Robust Acoustic Event Detection Methods In Driving Environment

Acoustic event is a continuous sound signal, which represents a single fullmessage and can be easily perceived. Meanwhile, acoustic event detection(AED)is the process of identifying the specific acoustic events, by analyzing theobserved audio features. Compared to traditional acoustic event detectionenvironments, driving environment meets the following challenges:(1)in theopen noisy environment, the signal to noise ratio is low, and the backgroundnoise can be arbitrary in magnitude, which changes randomly and often coversthe target sound signal;(2)Road conditions change complexly, leading to thecomplex target sound signals;(3)kinds of target sound need to be identified, eventhe unknown ones.Faced with above problems, we mainly study the robust acoustic eventdetection methods in road environment. We discuss a series of solutions, and themain contents are as follows:First of all, we research on AED methods based on signal periodcharacteristics. As a special kind of acoustic events in driving environment, thevehicle sirens have obvious periodicity. The sound signal period can be detectedwith short-time average magnitude difference function (AMDF) method. Withthe algorithm for recognition and rejection, it can be determined whether thesound signal is period, and whether it is some kind of sirens.Learning from the existing AED methods which are generally based ontraditional pattern classification framework, we study an AED method using melfrequency cepstrum coefficients (MFCC), with support vector machine (SVM)models. We introduce the steps to exact characteristics of MFCC from soundsignal, and the theories and methods to build SVM models. Results ofexperiments made on real sound of driving environment show a grossly declinein detection performance, due to interference and damage by substantiallyrandom noise. So how to exact acoustic features with robustness in sharp noiseconditions becomes the key point of event detection in driving environment.Then we study robust principal component analysis(RPCA) based methods for acoustic event detection, which are effective solutions to substantiallyrandom noise on detection performance. According to RPCA, low-rank matricesexacted from signal matrices can represent the intrinsic complexity of acousticfeatures, and are insensitive under certain noises, especially gross corruptions.lS1olving a convex optimization problem for a combination of nuclear norm andnorm, the RPCA features can be extracted with methods of augmentedLagrange multipliers (ALM) based on singular value thresholding (SVT).Experimental results show that the low-rank representation gets far more betterclassification performance and robustness.At last, we discuss the computation performance of RPCA featuresextraction, and then introduce a fast SVT method with polar decomposition. Asthe state-of-the-art extraction methods first find singular value decomposition(SVD) of matrices and then shrink singular values, the efficiency of the methodhighly depend on the computation performance of SVD. However, as the size ofmatrix increases, the computation of SVD increases dramatically. Byintroducing the dual problem of primal one, and taking the polar decompositionmethods with a range of optimization strategies, we get an effective solution tothe performance of extraction method. Experimental results show that the newmethod saves more than50%computational time from the method via the SVD,without loss of precision. Compared with the AED method based on SVD, theimproved method has the same performance.