Research On Continuous Discrete Poisson Multi-Bernoulli Mixture Multiple Extended Target Tracking Algorithm

In recent years,the application of multi-target tracking technology in portrait recognition and vehicle automatic driving has made rapid progress.In many multi-target tracking frameworks,the filter based on random finite set avoids the data association problem in the traditional method,and has been widely used due to its high tracking accuracy and stable performance.Among them,the Poisson Multi-Bernoulli Mixture(PMBM)filter can achieve high tracking accuracy in complex tracking scenes,and has attracted more and more attention.Based on the continuous discrete Poisson Multi-Bernoulli Mixture(CD-PMBM)filter in random finite sets theory,this thesis studies from two aspects: the extended target shape modeling and the extended target tracking method in a complex noise environment.The specific research content is as follows:Firstly,for the tracking scenario where the time of measurements obtained is non-uniform,we study the extended target tracking algorithm in the CD-PMBM filter framework.The basic theory of the CD-PMBM filter is introduced in detail,and its filtering algorithm is given.The filter realizes prediction and update on non-uniform time steps by establishing a continuous discrete model.In addition,the random matrix model is embedded in the CDPMBM algorithm,and the CD-PMBM filtering algorithm for elliptical targets is proposed.For the star-convex shape target,the EM algorithm is also introduced to estimate the shape based on the CD-PMBM filter.Simulation results show that in the case of non-uniform measuring time,the proposed algorithm can effectively estimate the state and shape of the extended target.Then,for the case of measurement noise and process noise with outliers,the student t mixture term is used to replace the traditional Gaussian mixture term,and an advanced extended target filter based on Student's t mixture continuous discrete poisson multi-bernoulli mixture is proposed.Due to certain properties of the Student's t distribution for better handling of heavy-tailed noise effect,which is combined with the CD-PMBM conventional filter solves the problem of noise in the presence of outlier performance degradation.Considering the unknown clutter,we introduce the finite mixture model and propose a continuous discrete poisson multi-bernoulli mixture filter based on the finite mixture model.The filter introduces a finite mixture model to describe the clutter,the EM algorithm and the Gibbs sampling algorithm is used to estimate the clutter parameters.It can greatly reduce the false alarms generated in the dense and uniform background through the clutter estimation algorithm,thereby improving the performance of the CD-PMBM filter.Simulation experiments show that the proposed algorithm can achieve effective and stable tracking of the target in the corresponding complex environment.At last,when the measurement noise covariance is unknown,we introduce the variational Bayes algorithm,and propose a new CD-PMBM filter based on a variational Bayes.The filter defines the noise covariance as a state variable,and it performs a variational Bayes approximation of the joint probability distribution composed of the measurement noise covariance and the extended target state,and uses the inverse Wishart distribution to represent the measurement noise covariance distribution.The simulation results show the effectiveness of the proposed algorithms in tracking multi-extended targets in unknown measurement noise covariance scene.