A Shape-driven Based Multiple Extended Target Tracking And Classification Method

As the resolution of the sensor increases,a target can occupy multiple resolution cells which is called extended target.In recent years,the research on extended target tracking has received extensive attention from scholars at home and abroad.Particularly with the appearance of Random Finite Set(RFS),it is widely used in multiple extended target tracking,which solves the problem of multiple extended target tracking with unknown and changing target numbers in complex environments,and also applies such technologies to civil,military,medical,etc.field.Although the multiple extended target tracking technology has been further developed,the problems of difficult to effectively divide the measurement and the estimation of irregular shapes are still the key and difficult problems in the field of extended target tracking.Focusing on the application of random finite sets in extended target tracking,this paper mainly solving problems such as the partition and tracking of closely spaced target measurement sets,prediction of unknown new targets,and track association.The main achievements are as follows:1.Aiming at the problem of partitioning the closely spaced target measurement set and subsequent tracking in multiple extended target tracking,a B-spline Shape Driven Probability Hypothesis Density(BSD-PHD)filtering method is proposed.This method first aims at the problem of high time complexity when traversing candidate positions in the measurement set by the traditional shape selection partitioning(SSP)algorithm,and introduced the Kernel density estimation(KDE)method into the original algorithm to find candidate positions,finally,the BSD-KDE-SSP method is proposed to select the best partition scheme by B-spline shape matching strategy.Afterwards,when the target is in closely spaced or cross motion,only the position information of the target is used to update it,which will cause missing or wrong tracking,but using the B-spline shape matching strategy in the algorithm to update it,the correct target state can be obtained,and then the shape category information of the target can be obtained accurately.Experiments show that BSD-PHD-SSP can effectively improve the efficiency of the algorithm.In the experiment,the BSD-PHD filtering algorithm was compared with ET-GM-PHD and Gaussian inverse Wishart PHD(GIW-PHD)respectively.The algorithm can more accurately estimate the motion,shape and shape category information of the target without changing other parameter values.2.In the multiple extended target tracking,a multiple extended target tracking algorithm based on the correlation matrix is proposed to solve the problem of large calculation of Gaussian mixture filtering algorithm and unknown new target.The traditional ET-GM-PHD filtering method needs to measure and traverse each Gaussian component in the filtering at each moment,which seriously affects the computational efficiency of the algorithm.Therefore,a multiple extended target tracking algorithm based on the correlation matrix is proposed.In this algorithm,the correlation matrix between the target measurement state and the predicted Gaussian component state is introduced into the filtering method,and the components with high similarity between the measurement and the Gaussian component state are retained,other components that have not been updated and matched may be clutter components directly discarded,which effectively improves the execution efficiency of the algorithm.When there is a new target during the movement,there is a new target measurement set in the measurement column in correlation matrix.Experiments show that the correlation matrix can not only improve the efficiency of the algorithm,but also complete the identification of unknown new targets.3.Aiming at the problem that accurate track association information cannot be provided in multiple extended target tracking,two Gaussian component track label association algorithms under two filtering frameworks are proposed.First,the Gaussian component label is introduced under the BSD-PHD framework,which solves the problem that the track information of the target cannot be accurately extracted when the multiple extended targets are in closely spaced or cross motion.Then,the Gaussian component label is introduced under the BSD-PHD filtering framework based on the correlation matrix,which solves the problem of track estimation in iteratively increasing the number of the Gaussian components in Gaussian mixture.Meanwhile,resolved issue with track estimation for newly born target.Experiments show that the proposed algorithm has a good effect on the track extraction of extended targets.