2007 IEEE International Conference on Signal Processing and Communications (ICSPC 2007), 24-27 November 2007, Dubai, United Arab Emirates TRACKING MANEUVERING SOURCES IN ICA S.A. Banani 1, MA. Masnadi-Shirazi 2, A. Masnadi-Shirazi3 IDept of Electrical Engineering, Shiraz University, Shiraz, Iran 2Dept of Electrical Engineering, Shiraz University, Shiraz, Iran 3Dept of Electrical Engineering, University of California, San Diego, USA ABSTRACT The problem addressed in this work is to separate the signals of moving sources with independent component analysis (ICA) and tracking the kinematics (position, velocity, acceleration) of each individual source in the working space using Particle filters. To identify the unpredictable movement of the speaker over time, the new proposed state switching scheme handles the uncertainty of the speakerzs motion by incorporating multiple motion models in the tracking process instead of using the conventional IMM algorithm. The algorithm performance has been verified by illustrating some simulation results. Index Terms- ICA; Particle filter; Source separation; tracking; positioning 1. INTRODUCTION The problem of tracking sources in reverberating environments is relevant in several applications, including seismology, sonar and speech. Localizing and tracking multiple speakers talking in the same room can be used, for example, to automatically steer camera sensors in video-conferencing applications. Several works with deferent strategies have been done in the field of tracking and source separation. Some works like [1] focuses on TRINICON algorithm in the noise free environment. But in real life, there is always some kind of noise present in the observations. Noise can correspond to the actual physical noise in measuring devices or accuracies of the model used. In [2] an algorithm based on IMM-PDA filters has been proposed. Each of the speakers, state equation describing their movement and the observation equation has been assumed to be a linear function of the state. However, any of the equations may be a nonlinear function of the states. In such a case, using Kalman filters are not suggested. Furthermore, when there are severe nonlinearities in either of the state or observation equations, extended Kalman filters falls beneath its suboptimal performance. In this cases, using Particle filters as nonlinear state estimators are more suitable. In this work we will present a novel, general framework that can deal with both cases, that is, dealing with the nonlinearities of the state and observation equations for tracking the sources and separating the voices of multiple, possibly moving, speakers in the noisy environment. In order to be able to cover the unpredictable movement of the speaker over time, the new proposed state inference scheme, handles the uncertainty of the speakerzs motion by incorporating multiple motion models in the tracking process. 2. PROBLEM DESCTRIPTON 2.1. ICA Model In the standard noisy ICA, the noise is assumed to be additive. This is a rather realistic assumption used in factor analysis and signal processing, and allows for a simple formulation of the noisy model. Thus, the noisy ICA model can be expressed as Ok = Aksk + Wk (1) where the vector Sk is the vector of independent sources and 0k iS the observation vector in each iteration and Wk is the vector of additive noise that in general can have any non-Gaussian distribution. Assume that we have L independent source components and M observations at each iteration. The indices k shows that in each iteration, the mixing matrix A is changing due to the movement of sources or possibly non-stationary environment of the work space. As the elements of matrix A, i.e. aij, are some parameters that depend on the distances between the microphones and the sources, we may write any desirable nonlinear relation between a, and the distances. Thus we may write a,j = fij (rij), i = 1,2,..., M j = 1,2,...,I L (2) where is the distance between source j and microphone i. Note that in general, r, depends on x, y and z which are the related distances in three dimensions of working space. For example if source i is located in the origin, i.e. (0, 0, 0), We may write 1-4244-1236-6/07/$25.00 © 2007 IEEE 991