Multi-speaker Tracking Method Based On Audio-visual Feature Fusion Under Intelligent Environment

In recent years, with the development of the remote video conference system andautomatic analysis system of the meeting, speaker tracking under the intelligentenvironment has become the focus problem in the field of human-computer interaction.Speaker tracking have a wide range of applications in the pervasive computingsysterm、video conference system and intelligent robot navigation system. With thedevelopment of artificial intelligence technology, speaker tracking system which basedon microphone array or computer vision can not meet the accuracy requirement of thetrack. Based on this, this paper mainly research multiple speaker tracking under theintelligent environment with audio-visual features fusion. And mainly made thefollowing several aspects of the research results:(1) Speaker tracking system has a typical nonlinear feature. In the view of thewidely applications of the particle filter algorithm in nonlinear non-gaussian system,this paper research on the particle filter algorithm for further.We know that the statetransition probability density function can’t achieve the latest measurementinformation updation, and easily cause the phenomenon of the weights degradation.This paper adopted the finite center difference method to update the probability densityfunction and introduced the values of the information gap noise. The number ofparticles could update real-time and was well to overcome the phenomenon of theweights degradation.(2) This paper proposed a new kind of particle filter method based on finitedifference to locate and track spesker based on the microphone array. This methodintroduced the adaptive finite difference particle filter into speaker tracking problem,and got the rubust results in the low signal-to-noise.(3) Head tracking by using single feature results in a poor performance inrobustness. To solve this problem, an head tracking method based fusing measurementsof head by using D-S evidence theory. The proposed method uses the color anddistance to maximum gradient point (DMG) features as the observation model, andefficiently avoids the unsatble problems via using single color feature in theillumination of mutation, posture change, greater distance and similar background.(4) At last, this paper proposed a new kind of speakers tracking method whichbased on audio-visual unded the adaptive finite difference particle filter frame. Thismethod established the dynamic movement model and the system observation model based on the theory of filtering by the the full analysis to the dynamic movementmodel discipline. This paper realized the audio-visual speakers tracking problem in theintelligent environment.