基于线性判别分析的室内声源定位方法

Indoor Acoustic Source Localization Method with LDA

在小信噪比和混响时间较长的恶劣环境下,基于模式分类的手段能够有效克服传统的声源定位算法鲁棒能力不足的缺点,其中朴素贝叶斯分类器定位的准确率高,计算量小,鲁棒能力强。在此基础上,为了获得更好的定位性能,提出使用线性判别分析(LDA)分类器进行声源定位。使用Matlab进行仿真,截取声源信号的相位变换加权广义互相关函数(PHAT-GCC)作为特征向量,通过投影变换,找到最佳的特征空间来区分特征数据,从而训练得到线性判别分析分类器。然后在不同的混响时间和信噪比的条件下,进行定位测试,比较了线性判别分析分类器和朴素贝叶斯分类器的性能。仿真结果表明,在环境恶劣场合更宜使用线性判别分析分类器,特别是混响严重时,线性判别分析分类器的定位准确率比朴素贝叶斯分类器高1%~2%。

The method based on pattern classification can overcome the deficiency of traditional acoustic source localization algorithms which has an insufficient robust ability in the harsh environment of small SNR and severe reverberation. Among them, Naive Bayes classi- fier has high location accuracy with a small amount of calculation and strong robustness. In order to achieve better localization perform- ance ,Linear Discriminant Analysis （LDA） classifier is adopted to locate acoustic source on the basis of former research. It has been test- ed by Matlab,while the Phase Transform Generalized Cross-Correlation （PHAT-GCC） function would be used as feature vector. LDA classifier has been trained through projection transformation which could help to find a better feature space to discriminate the feature da- ta. Subsequently, the source would be located in different reverberation and noisy conditions to compare the performance with LDA classi- fier and Naive Bayes classifier. The simulation results have demonstrated that LDA classifier is a better choice in harsh environment and that the location accuracy of LDA classifier is higher than that of Naive Bayes classifier by 1% to 2%, especially in severe reverberation environment.