Real-Time Implementation Of The Probability Hypothesis Density Particle Filter

Supported by the National Science Foundation of China (No.61171149), this thesis has conducted research on probability hypothesis density particle filter for real-time multi-target tracking (MTT).First, the algorithms on MTT problem are reviewed. Based on the multi-target Bayesian model, two of the popular MTT algorithms, namely, the Probability Hypothesis Density (PHD) filter and the Multi-Bernoulli (MeMBer) filter, are introduced, according to two different random finite set theories. Specifically, the main theory and the implementation problems of the PHD are described, and the evaluation metrics of the MTT algorithms are discussed.Second, the time delay analysis is conducted on the weight update in the particle PHD filter, and this thesis proposes an efficient measurement-driven particle PHD filter for real-time MTT of nonlinear/non-Gaussian system in dense clutter environment. Since the measurement-driven scheme eliminates most clutters in the scene and classifies the remaining measurements into survival measurements and spontaneous birth measurements, better real-time performance can be achieved. Extensive simulations validate the improvement of both the real-time performance and tracking performance of the proposed measurement-driven particle PHD filter in comparison with the traditional particle PHD filter.Third, due to the sequential feature of the resampling algorithm, the resampling can not start until the weights of all the particles are updated, which limits the fully-parallel FPGA implementation of particle PHD filter. To overcome this difficulty, this thesis proposes a non-sequential resampling for the particle PHD filter, in which the particle weights are all set below a proper threshold. This specific threshold is determined using a distinguishing feature of the particle PHD filters:The weight sum of all particles in weight update is equal to the total target number in the current iteration. Theoretic analysis indicates that, in comparison with traditional systematic resampling (SR) based particle PHD filters, a particle PHD filter employing the proposed resampling can reduce the processing time by33%around in a typical MTT scenario, while simulation results show that it can maintain the same level of estimation accuracy.