Acoustic Event Detection In Driving Environment Of Driverless Cars

Driverless cars need to construct the current surrounding environment model through visual perception and auditory perception during driving.Acoustic event detection is the core of the auditory perception system construction model.The siren sound is the main acoustic event in the driving environment.Therefore,it is necessary to conduct in-depth research on the detection of sirens.(1)This paper takes the siren sound signal from special vehicles as the research target.Firstly,we build a baseline system based on the Mel-Frequency Cepstral Coefficient(MFCC)acoustic characteristics and Support Vector Machine(SVM)pattern classification in the experimental environment,and simulate the correctness of the baseline system.We collected experimental data in the driving environment at the vehicle test site and tested the baseline system.The results show that the baseline system has a lower detection accuracy under driving conditions.(2)In order to solve the problem that the traditional acoustic characteristics MFCC is sensitive to noise interference,we propose a robust acoustic characteristic(HMFCC)of the Harmonic Mel-Frequency Cepstral Coefficient.The algorithm combines the harmonic model of the acoustic signal with the MFCC algorithm to extract the formant frequency in the target acoustic signal and enhance the medium and high frequency components of the target acoustic signal,so as to obtain the acoustic characteristics with better robustness under noisy environment.(3)We built a driving wind noise model and used spectral subtraction to reduce noise.Firstly,the correlation between the low frequency domain and the high frequency domain of the wind noise is verified by Mutual Information.Secondly,the Radial Wind Function Neural Network is used to establish the driving wind noise model.The low frequency domain is used as the input vector and the high frequency domain is used as the output vector.Finally,the estimated high frequency domain signal is denoised by spectral subtraction to remove the noise in the acoustic signal as much as possible(4)In order to improve the performance of the classifier,we use different algorithms to optimize the penalty parameter c and the kernel function parameter g of the SVM model,and verify the optimized SVM classification performance through experimental data.The results show that the detection accuracy of the optimized model is greatly improvedThe study on the removal of noise in acoustic events,that provides a theoretical significance for the development of driverless perception systems for driverless vehicles.