Based On The Improved Particle Filter Tracking Algorithm Before Detection

In recent years, small target detection and tracking in the field of infrared and radaris taken more and more attention, and the track after detect method can’t meet the caseof low SNR, many kinds of track before detect algorithms have been pronounced tosolve this problem, including the particle filter track before detect.Track before detect based on particle filter have a superior performance in radarweak target detection and tracking. But the particle filter still has some problems andlimitations, which will affect the performance of PF-TBD. This paper focuses on theimprovements of TBD algorithm based on particle filter. In single target case, researchon improved TBD algorithm including three aspects: re-sampling, the likelihoodfunction and multi-model. In the multi-targets case, describes the probability hypothesisdensity filter algorithm, analyzes and discusses the TBD implement based on thisalgorithm.Firstly, the resample methods of basic particle filter lead to lacking of particlediversity, which also make it difficult to guarantee the performance of TBD. This paperproposes a resample method based on particle optimization. The algorithm useobservations by Extended Kalman Filter in the resample process, so that the probabilitydensity function characterized by the particles is closer to the true posterior probabilityfunction, and then split and copy the particles by the method like sigma sampling, toensure the diversity of the particles and improve the performance of the TBD algorithm.Secondly, the likelihood function in the PF-TBD algorithm is closely related to theprobability density function of the measurement. The performance of the PF-TBDalgorithm can’t be guaranteed if the PDF of measurement is unknown. This paperpropose a convolution particle filter，while the likelihood function is constructing andoptimizing, the resample method samples particles from continuous distribution ratherthan discrete distribution. And this will improve the PF-TBD tracking performance androbustness.Thirdly, according to weak motor target, the model used in particle filter TBD can’tmatch the real model completely, which decrease the performance of the algorithm. This article discussed the interactive multi model algorithm and its implementation. And addthe IMM algorithm to the convolution particle filter TBD, which improved the modelmatching between the algorithm and the real target, thereby improving the performanceof the TBD algorithm.Finally, it is hard to implement the TBD algorithm by PF in the case of multi-target,particularly the number of targets is changing with time. This article discussed theprobability hypothesis density filter based on random finite set theory, and thenproposed a new method based on PHD to extract the target state, which improve theaccuracy of estimating the number of target in multi-target environment as well as targettracking accuracy.