Random Finite Set Multi-target Tracking Algorithms Based On Box Particle Filtering

Multi-target tracking(MTT)is the online estimation of target number,states,tracks through the processing of measurements,which are conventionally modeled as points in the measurement space.The use of wireless sensor networks(WSN)is increasing driven by the development of very-large-scale integration,microelectromechanical systems,and wireless communication technology.During data transmission in wireless sensor networks,measurements are generally quantized in order to save communication bandwidth.The quantized data are intrinsically multi-dimensional intervals in the measurement space and are referred to as interval measurements or box measurements.Box measurements are a kind of unconventional measurements and are of great interest to researchers in the multi-target tracking area.In recent years,particle filtering(PF),box particle filtering(BPF)and random finite set(RFS)theory emerge and are provoking fresh thought for the processing of box measurements by multi-target tracking systems.Because of the high degree of uncertainty involved in box measurements,particle filtering requires a large number of particles when processing box measurements to yield satisfactory results.This leads to a heavy computational burden on the multi-target tracking system and may even be detrimental to the system's quality of being operating in real-time.However,when box measurements are processed by box particle filtering,much fewer box particles are required,hence smaller computational burden.This dissertation is on a few multi-target tracking problems in the presence of box measurements,which are solved by box particle implementations of random finite set based multi-target tracking algorithms.What follow are the major contributions of the dissertation:1.Given box measurements,conventional state estimation performance criteria,inclusion and volume,are only defined for single target tracking and are not suitable for multi-target tracking.To solve this problem,this dissertation first redefines the inclusion and volume criteria for multi-target tracking and then presents formulas to compute the two criteria.The proposed criteria can conclusively indicate the multi-target tracking state estimation performance.Moreover,the labeled multi-Bernoulli filter(LMB)involves integrals that have no closed-form solution.This problem is solved in this paper by approximating a target state probability distribution by a set of weighted box particles,and a box particle implementation of the LMB filter is derived.The box particle LMB algorithm is well suited for the applications that involve box measurements and has a very light computational burden.2.The conventional box particle filtering resampling procedure may fail in the situation where the target state vector has elements that are hidden for sensors.To solve this problem,an improved resampling procedure is proposed.A new parameter called box resolution vector is defined in the improved procedure.This parameter is application dependent and user-specified.It measures the uncertainty represented by the interval in each dimension of a box particle and sets a limit to the sizes of box particles after resampling.Box particle filtering based on the improved procedure is more reliable and is able to deal with the situation where the conventional resampling procedure may fail.Moreover,in order to deal with the problem that the joint prediction and update generalized labeled multi-Bernoulli(JGLMB)filter has no closed-form solution due to complex integrals,a box particle implementation of the JGLMB filter is proposed.The box particle JGLMB algorithm has good performance and short run time.3.For the multi-target tracking problem in the presence of multiple sensors returning box measurements,a heuristic multi-sensor iterated measurement contraction procedure for box particle filtering is proposed.Measurement sets from the sensors are processed sequentially.After the likelihoods of the measurements from the sensor currently being processed are computed,a set of contracted versions of box particles are generated.Then likelihoods of the measurements from the next sensor are evaluated at these contracted versions.Moreover,the contracted versions resulted from the current iteration are contracted again by the measurements from the next sensor and result in a new set of contracted box particles.A box particle implementation of the multi-sensor JGLMB(MS-JGLMB)is developed based on the proposed iterated measurement contraction procedure.The box particle MS-JGLMB requires less run time than the particle MS-JGLMB does when reaching a similar level of performance.4.For the multi-target tracking problem in the presence of multiple sensors,some of which returning box measurements and the others returning point measurements,two approaches are proposed.One approach of the two first transforms the point measurements into box measurements using their supports and then process the box measurements with the likelihood function of a box measurement given a box particle.The other approach derives the formula of the likelihood of a point measurement given a box particle.Such likelihood and the likelihood of a box measurement given a box particle are employed to process point measurements and box measurements,respectively.The second approach avoids the contraction of box particles when point measurements are processed by the derived likelihood function and hence reduces the number of contracted versions of box particles.Simulation results show that both approaches can deal with the situation and output accurate estimates.Compared to the first approach,the second one is computationally cheaper and less accurate,though.