Two stage feature extraction using modified MFCC for honk detection

Robust and accurate detection of vehicle horn along with the rate of honking can estimate the traffic state of a street in an urban area. Participatory sensing using audio of users´ mobile phones is being used for monitoring the environment. In this paper, we propose a Spectral Based Mel-Frequency Cepstral Co-efficient (SBMFCC) feature for horn detection which considers the spectral characteristics of the sounds in feature computation. The proposed approach modifies the conventional Mel filter bank structure according to the varying nature of spectral energy distribution of the horn sound. A two stage feature extraction approach is also proposed to further reduce the processing on mobile device. Database of different traffic sounds including vehicle horns are collected under various road and traffic conditions to perform a comparative study of the performance. Experimental results prove the effectiveness of SBMFCC feature over the conventional MFCC feature while using a Gaussian Mixture Model (GMM) classifier. Substantial saving in processing load is achieved with the incorporation of two stage feature extraction process.