The Algorithm Improvement And Simulations Of Particle Filter Applied To Maneuvering Target Tracking And Its Architecture Design For Hardware Implementation

This dissertation presents two improved algorithms of particle filter used to solve the problem of tracking target occasionally hidden in the blind Doppler. The blind Doppler is a highly nonlinear problem, on which particle filter is more advantageous than the Kalman filter. The first improved algorithm sets two levels of noise during the particle generation process (a.k.a sampling), where one is fit for constant velocity motion, while the other is fit for motion with acceleration. Estimation of the current noise is based on the sum of particle weights, and we set a threshold for the sum, that is if it is larger than the threshold, we consider the noise level right and the target is moving with constant velocity; if it is less, we consider the noise level wrong and the target is moving with larger noise, maybe an acceleration. If the noise level is wrong, we adjust the noise to another noise level and sample the particles again, and also calculate the weights again. The second improved algorithm is the interacting multiple model particle filter, which uses two motion models, one is the constant velocity model; the other is the acceleration model. The estimation of current model is an estimation of target motion type. The advantage of the interacting multiple model particle filter is that if the model is close to the truth, the tracking performance is theoretically better than single model method. Nevertheless, it has much higher complexity and implementation difficulty. To verify the improvements, we conduct Matlab simulations to compare the improved algorithms with the extended Kalman filter and the general particle filter method proposed in literature. The second part of this dissertation is the design of the architecture of the interacting multiple model particle filter. The architecture we propose can be used in hardware implementation. We specially design for the parts that are most different with the general particle filter.