Data association techniques for bearings-only multi-target tracking using simulated annealing and implemented with Boltzmann machines.

摘要

In this dissertation, algorithms and parallel hardware architectures for passive estimation of the initial states of multiple targets are developed. The measurements are assumed to be bearing estimates from acoustic array preprocessors. The maximum likelihood principle is used to formulate the problem with multiple targets and clutter points considered at the same time. However, a traditional optimization algorithm is frequently trapped in a local maximum. The "simulated annealing" method is then exploited for locating the global maximum of this multimodal likelihood function.; A fast simulated annealing method, well-tailored for this particular optimization problem, is developed for targets in both clean and dense clutter environments. The solution space is decomposed into the combinatorial (data associations) and the continuous (target states) parts. Simulated annealing is applied to search the combinatorial part only, and nonlinear programming searches the continuous state space. This combinatorial space is further decomposed into sk independent subspaces, with s the number of sensors and k the number of scans, such that simulated annealing, performed in each subspace, has only moderate computational complexity. The cost surface of the likelihood function is analyzed experimentally. Monte Carlo simulation results for this algorithm are compared with theoretical lower bounds in order to measure the performance and adjust the system parameters.; The decoupling of the data association and state estimation problems, and the recent progress in analog VLSI technology make the proposed method suitable for implementation in hybrid analog/digital hardware. The nonlinear estimation is performed by a traditional digital processor while the data association problem is solved by parallel neural network structures, Boltzmann machines, or diffusion networks. Due to the fact that a Boltzmann machine converges to the global optimum only when each neuron is sequentially fired, a diffusion network is developed to implement the Boltzmann machine in full parallelism. The diffusion network consists of N by M interconnected stochastic neurons, each characterized by a stochastic differential equation, the Langevin equation, and is shown to be suitable for analog VLSI implementation.