Sound learning–based event detection for acoustic surveillance sensors

This study proposes an event detection technique for acoustic surveillance that detects emergency situations by using acoustic sensors. Most surveillance systems have widely depended on visual data recorded by closed-circuit television (CCTV) cameras, but more intelligent systems are now beginning to use audio information for more reliable detection of emergency situations. Most of the conventional studies on acoustic event detection adopt limited types of acoustic data and are based on simple algorithms, such as energy-based determination. Thus, these approaches are easily realized, but may induce serious detection errors in real-world applications. In this study, we propose an event detection technique based on a sound-learning algorithm to be adopted by real-time acoustic surveillance systems. One main process of this technique is to construct acoustic models via learning algorithms from sound data collected according to types of acoustic events. The models are used to determine whether audio streams entering an acoustic sensor refer to the events or not. In event detection experiments performed in an outdoor environment, the proposed approach outperformed conventional approaches in the real-time detection of acoustic events.