Research On Multitarget Tracking Algorithm Based On CHPD Filtering

PHD filter and CPHD filter represent the problem of multi-target tracking byrandom finite set under the framework of Bayesian filter, which changes the way thattraditional multi-target tracking decomposes multi-target into multiple targets, can dealwith the situation of target survival, birth, division and avoids complex data association.PHD filter and CPHD filter propagate a first-order moment of posterior probabilitydistribution rather than the posterior probability distribution in traditional Bayesian filter.This reduces computational effort greatly, so that PHD filter and CPHD filter become anew hot in the field of multi-target tracking research. Compared with PHD filter, CPHDfilter overcomes the disadvantage of incorrect estimation of target number of PHD filterin low signal-to-noise ratio (SNR) conditions, but increases computation complexity.This dissertation mainly studies the methods of state estimation, multi-sensor statefusion and track continuity in CPHD filter.1）State estimation of CPHD filter: In the general CPHD filter, the extraction oftarget state needs the help of the methods such as clustering. However, the STPHD filterwhich is proposed based on PHD filter transforms the PHD into the sum ofdecomposed-PHD components corresponding with observations, so we only needs toextract maximum peaks of decomposed-PHD components based on estimated targetnumber without external methods. As CPHD filter is quite similar to PHD filter in therespects of PHD prediction and update, we can easily extend STPHD filter to CPHDfilter by the correspondence function, which is called QSTCPHD filter. This dissertationuses particle filter to realize single-sensor CPHD filter and uses QSTCPHD filter toextract target states, which makes the state extraction of CPHD filter as simple andaccurate as STPHD filter.2）Multi-sensor state fusion of CPHD filter: To improve the tracking accuracy ofCPHD filter, we can extend it to a multi-sensor environment and monitor observationregion by multiple sensors. However, CPHD filter bases on random finite set theory, soit regards the targets at each frame as a random and unordered set, and can not give anytarget identities. That is to say, at each frame, although the target states estimated byCPHD filter are estimated states of true targets, we don’t know each estimated state isfrom which target. This means that if we want to realize multi-sensor CPHD filter, weneed a track association algorithm to identify the multi-sensor states belonging to thesame target, in order to fuse the states coming from the same target. This dissertationtakes the concept of intuitionistic fuzzy set into account and proposes a new trackassociation method based on intuitionistic fuzzy set clustering. This method firstintegrates the estimated target states and analyses the relationship among them especially the fuzzy correlation, then transforms the estimated states into intuitionisticfuzzy sets, finally constructs the correlation between any two states by the process ofclustering and takes this correlation as an important basis of track association. Thismethod not only extends CPHD filter to the multi-sensor aspect, but improves trackingeffect significantly.3）Track continuity: In a multi-target tracking system, it is far from enough for usto obtain target states at each frame, we have to know motion track of each target.However, without giving target identities, CPHD filter cannot obtain the statecorrelation among frames, which leads to the failure of construction of targettrajectories. This dissertation utilizes the particle weight vector in particle QSTCPHDfilter and target motion rules to construct particle weight correlation matrix and stateprediction correlation matrix in adjacent time and establish a “birth—>survival—>death”trajectory for each target with the leading of particle weight correlation matrix andassistance of state prediction correlation matrix. As so often, we build up a completemulti-target tracking system for CPHD filter, including multi-sensor fusion and trackcontinuity, which makes the theory system of CPHD filter more thorough.