| The techniques of passive multi-sensor multi-target tracking are important topicsin multi-sensor data fusion systems. Because of the wide applications in both militaryand civil areas, much attention has been paid to their developments by worldwilderesearchers and engineers. Sponsored by the National Natural Science Foundation ofChina, the dissertation mainly investigates the applications of the random finite settheory in the field of passive multi-sensor multi-target tracking. The main contributionsof the dissertation are as follows:1. To construct a proper cost function for the multi-sensor data associationproblem, a modified cost function is proposed based on the distance weighted leastsquares, which takes the distance information and the fusion covariance into account,with a great improvement of correlation accuracy. Then, to improve the computationalefficiency of the Lagrangian relaxation algorithm, a fast relaxation algorithm isproposed, which directly picks out a part of correct pairs by two statistic tests withoutredundant relaxation and enforcement processes. Finally, the modified cost function isemployed into the fast relaxation algorithm, which efficiently improves the correlationaccuracy and computational efficiency.2. For the problem of scale unbalance in the product multi-sensor PHD filter, animproved algorithm is proposed, which calculates the joint likelihood function in theproduct form while the scale factor in the summation form, respectively. Then, to solvethe particle degeneration problem in the multiple-model PHD filter for trackingmaneuvering targets, an interacting multiple-model PHD filter is proposed, whichapproximates the model conditional PHD of target states by particles, and makes theinteraction of survival targets by resampling. Finally, the idea of Rao-Blackwellized isintroduced into the improved algorithm to further enhance the performance for jumpMarkov systems with mixed linear/nonlinear state space models.3. For the problem of missed detection in the CPHD filter, an improved algorithmis proposed, which minimizes the effect of the weight shifting and subsequentestimation errors by a dynamic reweighting scheme after pruning and merging. Then,to solve the weight over estimation problem in the GMP-CPHD filter, a novel CPHDalgorithm based on nonlinear Gaussian filter is proposed, which propagates the meansand covariances of the Gaussian components by a group of quadrature points, and improves the tracking accuracy and computational efficiency.4. For the problems of target number over estimation in the MeMBer filter and themeasurement innovation weakening in the CBMeMBer filter, respectively, anIMeMBer filter is proposed by modifying the legacy rather than the measurementupdated tracks parameters, which can solve both the problems effectively. Then, toprovide a closed-form solution to the nonlinear problem occurred in the passivebearings-only multi-target tracking system, a set of Gaussian particles is employed toapproximate the distributions of the multi-Bernoulli RFS.5. For the problem of track management in the RFS based filters, an improvedestimate-to-track association algorithm is proposed. Firstly, a multi-step prediction ofcurrent target states is made, and the weighted labels are assigned to them according tothe inertia. Secondly, the fuzzy membership degrees of the predicted state estimatesbelonging to the current state estimates are obtained by utilizing the maximum entropyfuzzy clustering. Different from the traditional methods, the proposed algorithm doesnot update the track information by simply summing the log likelihood ratios betweenadjacent frames, but takes the entire multi-frame information into account, thus anexcellent performance of track continuity. |
PDF Full Text Request