration process of multi-object posterior probability density under the labeled random finite set architecture is presented in the end.The proposed method is evaluated against the one based on random matrix and multiplicative noise model and the results show that the proposed filter can achieve precise estimation of elliptic shape parameters.