Enhancing audio surveillance with hierarchical recurrent neural networks

The need for effective and reliable surveillance techniques is getting nowadays more and more of primary importance, especially in the actual scenario in which safety and security have become a priority. While classical techniques rely on video-based surveillance systems, such as Close-Circuit television, many studies show that also the audio signal can be effectively used for these purposes. There are many characteristics that make the audio signal particularly suited for this task and, above all, the fact that the analysis of the audio signal can greatly improve thanks to the introduction of automatic classification. Recently, a large focus has been on the use of Deep Neural Networks for classifying audio data and, in this work, we aim to test their performance in the audio surveillance field. In this contribution we propose an algorithm for audio events detection in noisy environments based on the use of deep recurrent neural network. The achieved results show satisfactory and improved performances with respect to state-of-the-art techniques.