Distributed Particle Filter Based Speaker Tracking In Distributed Microphone Networks

The problem of speaker tracking with microphone arrays aims at locating a moving s-peaker based on the signals acquired by multiple spatially distributed microphones. Speaker tracking with microphone arrays has played an important role in many speech scenarios, includ-ing surveillance systems, audio/video conferences, automatic speech recognition, hands-free systems and robotics. Nowadays, advances in wireless sensor networks, communication, mo-bile computing and integrated circuit have motivated the deployment of distributed microphone networks (DMNs), or so called wireless acoustic sensor networks. Correspondingly, speaker localization and tracking in DMNs has become a hot issue. So far, most existing approaches are based on traditional regular arrays, whereas they cannot work well in DMNs.The Bayesian filter has been widely used for speaker tracking in noisy and reverberant en-vironments, and it models speaker tracking with the state-space approach. When the state-space model is linear and Gaussian, Kalman filters can provide the optimal solution for Bayesian filter; for the nonlinear and non-Gaussian cases, particle filters have been proved to work well. The problem of speaker tracking in indoor environments is usually nonlinear and non-Gaussian. To address such problem, the current distributed particle filters were improved in this dissertation; moreover, based on the particle filtering theory a novel distributed particle filter was also devel-oped. Subsequently, the proposed distributed particle filters were adapted for speaker tracking in DMNs.The main contributions of this dissertation are summarized as follows:(1) In the weight consensus-based distributed particle filter (DPF), the computation of its likelihood requires the statistical knowledge of observation noise and the assumption that ob-servations among all individual nodes are independent conditioned on a given state, which is rather strict in realistic environments. To address such problem, a pseudo likelihood function is constructed based on the global coherence field (GCF) function from multiple microphone pairs. A global coherence field-distributed particle filter (GCF-DPF) is then derived and employed for speaker tracking in DMNs. The proposed method requires neither knowledge of the observation noise statistics nor the assumption that observations among individual nodes are independent for a given a state. Meanwhile, it is feasible for distributed implementation. Both the simulation and real-world experimental results show that the proposed speaker tracking method is capable to track the moving speaker in noisy and reverberant environments.(2) As to nonlinear Gaussian systems, a modified distributed Gaussian particle filter (DGPF) is proposed, and then introduced into speaker tracking in DMNs. In the proposed method, all individual nodes run local Gaussian particle filters simultaneously for local esti-mates of the speaker state. In the prediction stage, particles are employed to predict the state posterior, and a fusion scheme is then developed to incorporate all local particles for the global predicted density of the state. In the update stage, based on an average consensus filter, an op-timal fusion rule is employed to merge all local estimates and the common prior among local estimates are removed. Finally, each node possesses the global estimate of the speaker state. In the proposed method, all local estimates are allowed to be correlated, and only local commu-nication is required among neighboring nodes. Simulation and real-world experimental results demonstrate the validity of the proposed method.(3) In some tracking problems, the nonlinear and non-Gaussian state-space models usually contain linear and Gaussian substructures. To address such problem, a distributed marginal-ized auxiliary particle filter (DMAPF) is proposed. In the DMAPF, the linear state variable is marginalized out from the posterior and estimated with the distributed Kalman filter (DKF), whereas the remaining nonlinear variable with the distributed auxiliary particle filter (DAPF). Considering that the speaker state-space model also contains linear and Gaussian substructure, the DMAPF is creatively employed for speaker tracking in DMNs. Specifically, the speaker position is marginalized out from the state-space model and estimated by the DAPF, whereas the speaker velocity with the DKF. Moreover, to cope with the adverse effects of noise and re-verberation, a time difference of arrival (TDOA) selection scheme is also presented, based on the magnitude of the generalized cross-correlation (GCC) and a certain energy ratio. With the TDOA selection scheme, underlying unreliable observations are abandoned, and the reliable ones are exploited for tracking, indicating better estimation accuracy. Simulation and real-world experimental show that the proposed method obtains good tracking accuracy in noisy and rever-berant environments.