摘要
The present study describes a practical methodology for automatic space monitoring based solely on the perceived acoustic information. Our approach is based on a two stage recognition schema that detects and classifies sound events related to hazardous situations. The main objective is to identify on time abnormal events that may lead to life-threatening situations or property damage and forward detected sound events to an authorized officer for further evaluation of the cases. We consider the case where the atypical situations of screams, explosions or gunshots take place in a metro station environment. For describing the audio signals we constructed a feature vector which includes the Mel-frequency cepstral coefficients and three MPEG-7 low level descriptors. These are subsequently fed to hidden Markov models towards representing each sound category. The accuracy of the proposed method is tested under several SNR conditions.
- 出版日期2011-2