Research On Traffic Acoustic Environment Perception Method For Autonomous Vehicles

Environmental perception is the basis for cars to achieve autonomous driving.With the continuous development of self-driving car technology,the technical requirements for environmental perception are becoming higher and higher.Multi-sensor fusion has become an inevitable trend to realize the perception of autonomous driving environment.This paper achieves partial perception of the acoustic environment by recognizing the acoustic events of the traffic environment,and makes up for the blind spots of the traditional perception system.In the traffic environment,driving safety has high requirements for the recognition rate and robustness of acoustic events.Based on the experimental conditions of a research center,this paper completes the collection and identification of four traffic environment acoustic events of 110 sirens,120 sirens,119 sirens and screams.The main research contents are as follows:First,build a baseline system for traffic environment acoustic event recognition based on the MFCC and SVM.Experiments verify that the baseline system has a high recognition rate,but also the problem of poor robustness is exposed.Second,in order to improve the basic recognition rate of the system,by analyzing the time-frequency domain characteristics of the four sound signals,a three-dimensional time-frequency domain feature extraction method is proposed.And,the MFCC analysis is performed to remove the interference terms in the MFCC parameters.a feature parameter optimization method based on EMD is proposed to make the feature parameters have a filtering effect in the extraction process.The Optimization feature are recorded as EMD-MFCC optimized combination parameters.Experiments show that EMD-MFCC optimized combination parameters have better anti-noise effect and higher basic recognition rate.Third,a denoising method based on wavelet packet transform is proposed to improve the anti-noise performance.Analyze the commonly used threshold criteria and threshold functions from both theoretical and experimental aspects,and propose multi-threshold criteria and an improved threshold function to achieve a betterdenoising effect and achieve the adaptability of the threshold function to noise.At the same time,the wavelet packet basis functions and decomposition layers are optimized.Finally,the experiment verifies that the optimization of the denoising method of traffic environment acoustic signal based on wavelet packet transform can greatly improve the recognition rate of the recognition system for low signal-to-noise ratio and noise,so that the traffic environment acoustic event recognition system can adapt to complex traffic Environmental capacity.Fourth,through theoretical analysis of the SVM classifier,find out the factors that affect the recognition performance: kernel function and kernel function parameters.First,compare the impact of the commonly used kernel functions in the SVM model on the recognition performance through experiments.Second,compare the effects of different optimization algorithms on the penalty factors and kernel parameters.The experimental results show that the Gaussian kernel function has a higher recognition rate and the shortest training time;the particle swarm optimization algorithm takes less time to optimize parameters,and the obtained parameters have better adaptability to noise.Finally,the simulation verifies the effectiveness of the traffic environment acoustic event recognition system under driving conditions.The research results show that through the optimization of feature parameters,denoising algorithm optimization and SVM model optimization in this paper,the basic recognition rate of the traffic environment acoustic event recognition system for noise-free traffic environment acoustic events is increased to 99.17%.The recognition rate of acoustic events in traffic environment with 20 dB can reach 87.08%.and the low signal noise(0-10dB),the recognition rate of acoustic events in the traffic environment is improved more than 50%,which improves the basic recognition rate and robustness of the acoustic event recognition system in the traffic environment.Finally,the feasibility of the traffic environment acoustic event recognition system under driving conditions was verified by simulation.