A detection and classification system for sound surveillance is presented. A human/non-human voice classifier is firstly applied to separate the input sound into human voice sound or non-human emergency sound. It utilizes a sliding window Hidden Markov Model (HMM) with trained background, human voice and non-human sound templates. In case of human voice, a scream/non-scream classification is performed to detect screaming in an abnormal situation such as screaming for help during bank robbery. In case of nonhuman sound, an emergency sound classifier capable of identifying abnormal sounds such as gun shot, glass breaking, and explosion, is employed. HMM is used in both scream/non-scream classification and emergency sound classification but with different sound feature sets. In this research, a number of useful sound features are developed for various classification tasks. The system is evaluated under various SNR conditions and low error rates are reported.
[1]
Cheung-Fat Chan,et al.
Improving pitch estimation for efficient multiband excitation coding of speech
,
1996
.
[2]
Augusto Sarti,et al.
Scream and gunshot detection in noisy environments
,
2007,
2007 15th European Signal Processing Conference.
[3]
Nikos Fakotakis,et al.
On acoustic surveillance of hazardous situations
,
2009,
2009 IEEE International Conference on Acoustics, Speech and Signal Processing.
[4]
Chloé Clavel,et al.
Events Detection for an Audio-Based Surveillance System
,
2005,
2005 IEEE International Conference on Multimedia and Expo.
[5]
Janto Skowronek,et al.
Automatic surveillance of the acoustic activity in our living environment
,
2005,
2005 IEEE International Conference on Multimedia and Expo.