Algorithms And Hardware Implementation Of Particle Filters For Target Tracking

With recent increases in computational power, particle filters, also known as sequential Monte Carlo (SMC) methods, are receiving great interest in target tracking. This dissertation considers the algorithms and hardware implementation of particle filters for target tracking.Firstly, the resampling unit, regarded as a bottleneck of particle filter due to its sequential nature, is investigated. A compact threshold-based resampling algorithm and its architecture for efficient hardware implementation of particle filters are presented. By using a simple threshold-based scheme, this resampling algorithm can highly reduce the complexity of hardware implementation. Simulation results indicate that this algorithm has approximately equal performance with the traditional systematic resampling algorithm when the root mean square error (RMSE) and lost track are considered. The results conducted on FPGA platform establish the superiority of the proposed architecture in terms of memory efficiency, low power consumption as well as low latency.Secondly, hardware implementation of particle filter applied to bearings-only tracking (BOT) problem is studied. A simplified particle filter algorithm, which features lower computing power and hardware complexity, is proposed. Based on the proposed algorithm, this dissertation lays emphasis on the efficient hardware implementation of this algorithm on FPGA platform, and presents the hardware architecture of the system and implementation of its sub-modules in detail.The study shows that the implemented particle filter can be used to solve the BOT problem and has relatively fast processing rate.Additionally, in the BOT problem, the uncertainty of process model is smaller compared with the uncertainty of measurement and this will lead to severe sample impoverishment when using particle filter. In order to solve this problem, a novel roughening algorithm and its hardware architecture are proposed. The reasonable distribution of the roughening jitter is calculated from the innovation related to the resampled particles. Simulation results and experimental studies indicate that particle filter with this roughening method can be efficiently implemented in hardware and can effectively solve the BOT problem with fast processing rate and low complexity.When hardware implementations are considered, most of the multiple model particle filters (MMPFs) have a potential drawback:the number of particles used per model is time-varying.To solve this problem, an easy-hardware-implementation MMPF for maneuvering target tracking is proposed, where the sampling importance resampling filter is extended to incorporating multiple models that are composed of a constant velocity model and a "current" statistical model, and the prior model probabilities and transition probabilities are assumed constant. Thus, the implementation difficulty of predetermining sub-models for different maneuvering motions can be avoided and the proposed filter can keep a constant number of particles per model at all times.Thereby, it allows a regular hardware structure with deterministic execution time. In order to avoid the bottleneck introduced by the traditional systematic resampler, the Independent Metropolis Hasting (IMH) sampler is utilized in the resampling unit in each model to reduce the latency of the whole implementation. Hardware architecture of the IMH sampler and its corresponding sample unit are presented, and a parallel architecture of the system is described. The proposed architecture is evaluated on a Xilinx Virtex-â…¡Pro FPGA platform for a maneuvering target tracking application and the results show many advantages of the proposed MMPF over existing approaches in terms of efficiency, lower latency, and easy hardware implementation.Lastly, particle probability hypothesis density (PHD) filter for multiple targets tracking is studied. A novel multiple model PHD (MMPHD) filter for multiple maneuvering targets tracking is presented. This proposed filter shows similar tracking performance with the standard MMPHD without any knowledge of target models and model transition probabilities for different maneuvering motions, and has faster processing rate. Further more, the hardware implementation of the particle PHD filter is investigated.