Research On Rapid Acoustic Event Detection In Noisy Driving Environment

Under the given circumstances,acoustic event(AE) is that humans can identify and understand a period of continuous full meaning of the sound signal. Acoustic event detection(AED) means to detect the specific acoustic events by the acoustic features of the sound signal, which can be generally realized by the classification algorithm of acoustic event. The acoustic event detection tasks of this issue of acoustic environment is different from common testing environment, acoustic event be primarily affected by driving noise, the noise is mainly from wind noise and engine vibration, etc. Wherein the engine vibration noise always persists, in the processing of driving, the wind noise is generated by the friction of the vehicle and the airflow has complex and changing characteristics. Moreover, due to the vehicle is a closed environment, the reverberation of all kinds of noise in the vehicle interior is formed. This causes a large difficulty for the acoustic event detection tasks in driving noise environment. In addition, in order to the sound perception information can be provided to the Intelligent Decision Support System, the acoustic event detection system only timely, rapid detects the target acoustic event, too much delay time will affect the normal running of unmanned vehicles and finally lead to a traffic accident.For the spectral characteristics of the driving noise and the target of acoustic events, in this paper, we regard the interior noise environment as the background noise, and the implement of fast and accurate detection of the approach of acoustic events is designed as the main content of the paper. The main contents are as follows:Firstly, in order to remove noise components from the sound signal as much as possible to obtain a pure sound signal, we predict vehicular interior noise and reduce the noise by spectral subtraction. Specific contents include: we use Markov Chain to modeling the changes of the vehicular interior noise, and after wavelet analysis of the vehicular interior noise, using the linear prediction coefficients predicts the noise component of acoustic events sound signal, then uses spectral subtraction to reduce the noise, so as to remove the noise from the signal of acoustic events.Secondly, in order to reduce the impact of noise on the acoustic event detection as much as possible, improving the noise robustness of the acoustic features, this paper presents a robust acoustic feature extraction approach that is based on frequency selectivity of human auditory under the driving noisy environment. Human’s ears are more sensitive for vibration peaks at different frequencies, through perceiving the different frequencies of vibration peaks and enhancing frequency, thus declines the impact of noise on auditory perception. In this paper, we realized that the simulation of the human ear’s perception of frequency selective gain characteristics, and use the information of the formant to weight the Mel filter banks, thus extracts noise robust acoustic feature.Thirdly, in order to quickly detect the acoustic events and improve the accuracy of acoustic event detection and reduce the error rate, we test the performance of classification of acoustic events of traditional signle kernel svm(SK-SVM) and multiple kernel svm(MK-SVM) in the driving noise environment, and analyze the deficiencies of those methods in this project application, multi-scale RBF kernel function is introduced to create the base kernel sets, we propose a training algorithm of multi-scale RBF SVM classifier. This algorithm not only solves the problem of single kernel function mapping capability, but also effectively decreases the time of training traditional MK-SVM and improves the accuracy and reduces the error rate, it is noteworthy, the multi-scale RBF kernel proposed in paper can take less time to detect acoustic events, and achieve the aim of the rapid detection of acoustic events.This AED system is developed with C++ programming language and taked experiment on the unmanned vehicle under real driving noise environment. The experimental results show that the system achieves the desired acoustic event detection accuracy and error rate requirements, and increases the speed of acoustic event detection, addressing the rapid detection of acoustic events latency requirements and the system has a good performance in terms of noise robustness.