Research On Multi-Target Tracking Combined With The Probability Hypothesis Density Filter

The technology of multi-target tracking is widely applied in military and civil fields such as space-based warning system,defense of ballistic missile,video surveillance system.The Probability Hypothesis Density(PHD)filter is a newly developed solution to the multi-target tracking problem.It is based on the Random Finite Set Theory and is capable of handling the uncertainty of target number.The main work and achievements of this paper are as follows:1.An updated-weight adjustment method for state extraction combined with the Gaussian Mixture PHD(GM-PHD)filter is proposed.The performance of the original GM-PHD filter drops down when it comes to the situation of closely spaced targets and high clutter rate,thus results to the inaccurate estimated state.Therefore,it is of necessity to modify the updated weight of Gaussian component such that “one measurement is generated from more than one target and one target can be associated with more than one measurement”.2.A method of multi-target state extraction for free clustering particle probability hypothesis density filter is proposed.The PHD function is approximated by a set of particles with weight and k-means is a usual adopted solution to estimate the target state from those particles.The particles are incorrectly classified using the k-means method if the targets getting too close to each other,inevitably,the estimated states are inaccurate.A new method to extract target states without the need of clustering is proposed.Firstly,the updating step of the P-PHD is decomposed.Secondly,the observation categories coming of the real-target are selected and the chosen observation categories are assigned with the same number of new particles with new weights respectively.Finally,the target state is extracted from those brand new particle clouds directly and there is no need to execute the clustering and peak extraction operation.3.Particle-aliasing-based method for track continuity combined with the probability hypothesis density filter is proposed.The particle probability hypothesis density(P-PHD)filter gives estimate of target state for multi-target tracking;however,it keeps no record of target identities and is not able to generate target tracks.The corresponding particle clouds originated from the same target at two successive time steps overlap each other largely.Targets' state information is included in particle cloud;there is a one-to-one correspondence between particle cloud and estimated state.Whether two states are associated is determined by analyzing the aliasing situation of the two corresponding particle clouds.Thus,estimated states at different time steps from the same target are associated to generate tracks step by step.