Audition based event detection and classification for vehicular surveillance

Currently, sensors embedded in automobiles for surveillance purpose are mainly composed of video based and/or acceleration based sensors. However, there are surveillance situations wherein image and acceleration information obtainable from vehicles may not be sufficient for recognizing surrounding abnormal situations. Such applicable events considered in this investigation include screaming, police car sirens, and noise of glass breaking observable in both driving and parked scenarios. In this paper, a two-step event recognition system is proposed. The first step is event detection through a threshold in high-frequency acoustic band, based on an assumption that the events considered here have greater acoustic energy concentrated in the high-frequency region compared to the background noise coming from the vehicle engine. The next step is to precisely classify the detected event robust from surrounding false positives. We present a set of features which are beat spectrum, spectral flux and their combinations. To enhance the classification performance, we employed confidence measures derived from time normalized log-likelihood score.