| Summary: | The central focus of our work is on designing particle filters which use more efficiently their particles by seeding them instate space areas with greater significance and/or by varying their number. We begin by introducing the auxiliary local linearization particle filter (ALLPF) whose importance sampling density brings together the auxiliary sequential importance resampling technique and the local linearization particle filter (LLPF). A simulation study assesses it suitability for tracking manoeuvring targets. We next incorporate the prediction mechanism of the LLPF within a multi-target algorithm. The proposed particle filter (A-MLLPF) performs simultaneously the functions of measurement-to-track assignment and particle prediction while employing an adaptive number of prediction particles. Compared to the equivalent standard multi-target particle filter, we show that the A-MLLPF performs better both in terms of tracking accuracy and measurement association. The remaining of the thesis is devoted to vehicle tracking which exploits information from the road map. We first focus on the variable-structure multiple model particle filter (VSMMPF) from the literature and we enhance it with a varying particle scheme for using adaptively fewer particles when the vehicle travels on the road. Simulation results show that the proposed variation results in a similar performance but with a significant decrease of the particle usage. We then incorporate a gating and a joint probabilistic data association logic into the VSMMPE and use the resulting algorithm (MGTPF) to track multiple vehicles. Simulations demonstrate the suitability of the MGTPF in the multiple vehicle environment and quantify the performance improvement compared to a standard particle filter with an analogous association logic. Returning to the single-vehicle tracking problem, we introduce lastly the variable mass particle filter (VMPF). The VMPF uses a varying number of particles which allocates efficiently to its propagation modes according to the modes’ likelihood and difficulty. For compensating for the resulting statistical irregularities, it assigns to the particles appropriate masses which scale their weights. Other novel features of the proposed algorithm include an on-road propagation mechanism which uses just one particle and a technique for dealing with random road departure angles. Simulation results demonstrate the improved efficiency of the VMPF, since it requires in general fewer particles than the VSMMPF for achieving a better estimation accuracy.
|
|---|
