Acoustic based abnormal event detection using robust feature compensation

In addition to traditional video-surveillance applications, the use of acoustic surveillance is becoming increasingly important. The acoustic clues are appropriate for detecting target that are hidden or in a dark location and assumed to make a sound. In the conventional acoustic surveillance systems, the false alarms tend to frequently occur because the conventional acoustic surveillance systems are not robust against the interferences of other sounds. In this study, in order to overcome the conventional difficulty, we apply the Hidden Markov Model (HMM) based robust recognizer with feature compensation to the detection of glass destruction sounds. Experimental results confirmed that the feature compensation method is effective not only to speech signals but also to the material sounds like the glass destruction sound.