Multi-sensor Target Tracking Algorithm Based On Probability Hypothesis Density Filter Under Unknown Scene Parameters

In modern society,target tracking technology is widely used in navigation,guidance,transportation,monitoring and other fields.Therefore,it is of great significance on target tracking problems.With the development of sensor technology,multi-sensor system has become a hot point in multi-target tracking field because it has the following advantages over single sensor: higher reliability,greater coverage,higher credibility and robustness,etc.However,multi-sensor systems still have many problems to be solved in practical applications,such as excessive transmission cost and computational burden,reduce the influence of unreliable sensor in the system,and the selection of sensor network types,etc.Since the 21 st century,the Random Finite Set(RFS)theory under the Bayesian framework has received widely concerned by scholars because it avoids the combined explosion problem caused by data association in the traditional tracking method.In this paper,the multi-sensor technology under unknown scene parameters is studied deeply based on the RFS theory,and the main contributions of the dissertation are given as follows:1.Problem of the unknown clutter environment.The algorithm under the Bayesian framework usually assumes that the clutter intensity is known as a priori.If the preset parameters of the clutter are not accurate,or the clutter is dynamically changing.Thus the precision of target tracking will be greatly affected.In the iterated corrector cardinalized probability hypothesis density(IC-CPHD),this error will be amplified with the iteration proceeding,which will further lead to filter divergence and target lost.Based on the ICCPHD algorithm,this paper proposes an IC-CPHD algorithm for clutter estimation using Dirichlet Process Mixture Model(DPMM).Firstly,the measurements obtained by all sensors is used to form the sample set.Then the DPMM non-parameter cluster method is adopted to estimate the clutter model flexibly and caculate the spatial distribution parameters.Finally,the sample set for clutter eatimation is corrected by the filter estimation of IC-CPHD.The simulation results show that the proposed algorithm can effectively estimation the clutter distribution,and can suppress the influence of the measurements generated by the real target for clutter estimation.2.Problems with unknown birth target RFS.The unknown birth RFS is always modeled by the historical measurements.This method driven by mesurements has two main problems.On one hand,there is a certain delay when new targets appear.On the other hand,the measurements with many clutters will affect the precision of birth RFS and lead to large computational burden.For multi-sensor systems,the method driven by mesurements will results in interference items around the real target,which may affect the track association.Therefore,it is necessary to divide the measurement set.This paper proposes a measurement set division algorithm based on IC-PHD filter.The algorithm first uses a single sensor measurement set to model the PHD of RFS with labels,and the measurement set is divided by iterative filter method.The measurement sets generated by the new target is used to model the birth target PHD.At the same time,the measurements of clutter can be removed from the measurement set.Thus,the predicted PHD can be updated with the measurement set without clutters.The simulation results show that the proposed algorithm can effectively model the new RFS,improve the target tracking accuracy and reduce the computational burden.3.Problem of unknown measurement noise.The measurement noise is usually assumed as a zero-mean Gaussian white noise.However,sensors in one system may have unknown translation error,and unknown measurement noise covariance,which means the measurements obtained by different sensors for the same target may have large difference.If the preset parameters of the measurement noise is not accurate,it will lead to a decline of target weight or even target loss during the iterative filter.In order to solve this problem,this paper proposes a label IC-PHD filter which can estimate the mean and covariance of the measurement noise simultaneously.The variational Bayesian method is used to decouple the measurement noise and the multi-target states in the likelihood function,and the translation error is approximated by the difference between the sensors.The simulation results show that the proposed algorithm can correctly estimate the parameters of unknown non-zero mean measurement noise and obtain accurate fusion track.4.Problem of unknown sensor reliability.Although iterative filter has the advantages of easy expansion and low computational complexity,it still has many problems,such as high transmission bandwidth,the order related results.In addition,if there is a poor sensor in the multi-sensor system,the accuracy and reliability of the iterative filter will be greatly reduced.In order to solve these problems,this paper proposes a distributed multi-sensor fusion algorithm based on D-S(Dempster-Shafer)evidence theory.Firstly,the local sensor is filtered by the label VB-PHD.In this step the track information and the measurement noise covariance are estimated simultaneously.Then,the D-S inference is used to determine the track-to-track correspondence.Finally,based on the estimated noise covariance,a track quality parameteris calculated as the weight of the state fusion.The simulation results fully prove that the proposed algorithm can obtain accurate fusion results in complex scenarios such as low detection probability,high measurement noise and target intersection,etc..The above four parts constitute a systematic series of multi-sensor multi-target tracking algorithm under unknown parameters of scene,which provides theoretical basis and technical support for solving multi-sensor system architecture and algorithm selection in complex environments.