An application of a particle filter to bayesian multiple sound source tracking with audio a.pdfVIP

An Application of a Particle Filter to Bayesian Multiple Sound Source Tracking with Audio and Video Information Fusion Hideki Asoh, Futoshi Asano, Takashi Yoshimura Kiyoshi Yamamoto Yoichi Motomura, Naoyuki Ichimura, Isao Hara, Jun Ogata Institute of Information Technology RI Information Science Electronics National Institute of Industrial Science and Technology University of Tsukuba AIST Central 2, 1-1-1 Umezono 1-1-1 Tennodai Tsukuba, Ibaraki 305-8568, Japan Tsukuba, Ibaraki 305-8573, Japan h.asoh@aist.go.jp Abstract – A particle filter is applied to the problem of detect- ing and tracking multiple sound sources by Bayesian inference using combined audio and video information. The problem is for- mulated within a general framework of Bayesian hidden variable sequence estimation by fusing observed information. The parti- cle filter is then introduced as an approximation of Bayesian in- ference. Experiments using real-world data demonstrate that the proposed method works well in ordinary environments such as a meeting room. The computational cost of estimation is reduced significantly compared to exact Bayesian inference, while main- taining the quality of estimation. Keywords: Particle Filter, Bayesian Multiple Target Tracking, Sound Source Tracking. 1 Introduction The importance of statistical inference using probabilistic models in multi-modal information fusion has been pointed out by many researchers, and the Bayesian approach in par- ticular has been promoted and applied to various problems in recent years. In the Bayesian approach, all variables in models are treated as random variables and both inference and learning can be treated in a unified way as a computa- tion of conditional probabilities of the target (hidden) vari- ables or parameters conditioned by observed information. Although the Bayesian approach is very simple and pow- erful in principle, the central drawback in practice is that it often requires intractably large amount of computations, main