Research On Key Technologies Of Target Tracking Based On Particle Flow And Random Finite Set

Target tracking is a challenging research field emerging from various disciplines such as image/signal processing,biomedical engineering,and econometrics.The core of the problem lies in combining observations and prior knowledge to provide a reliable,accurate,and timely estimate of state and continuous trajectories.Early research has focused on the single-target tracking problem,where the main challenge is to accurately estimate the target state.Compared with single-target tracking,multi-target tracking(MTT)is more complicated.The essence of MTT is to estimate the state and trajectories of an unknown and time-varying number of targets in a tracking scenario based on a series of observations.In addition to the random variation of the number of targets over time,there are problems such as noise contamination of the measurements,unknown sources,missed detections,and false alarms.These factors lead to the great challenge of MTT.In recent years,the random finite set(RFS)theory has opened up new research directions for the MTT problem and has shown excellent tracking performance.However,a series of Bayesian filtering methods in the RFS framework still suffer from performance degradation,estimation instability,and low computational efficiency and other problems.This paper introduces particle flow into the RFS framework and focuses on the problems of target state estimation and trajectory estimation based on the theory of homotopy sampling.The main research contents are as follows:1.Aiming at the large computation burden and assumption of differential measurement caused by particle migration operation in particle flow filter,this dissertation proposes two improved particle flow filter methods: statistical linear regression(SLR)particle flow filter and marginal particle flow filter(MPF).The purpose of MPF is to improve the efficiency of the original particle flow filter through state decomposition.The purpose of SLR particle flow is to approximate the non derivative measurement equation into linear equation,so as to avoid direct derivation operation and expand the application scope of original particle flow filter.2.Aiming at the problem of performance degradation problem of probability hypothesis density(PFD)filter under the nonlinear multi-target tracking environment,this dissertation proposes a particle flow PHD filter.The particle flow PHD filter utilizes the Gaussian mixture to approximate the posterior probability density.This method uses the local cumulative distribution sampling method to sample particles from each Gaussian term.These particles are driven to the high likelihood region by particle flow,which can more accurately simulate the posterior probability density of multi-target state,and show better estimation accuracy than the traditional filtering methods.3.Aiming at the problem of performance degradation of Cardinality Balanced Multi-Bernoulli(CBMe MBer)filter in high clutter environment,this dissertation proposes a new Bayes filter based on optimal transport theory,namely optimal transport filter(OTF).The OTF views the flow operation from prior distribution to posterior distribution as an optimization problem.Besides,this dissertation also proposes a OT-CBMe MBer filter by integrating the OTF into the CBMe MBer filter framework for the multi-target estimation problem.The OT-CBMe MBer filter redefines the measurement sampling and Gaussian component sampling technology according to the posterior multi-Bernoulli density.As a result,the OT-CBMe MBer filter can use the OTF to guide the sample to the high likelihood region according to the selected measurements and Gaussian term.Therefore,we can get more accurate posterior probability density of multi-target state.4.This dissertation proposes a Gaussian particle flow importance sampling labeled Bernoulli filter(GPFIS-LMB)to estimate the trajectories of targets.In order to avoid weight degeneracy,we employ the particle flow to produce a measurement driven importance sampling distribution in the LMB filter.This method can alleviate the problem that the target can not be confirmed in time when the prior uncertainty is very large.At the same time,we also propose a data association method based on deep reinforcement learning to reduce the computation burden of LMB filter.The method extends the data association problem to the image pixel classification problem.Then,we employ the convolutional neural network to learn the optimal solution for the data association.This method can reduce the time complexity of data association and improve the speed of the LMB filter.