Study On Deep Learning And D-S Theory Based Speaker Tracking Algorithm In Distributed Microphone Array

Speaker localization and tracking have been widely applied in smart systems,includingaudio/video conference system,smart monitoring system and smart home.Compared to traditional regular microphone arrays,distributed ones have advantages of low data transmission,strong fault-tolerant ability and higher stability.Current methods mostly take Kalman filter or particle filter to estimate speaker's state transition under the assumption that speaker's motion model accords with Langevin model.However,in real scenarios,one or several fixed models can't accurately represent speaker's motion model.Besides,distributed tracking algorithm allows data to be processed within each node,thus greatly reducing the data transmission volume,communication bandwidth and energy.Therefore,distributed data fusion,as an important part of distributed tracking algorithms,has attracted much attention.However,so far,consensus algorithm is still a more applied method of data fusion.Although it is simple in theory and easy to be implemented,the accuracy is worse than expected.The main contributions of this thesis are as follows:(1)In allusion to Langevin model failing in accurately representing speaker's motion,a distributed Kalman filter based speaker tracking method with learned motion model using Long Short-term Memory is proposed so as to improve the precision of speaker's state transition estimation.Simulation results show that the proposed method performs well in noisy and reverberant environment and is robust to broken nodes in network.(2)To deal with slow convergence rate and low accuracy of consensus algorithm caused by simple weighting coefficients,a D-S based distributed data fusion method is proposed.Belief values converted by TDOAs with basic probability assignment function(BPA),rather than TDOAs themselves,are fused within the network.Simulation results show that the proposed method can effectively improve the anti-noise and anti-reverberation performance.(3)Since there is no favorable and unified BPA generation method,a kernel density estimation based BPA generation method is proposed.It can be applied to different scenarios and effectively improve the precision of belief value transition.