Detecting abnormal events in multimedia sensor networks (MSNs) plays an increasingly essential role in our lives. Once video cameras cannot work (e.g., the sightline is blocked), audio sensor can provide us with critical information (e.g., in detecting the sound of gun-shot in the rainforest or the sound of car accident on a busy road). Audio sensors also have price advantage. Detecting abnormal audio events in complicated background environment is a very difficult problem; only few previous researches could offer good solution. In this paper, we proposed a novel method to detect the unexpected audio elements in multimedia sensor networks. Firstly, we collect enough normal audio elements and then use statistical learning method to train them offline. On the basis of these models, we establish a background pool by prior knowledge. The background pool contains expected audio effects. Finally, we decide whether an audio event is unexpected by comparing it with the background pool. In this way, we reduce the complexity of online training while ensuring the detection accuracy. We designed some experiments to verify the effectiveness of the proposed method. In conclusion, the experiments show that the proposed algorithm can achieve satisfying results.
展开▼
