摘要
随着机动车违法鸣笛现象日益严重,汽车鸣笛声识别可以识别违法鸣笛车辆,并对该行为给出科学有力的证据,因此对城市交通治理有着重要意义。传统方法主要包含基于GMM-HMM的概率模型算法、支持向量机等。但其准确率较低,且过程麻烦,给交管部门进行人工复核造成了很大困难。针对此问题,以城市交通汽车鸣笛声识别为背景,结合深度信念网络(DBNs)强大的非线性建模和特征提取能力,提出了一种优化的声音识别方法。该方法采用汽车鸣笛声信号的梅尔频率倒谱系数(MFCC)以及其一二阶差分作为特征参数,用于DBN网络的输入,对样本数据进行建模并提取更深层的特征,最后加入softmax分类器来实现汽车鸣笛声信号的匹配和识别。该方法获得比GMM-HMM更好的识别效果。并通过仿真实验证明了该方法的有效性。
With the increasingly serious whistle of motor vehicles,the whistle recognition of the car can identify the whistle-blowing vehicle and give strong scientific evidence to the behavior,which is of great significance to the urban traffic control.The traditional methods mainly include probabilistic model algorithm based on GMM-HMM,support vector machine and so on,but their accuracy is low and the process is troublesome,which makes it difficult for the traffic control department to carry out manual review.In order to solve this problem,based on the recognition of whistle sound in urban traffic vehicles,we present an optimized sound recognition method in combination of the powerful nonlinear modeling and feature extraction capabilities of DBNs(deep belief networks).The MFCC(Mel-frequency cepstrum coefficients)and its first and second order differences are used as the characteristic parameters for the input of the DBN network,and the sample data are modeled and further features are extracted,Finally softmax classifier is introduced to achieve the car whistle signal matching and identification.This method gives better recognition than GMM-HMM,and its effectiveness is proved by the simulation.
作者
郑皓
赵庶旭
屈睿涛
ZHENG Hao;ZHAO Shu-xu;QU Rui-tao(School of Electronics and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《计算机技术与发展》
2019年第2期60-64,共5页
Computer Technology and Development
基金
甘肃省自然科学基金(1504GKCA018)
关键词
神经网络
深度信念网络
特征提取
梅尔频率倒谱系数
汽车鸣笛声识别
neural networks
depth belief network
feature extraction
Mel-frequency cepstrum coefficients
car whistle sound recognition
