High Precision Sound Source Localization Method Under Low Fingerprint Density

The process of using the auditory information to determine the location of sound source is called sound source localization.In recent years,sound source localization based on microphone array is playing a key role in human-computer voice interface,video conference and military radar detection,etc.The technology becomes popular in the area of localization.At present,the common sound source localization methods are divided into geometric method and fingerprint method.Among them,the sound source localization method based on location fingerprint effectively improves the defects of the geometric localization method,and has the advantages of low model dependence and high positioning accuracy.However,in order to ensure the accuracy,the fingerprint-based localization method needs to establish high fingerprint density database,which not only increases the workload of the offline sampling stage,but also seriously affects the real-time performance of the online positioning.Aiming at the problem,this paper optimizes and improves the sound source localization method based on location fingerprint.Firstly,a distributed microphone array with four elements is used to construct sound field environment covering the location area.The mobile robot is driven to collect sound information at a small number of location calibration reference points.And the time difference of arrival(TDOA)characteristics of each reference point are obtained through dual-threshold endpoint detection technology,which reduces the multi-peak and function peak weakening phenomenon existed in the classical speech endpoint detection method.Secondly,from the actual application scenario,the feature vector with location identification function is processed by average operation,which can suppress the influence of measurement error on fingerprint quality.The sound location fingerprint is stored in the form of row vectors,and the clustering algorithm is adopted to process database,which reduces the time and complexity of searching high-dimensional database and improves the efficiency of the localization.Thirdly,the existing location fingerprint-based localization method mostly uses the K-nearest neighbor algorithm to calculate the target location,which lead to a serious waste of fingerprint resource.For this,a fast search strategy based on TDOA vector is proposed to solve this problem.In this strategy,the TDOA features collected in real time at the sound source to be located are arranged into vectors.Feature matching degree between the TDOA vector of the cluster center and the online TDOA vector is used as the correlation coefficient,and the cluster of the sound source is selected according to the value of correlation coefficient.The sound source location is calculated in the selected cluster,which further accelerates the speed of online matching and locating.In addition,in order to improve the localization accuracy,the database is updated by interpolation algorithm in the selected cluster,which can greatly reduce the workload of sample collection and make the fingerprint-based localization method more suitable for practical application.Finally,for the affiliation clustering problem of the sound source to be located at the cluster boundary,an iterative-based localization error correction method is proposed.By generating virtual adjacent reference points,the location of sound source with low accuracy is corrected,thereby improving the localization accuracy of system.The sound position fingerprint-based localization method is explored in this paper,which lays the theoretical foundation for the realization of more efficient and accurate indoor sound source localization method under the low fingerprint density.