Research On Sound Source Localization Based On Robot Auditory System

As greater and greater importance is attached to the public utilities such as national security and public order, the research on the system of the robot with multi-sensor and highly intelligent that can replace human beings in the hazardous environments to perform missions so as to free them from working in danger is of great theoretical and practical significance. Because sound source is the most common target handled by robot in dangerous environment, rapid and accurate sound localization is not only the basic need for humoid robot, but also the basis of sound target separation and speech recognition. For hazardous environment, real time and localization accuracy are the basic and most important targets; therefore, complex sensory devices and location algorithms will affect the efficiency of robot operations which cannot meet the special environment of humanoid robot.This paper, aiming at achieveing the high accuracy of sound source direction, is supported by the Hi-tech Research and Development Program of China -- the Multi-Sensory Robot System of Detecting Dangerous Goods in Extreme Environments (No. 2006AA04Z221). The overall idea of paper is to use the binaural microphone localization system of the humanoid robot with multi-sensory and multi-DOF (degree of freedom) designed by ourselves, and estimate sound source direction from azimuth degree, elevation degree and sound distance with binaural effect, pinnae effect and head-related transfer function. Also, in order to optimize the error in high degree direction of sound in azimuth and elevation, the DOF will be involved in sound localization.Overall, this paper includes the following original points:1. According to the characteristics of robot auditory localization system, paper proposes the method of signal acquisition synchronization with using double buffering and multi-thread technologies between multi-microphone sensors. In the signal acquisition process, also use normalized method to converte the collected data into the format for Matlab analysising directly, while use signal pre-processing technologies, enframe, windowed and dual discriminant signals endpoint detection method, to extract effective signal fragments. Thereby with these pre-processing measures, greatly reduce the redundant and invalid signal data in transmission and analysis, which will improve the speed of sound source loacalization. 2. For signal denoising, paper uses threshold denoising method based on wavelet transform, and how to chooce the threshold is the most key factor in denoising. Paper adopts the block thresholding way which can be obtained more effective result than traditional soft-threshold denoising method by defining the appropriate size of the block. Also the SNR (signal-noise-rate) can be improved with shift invariant theory of wavelet which can suppress Persudo-Gibbs Phenomenon and minish root-mean-square error between the original signals and estimated one.3. With the sound pressure changes caused by pinnae devices and head shadow in rotation, paper gives the rules to determine front-and-back fuzzy localization. And the TDOA theory is used to estimate the azimuth of different frequencies sound source in horizontal plane with binaural effect including ITD, ILD and IPD factors. Also notch frequencies of pinnae principle, signal elevation degree can we estimated with elevation-frequency evaluation functions established in experiments. Meanwhile, the paper also put forward the best pinna device options based on elevation localization which can be referred by other applications with pinnae.4. We know that the localization error is larger than other directions when sound is close to one side of robot head. Therefore, paper provides that the rotating and swing DOFs of robot head, main feature of project robot, can be used to optimize these direction errors. Also, we clarify the algorithm feasibility that paper proposed through system simulation experiments.