基于改进小波包去噪与梅尔倒谱系数的低信噪比交通环境声音识别

随着自动驾驶汽车研究的不断深入,对其环境感知系统提出了更高的要求。为了使自动驾驶汽车适应更复杂的交通环境,本文研究了低信噪比声学环境感知技术,提出改进的小波包去噪方法；采用经验模态分解(EMD)的方法改进梅尔频率倒谱系数(MFCC)的提取；采用支持向量机(SVM)识别模型完成低信噪比交通环境声音识别。实验结果表明,本文提取的去噪方法提高声音事件信噪比的同时保持声音特征,且对噪声有自适应性；改进的MFCC提取方法一定程度上提高了特征参数的抗噪性能。通过对低信噪比交通环境声音去噪和特征参数优化后,其平均识别率比优化前提高了33.34%,并改变了识别率骤降的趋势。

Low SNR Traffic Environment Acoustic Recognition Based on Improved Wavelet Packet Denoising and Mel Cepstrum Coefficient

With the deepening research on autonomous vehicles, higher requirements were placed on their environmental awareness systems. In order to adapt the self-driving car to the more complicated traffic environment, this paper studied the low signal-to-noise ratio (SNR) acoustic environment sensing technology and proposed an improved wavelet packet denoising method. In addition, the empirical mode decomposition (EMD) method was used to improve the Mel frequency cepstral coefficient (MFCC). Support vector machine (SVM) recognition model was used to complete the low SNR traffic environment acoustic recognition. The experimental results showed that the denoising method extracted in this paper improved the SNR of acoustic, maintained acoustic characteristics, and was adaptive to noise. The improved MFCC extraction method improved the anti-noise performance of the characteristic parameters to some extent. After denoising the low SNR traffic environment acoustic and optimizing the characteristic parameters, the average recognition rate is 33.34% higher than that before optimization, and the trend of sudden drop in recognition rate is changed.