Research On Abnormal Audio Event Surveillance

Abnormal audio event surveillance is a new hot topic in the field of security. In public places and government organs and departments, security equipment has been mainly based on video surveillance. But when using video surveillance, security staff have to concentrate on monitoring all the time. Using abnormal audio event detection system combined with video surveillance system is a hot research direction in the future.However, detection of abnormal audio event has several difficulties. Firstly, we haven’t got a effective attributes for such a classification problem. Secondly, for the demand of security, the detection accuracy must be higher than any other field, we set the target detection rate as99%, while the false alarm less than10times per hour. Finally, security devices have real-time requirements, and therefore need to consider the efficiency of the algorithm.This paper describes an abnormal audio event surveillance system which automatically detects abnormal audio events such as screams or gunshots. For the existed problem and difficulties, the paper does research as follows:1. Build an abnormal audio event surveillance system based on GMM-HMM. Try to optimize the parameters of the model through experiments. The experimental results show our baseline system of5-states-16-gaussian-mixtures can detect abnormal audio event with a detection rate of94.4%at a false alarm rate of12per hour.2. Research on attributes and feature selection. In the paper, we extract a190-dimensions feature vector and use a feature selection algorithm of Relief-F to reduce the dimension. The results show this can increase the detection rate to99.5%and reduce the false alarm rate to3.3per hour.3. This paper also try to propose modules for improving performance of audio event detection, such as rejection, self-adaption and likelihood value adjustment. After adding the rejection module, the real-time rate reduce from0.35to0.22. And with the likelihood value adjustment module, we can find an optimal balance between the detection rate and false alarm rate.