Multi frame size feature extraction for acoustic event detection

This paper addresses the problem of detection and recognition of impulsive sounds in surveillance system, such as door slams, footsteps, glass breaks, gunshots and human screams. We build an acoustic event dataset of about 1k sound clips and a ground truth dataset of a surveillance system. We investigate the influence of different frame size in audio feature extraction when classify acoustic events and our result show that the classification accuracy differs from different audio frame sizes. Based on this result, we propose an approach to integrate multi frame size features to generate a new feature set, which can achieve better performance. We build an abnormal acoustic event detection system for surveillance using this feature and adopt a smoothing post process. The experiments show the effectiveness of our proposed approach.