Pitch-Range Based Feature Extraction for Audio Surveillance Systems

Most security systems detect abnormal events using visual clues. However in some cases, we can obtain more accurate and efficient data from audio information. This study analyzes the acoustical properties of audio signals for audio surveillance systems and presents a novel feature extraction method. The purpose is to detect unusual and unsafe sounds such as gunshot, dog barking, and breaking glass using pitch range based feature parameters. Support Vector Machines (SVMs) are used as a recognition algorithm to evaluate the performance of the proposed feature parameters. Recognition rates are found in the range of 79 to 92%.