A bottom-up audio attention model for surveillance

This paper proposes a bottom-up audio attention model based on spatial audio cues and sub band energy change for unsupervised event detection in stereo audio surveillance. Firstly, the spatial audio parameter Interaural Level Difference (ILD) is extracted to calculate and represent the attention events, which are caused by rapid moving sound source. Then the sub band energy change is computed to present the salient energy distribution change in frequency domain. At last, an environment adaptive normalization is used to assess the normalized attention level. Experimental results demonstrate that the proposed audio attention model is effective for audio surveillance event detection.