Application Research Of Particle Filter For Maneuvering Target Tracking

Target tracking is a typical problem of dynamic system state estimation. Under linear Gaussian conditions, the Kalman filter is the optimal estimation. However, in the actual application, the moving target rarely meets a single model and a linear Gaussian conditions, especially in the strongly non-linear/non-Gaussian environment, the accuracy of the Kalman series filters will be decreased or even divergent, and can not satisfy the requirements. Particle filter has provided an approximate Bayesian solution for discrete-time recursive filtering, which can handle arbitrary non-linear/non-Gaussian system and has been concerned and researched for its great theoretical and practical significance.The key idea of particle filter is to represent the required posterior density function by a set of random samples with associated weights and to compute estimates based on these samples and weights. Sequential importance sampling (SIS) algorithm is one of the main algorithms of particle filter, but a common problem with the SIS particle filter is the degeneracy phenomenon. In addition, the Rao-Blackwellized particle filter is a powerful combination of the particle filter and the Kalman filter, which can be used when the underlying model contains a linear sub-structure, subject to Gaussian noise. But each particle runs a Kalman filter, it is hard for actual application because of the immense calculation.Under the background of multi-target tracking, a few major issues of the particle filter were analyzed in this thesis. In view of the above-mentioned two aspects, the resampling algorithm and Rao-Blackwellized particle filter were improved respectively. Based on the systematic re-sampling algorithm, the improved resampling algorithm replacing the random number generator process with specific values, thus reducing the cost of the algorithm; for the underlying model contains a linear sub-structure, subject to Gaussian noise, the improved Rao-Blackwellized particle filter use the Gaussian particle filter in place of the general particle filter algorithm, which does not require resampling step, not only reduced the computational cost, but also to avoid the affect of the resampling issues such as degeneracy of the samples. For the linear part, only one Kalman filter was used, rather than each particle run a Kalman filter, and reduced the computational complexity.The research results of this thesis show that the improved algorithms reduced the computational complexity while as accuracy as the original algorithm, and shorten the running time of the algorithm. Through the simulation and combination of the interactive multi-model approach to maneuvering target tracking, results show that the improved algorithm performance in line with the theoretical analysis, shorten the running time of the algorithm and has made the expected tracking results.