Research On Theory And Applications Of Gaussian Mixture PHD Filter

With the Multi-target tracking getting mature and developing,a lot of algorithms about multi-target tracking are coming out,and the solutions to multi-target tracking problems are abundant.According to the development of the theory for multi-target tracking,multiple-target tracking algorithm can be divided into two categories,namely the classical algorithms based on the data association,and the probability hypothesis density(PHD)filter algorithm based on the stochastic finite set.The latter use a random set to model,in order to avoid the computational complexity caused by data association.In this paper,the probability hypothesis density(PHD)filter method has been mainly studied.Firstly,the traditional multi-target tracking methods,such as Bayes filtering,Kalman filter,particle filter are introduced.These target tracking algorithms are only suitable for simple situations,so it cannot solve the real problem.The PHD filter algorithm is proposed by Mahler.R derived under strict mathematical framework,based on a solid mathematical theory,so it is easy to be implemented,and the estimation is more accurate.The algorithm promotes the development of multi-target tracking in this field.Before the introduction of the algorithm,some theories of the stochastic finite set are described,such as the definition of stochastic finite set,the set integral and the set differential,and the evaluation criterion of the effective tracking performance for Multiple target tracking algorithm based on random set are introduced in detail etc.,as the basis for the study below.In the second part of this paper,the Probability Hypothesis Density(PHD)filter algorithm and the Cardinality Probability Hypothesis Density(CPHD)based on a random set are introduced,and the recursions of the algorithm in Bayesian framework are elaborated.Both the implementation of particles filtering and Gaussian mixture filtering for PHD filter,and the realization of Gaussian mixture filtering for CPHD filter are studied in depth.The tracking performances of two algorithms are compared through simulation.Due to the Gaussian mixture implementation of PHD/CPHD filter are considered to be known a prior probabilities,newborn target Gaussian mixture components require to appears in the whole monitoring area,which is inefficient,and it restricts the engineering applications.The GM-PHD algorithm prones to make mistake tracking in the strong clutter.To solve this problem,a plug is added in the original algorithm which distinguishes between the persistent and the newborn targets in the prediction and the update step,the two steps of both targets are carried out respectively.Through the measurements received from each scan to obtain the new intensity function adaptively,so that the hypothesis about prior probability can be avoided according to the measurement-drive method.The algorithm is named Adaptive Gaussian Mixture probability hypothesis density(AGM-PHD)filter in this paper.Finally the GM-PHD filter algorithm and the AGM-PHD filter algorithm are both applied into dolphin whistle tracking.There are a number of challenges.The various aspects of the GM-PHD and AGM-PHD filter are discussed by setting the different parameters,to research the performance of tracking whistle signal,and the rules are explored through a lot of experiments.