Trainable COPE Features for Sound Event Detection

摘要

Systems for automatic analysis of sounds and detection of events are of great importance as they can be used as substitutes of or complement to video analytic systems. In this paper we describe a flexible system for the detection of audio events based on the use of trainable COPE (Combination of Peaks of Energy) features. The structure of a COPE feature is determined in an automatic configuration process on a single prototype example. Thus, they can be adapted to different kinds of sounds of interest. We configure a set of COPE features in order to account for robustness to variations of the characteristics of sounds within a specific class. The proposed system is flexible as new features (also configured on examples drawn from new classes) can be easily added to the feature set. We performed experiments on the MIVIA road events data set for road surveillance applications and compared the results that we achieved with the ones of other existing methods.