Research On Particle Filter Based Track-before-detect Algorithms For Weak Targets

Track-before-detect (TBD) is an important method for the detection and tracking of weak targets. Different from classical track-after-detect (TAD) processing, TBD directly uses unthresholded or low thresholded measurements of sensors for adequate information utilization. It gains the increasing of signal to noise ratio (SNR) by temporal cumulation of measurements and realizes the detection and tracking (track extracting) of weak targets simultaneously. Since TBD usually behaves as a hard nonlinear problem, the particle filter (PF) technique becomes a reasonable solution. In this dissertation, the particle filter based track-before-detect (PF-TBD) algorithms are investigated and the research conclusions are summarized as follows:Firstly, the models of TBD processing for radar and infrared (IR) are established. The uniform framework and principle for PF-TBD are studied on the basis of these models. Following the principle, a simple PF-TBD algorithm is presented. The simulation results prove the feasibility of using particle filter to realize TBD processing. Besides, the outlines of implementing PF-TBD algorithm for multiple weak targets are also discussed.Secondly, a SPRT-FSS likelihood ratio united test algorithm is proposed, which combines sequential probability ratio test (SPRT) with fixed sample size (FSS) likelihood ratio test. The employment of SPRT-FSS likelihood ratio united test and particle filter makes the realization of TBD processing for weak targets more efficient. Moreover, an autonomous multiple model (AMM) based particle filter algorithm is developed, which resolves the TBD processing for maneuvering weak targets via integrating the SPRT-FSS likelihood ratio united test.Thirdly, considering the deployment of multiple sensors, the thesis develops the multiple sensors distributed fusion based PF-TBD algorithms in the following two routes:⑴improving the fusion algorithm of particles'states. A predigested formula of calculating fused particles'weights is deduced. And the Gibbs sampler is also adopted to estimate the correspondence between particles'states of sensor nodes;⑵presenting the density fusion algorithm of estimated PDFs between sensor nodes. Kernel density estimate (KDE) technique is used to compute the conditional PDF based on the particles set at each sensor node. Density fusion method fuses these conditional PDFs to get fused particles set. In this thesis, it is proved that the fused particles'weights in aforementioned two fusion algorithms both satisfy the condition of SPRT-FSS likelihood ratio united test. Comparing with single sensor deployment, the proposed distributed PF-TBD algorithms not only reduce the time delay of detection, but also improve the precision of estimation to a certain extent.At last, the gradient information is introduced for the optimization of PF-TBD algorithms. A dual gradient particle filter algorithm is presented for the improvement of PF, which is used for realization of TBD. In order to decrease number of particles for detection and tracking, this algorithm uses gradient information to modify the transfer of particles'states. Two kinds of gradient information are adopted in this algorithm:⑴the gradient information of measurement model;⑵the gradient information of posterior PDF. The experiments indicate that the use of gradient information can ensure the detection performance and estimation precision, while markedly reduces the number of particles.