Research On Vehicle Type Classification Based On Acoustic Signals Using CNN,CRNN And SVM

With the development of intelligent cities and transportation technology,people have higher requirements for the development of vehicle type classification technology.Vehicle type classification technology can not only realize the functions of monitoring traffic flow and highway toll collection,but also assist cars to perceive the surrounding environment information and provide further safety guarantee for the driverless cars.How to classify vehicle types simply,accurately and fast is the current research hotspot of vehicle type classification technology.This thesis addresses complex,multi-source and non-smooth characteristics of vehicle acoustic signals and improves the classification algorithms based on the acoustic signals’ feature extraction,increasing the classification accuracy through experimental verification in the end.The main contents and innovations mainly include:(1)Establish a multi-type vehicle acoustic signal database,analyze the acoustic signals’ characteristics and select the corresponding feature extraction algorithmWith the reference to the domestic standards of vehicles’ types,four vehicles’ types are decided as the classification objects.Different vehicle types’ acoustic signals have been collected to consist of the database,and their spectral characteristics are analyzed.This paper selects Mel spectrograms method as feature extraction method which can amplify the energy in the low frequency band according to the characteristics of vehicles acoustic signals spectrum that most of their energy are in the low frequency band.(2)Improve convolutional neural network algorithm and convolutional recurrent neural network algorithm to achieve improved classification performanceBy considering that the convolutional neural network can effectively extract and process the deep typical features of signals through convolution and pooling operations,this paper proposes a new suitable convolutional neural networks model to do classification tasks with the basis of improving the classical convolutional neural network Le Net-5 structure and combining with the characteristics of the established database.Because acoustic signals are sequential signals,this paper also proposes a recurrent convolutional neural networks model with the basis of the proposed convolutional neural networks model.Through experiments,the improved classification algorithms have been validated that they can improve classification performance compared with other classification algorithms from the similar papers.(3)Propose and validate new classification schemes that filters combined with classification algorithms to improve classification performanceSince the collected vehicles’ acoustic signals contain background noises which may affect classification performance,two filter schemes are designed in this paper: filter one is to remove part of the background noises,filter two is to remove part of the background noises and retain most of the energy of the vehicle acoustic signal.By comparing experimental results,the overall F1 value of the filter one with CNN classification algorithm increases4.9%,and the overall F1 value of the filter two with SVM classification algorithm increases6.5%.The new classification schemes show better classification performance.