Research On Distributed Kalman Filter Based Speaker Tracking Method In Microphone Array Networks

Speaker localization and tracking with microphone arrays have played an important role in many scenarios,including audio/video conferences systems,public surveillance systems,auto-matic speech recognition,vehicle telephone and robotics.Recently,with the advance in wireless sensor networks,mobile computer,integrated circuit and embedded processor,the distributed speaker localization and tracking techniques using microphone array networks have been mo-tivated in speech signal processing field.Employing the distributed data processing manners,the distributed speech source localization approaches do not restrict the geometry structure of microphone arrays to be regular,are robust against lost data or link failure,and possess high-er reliability and stability for speaker tracking.However,these distributed speaker localization methods only depend on signals in the current frame,hence are not robust against high room reverberations and background noises.To deal with these problems,it is necessary to adapt the distributed speaker tracking methods to estimate speaker's positions robustly.Distributed speaker tracking is a typical nonlinear filtering problem.In this dissertation,several distributed nonlinear filtering approaches are proposed and introduced into the speaker tracking problem in microphone array networks.Therefore,the speaker's state estimates can be obtained accurately and stably.The main contributions are as follows:(1)To deal with the problems that the distributed localization methods are not robust a-gainst high room reverberations and deteriorated background noise,a distributed Kalman filter(DKF)based speaker tracking method is proposed.To solve the problem that the estimated time delay estimations(TDE)at each node in microphone array networks may be invalid because of the room reverberations and background noises,a TDOA validation scheme is developed.In speaker tracking,the valid are gathered TDE from neighboring noses,Langevin model are em-ployed to describe the speaker's movement mode,and the DKF is used to estimate the speaker's state accurately.The proposed speaker tracking method can improve the tracking robustness significantly,has performance characteristics in low communication loads,high reliability and stability,and flexible use,etc.(2)To deal with the nonlinearity in speaker tracking,a distributed unscented Kalman filter(DUKF)is first proposed,which has second order accuracy.After that,the interacting multiple-model(IMM)algorithm is introduced to describe speaker's difference movement mode and a distributed IMM unscented Kalman filtering(D-IMM-UKF)algorithm is proposed for the de-termination of speaker's positions in microphone array networks.For speaker tracking,the pro-posed speaker tracking method gathers the valid TDE from neighboring noses,models speaker's different motion modes,such as standing,walking slowly,walking,and running,and fuses the state estimates from multiple-model based filter.In this way,the tracking accuracy can be im-proved more effectively.(3)To deal with the problems that the initial target state may be far from optimal value in speaker tracking,a distributed iterated extended Kalman filtering(DIEKF)algorithm is proposed and introduced into the microphone array networks for speaker tracking.Meanwhile,in order to guarantee that the iterations increase the likelihood,i.e.iterations move towards the maximum likelihood solution,a iteration termination strategy is delivered.By performing several local iterations,the DIEKF can obtain the speaker's position more accurately in microphone array networks.Especially,when the initial target state is far from the optimal value,the DIEKF can capture the speaker's moving trajectory faster by performing several local iterations.(4)Taking into account the merit of the cubature Kalman filtering and the iterated extended Kalman filtering algorithms,which have high accuracy and fast convergence rate,respectively,a iterated cubature Kalman filtering(ICKF)algorithm is proposed,which has second order ac-curacy.After that,a distributed iterated cubature Kalman filter(DICKF)is proposed and intro-duced into the microphone array networks for speaker tracking.The DICKF has second order accuracy,low computational complexity and fast convergence rate.Hence,it can overcome the nonlinearity in speaker tracking and improve the tracking accuracy more effectively.