Research On Moving Multi-target Segmentation And Tracking Method

Advances in modern technology and sensor are driving the complexity of targets tracking scenarios.Recently,it appeals to scholars focus that group targets tracking suitable for actual complex environment.Group targets are composed of multiple targets with similar motion and certain distance.For multi-target tracking,the Gaussian Mixture Probability Hypothesis Density(GM-PHD)is one of the most mainstream algorithms.Therefore,many improvements on GM-PHD algorithms are proposed for group targets tracking,but such algorithms cannot obtain trajectories of targets with detailed information directly and accurately.Moreover,the tracking of group targets should take into account the group segmentation of targets and the complex dynamics of crossing,splitting and merging that may occur during the movement.To solve the above problems,the Gaussian Mixture Trajectory Probability Hypothesis Density(GM-TPHD)filtering algorithm is introduced in this thesis.The tracking problem of resolvable group targets and unresolvable group targets is studied respectively.The main work is as follows:(1)Research on resolvable group target tracking algorithm.The existing PHD-based series of resolvable group target tracking algorithms are not able to achieve real-time tracking of group target trajectories with high accuracy,and the tracking effect is relatively poor in the complex scenarios such as multiple groups in close proximity,overlapping regions,splitting and merging.The GM-TPHD was introduced into the tracking method for resolvable group targets to address the problem of not being able to track group targets directly and accurately.Aiming at the problem of poor tracking effect in complex scenarios,The GM-TPHD algorithm based on motion correlation and the GM-TPHD algorithm based on evolutionary network are proposed respectively.The GM-TPHD algorithm based on motion correlation obtains the motion information in the target trajectories,and combines the manifold distance,motion direction and magnitude to achieve accurate group segmentation in complex scenes by focusing on targets motion characteristics.The GM-TPHD algorithm based on evolutionary network uses the group internal information updated in real time by evolutionary network based on targets spatial characteristics to modify the targets state in the group through two methods of covariance weighting and affiliation,and then feeds back to the filtering,to improve the tracking accuracy of the algorithm in complex scenarios.The complex scenarios of multi-group proximity,splitting and merging shows that the two algorithms proposed can achieve accurate group segmentation and stable tracking.(2)Research on unresolvable group target tracking algorithm.For the unresolvable group,the idea of holistic tracking is adopted.The GM-TPHD group target tracking algorithm based on the improved quadratic partitioning model is proposed to address the problem that the existing algorithms cannot accurately update group target trajectories and low tracking accuracy because the number of targets is mis-estimated under the complex scenarios such as proximity,intersection,splitting and merging.Through the combination of GM-TPHD filtering algorithm and elliptic Random Hypersurface Model(RHM),real-time trajectories with high accuracy updating are realized while estimation state of group targets is achieved.At the same time,an improved quadratic partition model is proposed to solve the problem of mis-estimation and under-estimation of group target.In the measurement division step,whether the quadratic partition is needed is determined according to the diffusion pattern of measurement set,so as to solve the problem of low tracking accuracy caused by mis-estimation in complex scenarios.The simulation results verify that the proposed improved algorithm solves the problem of underestimation of the number of group targets in complex scenarios and improves the tracking accuracy of the algorithm compared to existing algorithms.