Multi-target Tracking Algorithms In Unknown Scenarios Based On Multi-bernoulli Filter

With the development of the computer and the sensor such as radar,sonar,infrared,electro-optical,etc.,multi-target tracking(MTT)has been widely applied in diverse fields such as air traffic control,surveillance,missile defense,autonomous vehicles,computer vision,biomedicine,oceanography,etc.In unknown and dynamic scenarios,how to maintain the validity of the MTT algorithms is one of the most important and difficult tasks.Traditional Bayesian MTT filtering algorithms are mainly based on the thought of data association,i.e.,making hypothesized associations between the target and the measurement.When the number of targets and measurements is large,the association complexity increases dramatically,seriously influencing the applicability of the filtering algorithms.The random finite set(RFS)theory can effectively solve the computing intractability induced by the data association of traditional filtering algorithms,which makes the MTT methods possible to deeply investigate the tracking problems in unknown and varying scenarios.Based on the Bayesian RFS tracking algorithms,the research of this dissertation puts its emphasis on modeling the parameters in unknown target birth background,unknown maneuvering scenarios,unknown clutter background and unknown detection probability background.The main contents and contributions of the dissertation are listed as follows:1.The target birth intensity modeling problem with unknown birth information has been studied.The unknown target birth density models the possible target appearing regions by using the measurement information of current time step,which extends the algorithms adaptable to unknown target birth cases.Nevertheless,the accuracy of the birth model is decreased due to the coarse assignment of the expectation number of target birth.Therefore,under the cardinality balanced multi-target multi-Bernoulli(CBMe MBer)filtering framework,this dissertation proposes an improved unknown target birth density tracking algorithm.The algorithm constructs an assignment function by using current prediction information of surviving targets,which makes a more reasonable allocation of the target birth expectation number and thus improves the accuracy of the unknown birth model.The simulation results indicate that the improved algorithm can achieve favorable tracking accuracy under unknown target birth background.2.The target birth probability modeling problem and the real-time property degrading problem induced by birth model have been studied.In the existing birth models,the priori known and fixed target birth expectation number is far from the truth,which makes it is impossible to get accurate target birth probabilities.In order to solve this problem,a pre-processing step is proposed,which models the target birth probability more accurate by using the known information of current time step.Additionally,compared with the traditional birth model,the modeling process of unknown birth model is more complex,which includes more birth components and induces large number of likelihood calculations in the update step,resulting a decrease in real-time performance of the algorithms.Therefore,a measurement noise based threshold method is proposed,which greatly improves the operation efficiency of the algorithms by reducing large number of useless measurement likelihoods.The simulation results indicate that the proposed filter can better capture new-born targets and maintain favorable tracking accuracy and real-time property.3.The unknown parameter modeling problem of target motion model in maneuvering scenarios has been studied.When the parameter of the target motion model abruptly changes,the interacting multiple model(IMM)MTT algorithms are hard to match the target maneuvering,unless the target motion model set is expanded to cover all possible abrupt parameter changing.However,such operation is inefficient.To solve this problem,this dissertation proposes an unknown parameter estimation method based on Liu and West(LW)filter,which can adapt the abruptly changing parameter using the particles of pervious time step.Then this method is applied into the CBMe MBer filter to estimate the motion model in real time,so as to match the actual target maneuvering.The simulation results indicate that the unknown motion model parameter filtering algorithm has better applicability and tracking accuracy in unknown maneuvering scenarios.4.The problem of unbalanced cardinality allocation in unknown clutter rate cardinalized probability hypothesis density(CPHD)filter and the problem of detection probability estimation delay in unknown detection probability filter have been studied.Due to the special structure of the partial fraction in the legacy tracks,the unknown clutter rate CPHD filter will lead to a prediction mixture when processing the “pseudo targets” and the true targets,which causes an unbalanced cardinality allocation between the clutter and the target.Since the legacy tracks in the CBMe MBer update step only contains the prediction function and the un-detection probability but not contains the special partial fraction which leads to the prediction mixture,the unbalanced cardinality allocation problem can be avoided by implementing the clutter processing method under the CBMe MBer filtering framework.The unknown detection probability filter adopts Beta distribution to fit the detection probability whose current estimation is not used in the current filtering operation,which results an estimation delay.To avoid this,the joint estimating and filtering algorithm is divided into two steps.First,the real-time estimation of the detection probability is obtained by using the Beta distribution and the current information.Then,the estimated detection probability is applied into the formal filtering operation to update the target states.The simulation results indicate that the unknown clutter rate CBMe MBer filter corrects the unbalanced cardinality allocation,which gives favorable tracking accuracy,and the improved unknown detection probability filter corrects the detection probability estimation delay,which obviously improves the detection probability estimation and the target tracking accuracy especially in the information accumulation stage.The above four sections complement each other and provide new views and technical supports for the multi-target tracking in complex scenarios.