Abstract

This paper introduces a new concept of surveillance, namely, audio-visual data integration for background modelling. Actually, visual data acquired by a fixed camera can be easily supported by audio infor- mation allowing a more complete analysis of the monitored scene. The key idea is to build a multimodal model of the scene background, able to promptly detect single auditory or visual events, as well as simultaneous audio and visual foreground situations. In this way, it is also possible to tackle some open problems (e.g., the sleeping foreground problems) of standard visual surveillance systems, if they are also characterized by an audio foreground. The method is based on the probabilistic modelling of the audio and video data streams using separate sets of adaptive Gaus- sian mixture models, and on their integration using a coupled audio- video adaptive model working on the frame histogram, and the audio frequency spectrum. This framework has shown to be able to evaluate the time causality between visual and audio foreground entities. To the best of our knowledge, this is the first attempt to the multimodal model- ling of scenes working on-line and using one static camera and only one microphone. Preliminary results show the efiectiveness of the approach at facing problems still unsolved by only visual monitoring approaches.