| Traffic whistle is a common type of noise in the urban traffic environment,and it is one of the main sources of urban noise.With the continuous and rapid development of China's economy and society,the problem of traffic whistle pollution has become increasingly serious.The continuous increase of traffic whistle has brought more and more bad effects on urban traffic and residents' lives.Therefore,it is imperative to control traffic whistle.At present,vehicle whistle capture systems are generally installed on urban arterial roads to capture whistle violations,while in non-main road areas,illegal traffic whistle is mainly rectified by manual control.This method is not only inefficient but also consumes A lot of manpower and material resources.Therefore,intelligent whistle detection and identification for such areas is of great significance for controlling urban traffic and improving the happiness of urban residents.This article takes traffic whistle sounds around campus as a research object,uses low-cost portable recording equipment to collect traffic whistle sound data,discusses in detail the relationship between the spatial and temporal distribution of traffic whistle sounds and environmental and traffic factors,and implements a single channel Classification of traffic whistle,and further combined with deep clustering algorithm to achieve single-channel concurrent traffic whistle classification.Specifically,traffic whistle samples were first collected from low-end portable recording equipment from road sections around the campus.The spatial and temporal distribution of traffic whistle samples was statistically analyzed,and detailed descriptions of traffic whistle sounds on the sample collection roads were collected.The frequency distribution of different seasons,one week,and one day,combined with factors such as visibility,traffic flow,road characteristics,etc.,were used to analyze the frequency fluctuation of traffic whistle.Secondly,according to the collected traffic whistle,two common features of Mel Frequency Cepstral Coefficient(MFCC)and wavelet packet energy spectrum are extracted,and two commonly used classifiers: Support Vector Machine(SVM)and Random Forest are used to whistle traffic.The whistle sounds are classified into whistle sounds of large vehicles,whistle sounds of small and medium-sized vehicles,and whistle sounds of twowheeled and three-wheeled vehicles.The average classification accuracy rate is 90.3%.Considering that the environmental noise contained in the traffic whistle collected in the real environment has a certain effect on the classification effect,three types of noise reduction methods such as spectral subtraction,Wiener filtering,and non-negative matrix factorization(NMF)are used to whistle The samples were subjected to noise reduction processing.Through comparison,it was found that the signal classification performance after noise reduction using the NMF method was greatly improved,reaching 2.3%.In addition,a single-channel concurrent whistle separation method based on deep clustering was studied and implemented in response to the high probability of concurrent traffic whistle in the real environment.This method first obtains the embedding features of traffic whistle by training a bi-directional long short-term memory(BLSTM),and then clusters the embedding features based on the K-means algorithm to achieve the separation of concurrent traffic whistle.Finally,the sirens classification and performance analysis were performed on the separated signals using the aforementioned classification method,and the average classification accuracy reached 81.3%,which effectively achieved single-channel concurrent whistle classification of traffic.Based on the research results of traffic whistle sound data collected around campus on low-cost portable devices,this article provides a convenient and efficient auxiliary method for traffic management and control in areas where there is no monitoring system for whistle whistle of vehicles. |
PDF Full Text Request