Discriminative Acoustic Event Recognition in multimedia recordings

In this paper, we describe an Acoustic Event Recognition (AER) system for locating events of interest in the audio stream of multimedia recordings. We focus on two non-speech acoustic events; bomb explosions and gunfire, which typically exist in surveillance videos and are of importance in monitoring and alerting applications. Recognition is performed using a discriminative approach based on Support Vector Machines (SVM). We compare the new approach to a baseline system that utilizes a Hidden Markov Model (HMM)-based classification approach. We performed experiments on a corpus of publicly available video files containing gunfire and explosion events. Our results show that the new discriminative approach, when configured to use a rich combination of acoustic features, achieves a high retrieval precision at a notable recall under noisy conditions. As compared to HMM-based system, we achieved 54% relative improvement in F-score for explosion recognition with 1.5% relative improvement in F-score for gunfire recognition.