Research On Multi-target Tracking And Trajectory Maintenance Algorithms Based On Random Finite Sets

As one of the most active research fields in information fusion theory and advanced filtering method,multi-target tracking technology is widely applied in both military and civil fields represented by aviation and aerospace.Without the complex data association technique in traditional tracking method,the multi-target tracking approach based on the random finite set has attracted considerable attention from researchers and engineers in various countries.Based on the random finite set theory,this dissertation carried out a more in-depth,systematic research in the multi-target tracking and trajectory maintenance under some complex tracking environments by using the probability hypothesis density filter,which mainly includes the following aspects.1.For the problem of estimating the target states and their number in closely spaced targets tracking scenarios,a close multi-target GM-PHD tracking algorithm is proposed under the linear Gaussian assumption.After the target predicted intensity being updated by using the measurement set at each time step in the standard GM-PHD filter,the target weight redistribution method is used to detect and reallocate the unreasonable weights of the components of the targets in the target posterior intensity.Additionally,in the stage of pruning and merging of the Gaussian components in multi-target posterior intensity,a pruning and merging approach of target components is proposed based on both the labeling method and weight competition method of the components,which can avoid the fusion error problem caused by the important components of different targets to some extent.Compared with the existing correlated close multi-target PHD filters,the proposed algorithm can achieve high accuracy of target state estimates and accurate target number estimation.2.For the problem that it is difficult for the standard PHD filter to estimate the target states and their number in multi-target tracking environment when the newborn target prior intensity is unknown,a multi-target GM-PHD filtering algorithm based on the adaptive newborn target intensity estimation is proposed.By using the PHD pre-filtering technology and target velocity characteristic scheme,the adaptive newborn target intensity estimation method can obtain the measurements originated from the real newborn targets,and the unknown newborn target intensity can be modeled by using these measurements.In addition,a measurement-driven update scheme is introduced into the update step of the standard GM-PHD filter.At each time step,the latest measurement set is divided into survival target measurement set,newborn target measurement set and clutter set.In the update step,different types of the target predicted intensities are updated with their respective corresponding measurement set,and the clutter set is banned from updating the target predicted intensity.Simulation results show that the proposed algorithm not only has superior accuracy of target state estimation and low target number estimation error,but also has a stable and low computational burden.3.To solve the multi-target tracking problem in imperfect probability of detection environment,a multi-target GM-PHD filtering algorithm is proposed based on the update strategy of target weights being integrated the exponential attenuation function and multi-frame target state extraction scheme.By using the exponential attenuation function and the previous weights of targets,update strategy of target weights attenuates the weights of the pseudo missed-detection targets caused by the unreasonable distribution of the measurements originated from the targets in the state space such that each target component in target posterior intensity can obtain a reasonable and effective weight.By referencing the multiple historical weights of each target,multi-frame target state extraction scheme extracts the state estimates of the true missed-detection targets caused by the loss of the target measurements in the low probability of detection environment from the target posterior intensity at each time step.Simulation results show that the proposed algorithm can obtain better estimation accuracy in terms of the target states and their number,and the filtering performance is relatively stable in imperfect probability of detection environment.4.To realize the multi-target trajectory maintenance in closely spaced targets tracking environment,a multi-target track continuity algorithm is proposed within the framework of the GM-PHD filter.Compared with the classical GM-PHD tracker,the proposed multi-target trajectory continuity algorithm is composed of the association and update scheme of target states and the irregular window based track management scheme at each filtering iteration.Based on the multiple previous states of the target components in target predicted intensity and the latest obtained measurement set,the association and update scheme of targets constructs an association-update factor matrix for updating the target predicted intensity at each time step.Then,the association-update factor matrix is utilized to realize the synchronization of the target predicted intensity update and the optimal association between the targets and measurements.By making full use of the state estimates of the target trajectory in a period of time,the irregular window based track management scheme not only effectively maintains the track continuity of the real targets,but also effectively solves the false target tracks generated by the false alarms or clutter in the filtering process.Simulation results from the multi-target cross and parallel tracking scenarios demonstrate that the proposed target trajectory maintenance algorithm cannot only improve the estimation accuracy of the target states and their number,but also has a relatively low computational cost and an excellent target track maintenance performance.