Research On Multiple Extended Targets Tracking Algorithm Based On GLMB Filter

Random finite set theory provides a simple Bayesian recursive form for multiple target tracking,which can effectively avoid data association and reduce the computational complexity.On the basis of the labeled random finite set framework,the generalized labeled multi-Bernoulli(GLMB)filter can solve the problem of target trajectory recognition and realize multiple target tracking in real sense.With the improvement of sensor resolution,the traditional point target theory is not applicable any more.Thus,multiple extended target tracking,which takes the shape of target into account,has drawn the attention of domestic and foreign scholars and then has become a hot research field of target tracking.This dissertation focuses on the research of multiple extended target tracking algorithm based on GLMB filter.The main achievements are as follows: Firstly,aiming at the problem of poor real-time tracking performance caused by high computational complexity,a FDPC-GLMB multiple extended target tracking algorithm is proposed.The algorithm calculates the local density and density distance of measurements,determine the measurement source according to these two parameters,then eliminates the clutter and clusters the measurements from the same target.Compared to the traditional partition algorithm,the proposed approach only uses a most effective partition to update the target state,which can effectively reduce the computational burden and ensure the real-time performance of target tracking.In addition,aiming at the underestimation of target number and the degradation of target centroid estimation caused by the inaccurate measurement partition in the scene of target neighbor,using K-means clustering and fuzzy C-means clustering to improve the FDPC-GLMB algorithm.Simulation results show that the proposed algorithm can partition the measurements in the neighborhood scene accurately,and improve the estimation accuracy of target number and the centroid state effectively.Secondly,aiming at the problem that the random matrix method can not model the extended shape of irregular target,a GLMB algorithm based on Star-convex random hypersurface model is proposed.On this basis,a GLMB algorithm based on EM Star-convex random hypersurface model is proposed by using the EM algorithm to update the shape parameters repeatedly in ordet to solve the problem of poor shape estimation caused by insufficient measurement information in low SNR environment.The proposed algorithm models the extended target shape by random hypersurface model,finds the global optimal association hypothesis in the GLMB framework and update the centroid state and shape respectively by the measurements in the hypothesis.Simulation results show that the proposed algorithm outperforms the traditional extended target tracking algorithm based on random matrix in the estimation accuracy of the centroid state and extended shape.Finally,aiming at the problem that the estimation accuracy decreases due to the improper selection of motion model,the joint tracking and classification algorithm is applied to GLMB extended target tracking theory,a GLMB algorithm based on random matrix and a GLMB algorithm based on target extend measurement are proposed respectively.The proposed algorithms classify targets by different features,adapt the probability of the target categories,and update the target centroid state by the corresponding multi-model GLMB filter.Simulation results show that both the proposed algorithms can effectively improve the accuracy of target state estimation.