Feature Selection for Audio Surveillance in Urban Environment

Abstract

This paper presents the work leading to the acoustic event detection system, which is designed to recognize two types of acoustic events (shot and breaking glass) in urban environment. For this purpose, a huge front-end processing was performed for the effective parametric representation of an input sound. MFCC features and features computed during their extraction (MELSPEC and FBANK), then MPEG-7 audio descriptors and other temporal and spectral characteristics were extracted. High dimensional feature sets were created and in the next phase reduced by the mutual information based selection algorithms. Hidden Markov Model based classifier was applied and evaluated by the Viterbi decoding algorithm. Thus very effective feature sets were identified and also the less important features were found.

Keywords

• acoustic event
• feature extraction
• feature selection