Researches On Particle Filter Track-before-detect And Implementation On GPU

The techniques of detecting and tracking dim targets are widely used in manydifferent fields, such as radar warning, infrared detection and flight navigation. Thetrack-before-detect (TBD) technique is proposed on the basis of classicaltrack-after-detect (TAD). TBD uses the raw measurements of sensors directly withoutapplying a threshold and accumulates target energy by processing more consecutivemeasurements to estimate the target track when the detection is declared. TBD iscommonly used in nonlinear and non-Gaussian scenarios; moreover, TBD based onparticle filter has a good performance on detecting and tracking dim targets, so thetechnique of particle filter (PF) turns into a reasonable solution for related TBDproblems.Since TBD makes decisions by the raw measurement data, its computationalcomplexity increases extremely, and then limits its real-time implementations. In recentyears, the graphic processing unit (GPU) develops rapidly. Thus, with a hybridutilization of GPU and CPU, various parallel algorithms can be implemented moreefficiently on the parallel structure of the GPU. In this dissertation, the implementationof track-before-detect based on particle filter on GPU is researched and the maincontributions are summarized as follows:Firstly, the theory of particle filter track-before-detect based on recursive Bayesestimation have been studied. According to simulations, the dissertation analysis andverifying the two algorithms of particle filter TBD. What’s more, the models of infraredand radar on TBD are established.Secondly, several improved algorithms on particle filter are studied and keyresearches are carried out on TBD based on auxiliary particle filter. Simulation resultsshow that various numbers of particles and different size of likelihood ratio areas haveinfluence on the detection and tracking performance of the algorithm. Since there is noeffective evaluation criterion for TBD, a new set of evaluation criterion has beenproposed. Theoretical analysis and simulation results indicate that this method can makeeffective comparisons among different TBD algorithms. Thirdly, concerned with the bad performance on real-time implementations of TBDthat leaded by computational complexity, a novel implementation of TBD algorithm onGPU is proposed. The CUDA programs based on the C language optimize the parallelthreads in GPU, furthermore, the speed-up ratio is improved and the performance onreal-time implementation is better. At the same time, the solution to the application oflarger size of likelihood ratio area which exceeds the number of threads in one block isgiven. The dissertation analyzes the performance of two GPU cards, and the simulationresults indicate that the memory interface width and processing cores are the keyparameters which affect the speed of computation.