基于改进VGG-16网络的交通声音事件分类方法研究

Research on Traffic Sound Event Classification Method Based on Improved VGG-16 Network

交通声音事件分类是提升城市智慧交通系统环境感知能力的关键技术之一.针对传统交通系统的环境声音感知能力弱、效率低、鲁棒性低、可分类数量少等问题,研究了一种基于VGG卷积神经网络的交通声音事件分类方法,该方法使用语谱图(spectrogram image features,SIF)作为交通声学特征,建立并优化了卷积神经网络(convolutional neural networks,CNN),从而实现交通声音的智能分类.首先,使用实验室采集的10种交通声音,构建了交通声音数据集.其次,利用语谱图方法对交通声音进行声学特征提取,搭建VGG-16分类算法主模型,通过双卷积层融合算法和块间直连通道对网络进行改进,得到了VGG-TSEC网络.该优化网络的交通声音事件分类准确率可达97.18%,与优化前相比准确率提升4.68%,其权重参数降低72.76%,占用空间降低384MB.同时,将该优化模型与K邻近(KNN)、支持向量机(SVM)等机器学习方法进行对比,其准确率分别提高了19.68%和4.41%.结果表明,VGG-TSEC交通声音分类方法可以实现警笛音、事故碰撞、行人尖叫、卡车等交通声音的高效分类,为交通声音事件分类提供参考.

Traffic sound event classification is the most important step to improve the environmental perception ability of transportation system.Aiming at the problems of traditional traffic system,such as weak sound perception,inefficiency,low robustness and few detectable types,a traffic sound event classification method based on VGG was studied.This method used Spectrogram image features(SIF)as traffic sound features established the Convolutional neural networks(CNN)to complete intelligent classification of traffic sounds.Firstly,a traffic sound dataset was constructed using 10 sounds collected in the laboratory.Then,the SIF method was used to extract the acoustic features of traffic sounds,and the main model of VGG-16 classification algorithm was built.Finally,the VGG-TSEC network is improved by fusion algorithm with two convolution layer and inter-block channel algorithm.The final experiment shows that traffic sound event classification accuracy of the optimized network can reach 97.18%,which is 4.68%higher than that of before optimization.The weight parameter is reduced by 72.76%and the resource consumption is reduced by 384MB.At the same time,the optimization model is compared with machine learning such as K-nearest neighbor(KNN)and support vector machine(SVM),and the final accuracy was improved by 19.68%and 4.41%,respectively.The results show that the VGG-TSEC traffic sound classification method can achieve efficient classification of traffic sounds such as siren sounds,accident collisions,pedestrian screams,and trucks sounds,etc.,which can provide a reference for the traffic sound event classification.