Research On Abnormal Audio Event Surveillance In Real-world Scenes

Nowadays,the traditional video surveillance system can no longer meet the demand of security,while audio monitoring technology shows advantages of surveillance system,which can detect dangerous sounds and locate the sound sources effectively.The human activity that takes place in meeting-rooms or class-rooms is reflected in a rich variety of acoustic events,either produced by the human body or by objects handled by humans,so the determination of both the identity of sounds and their position in time may help to detect and describe that human activity.Dangerous situation can be quickly alarmed,if the audio monitoring system can analyze audio information effectively.The key point of audio monitoring is detection of audio events,although there are still some troubles in this area,such as problems of system adaptive and separation of voice sources.This thesis focuses on the methods of abnormal audio events detection.Through the comparison of method of each part,the appropriate algorithm used to structure the system.The system shows high accuracy using only few auditory scenes.Main work and contributions of this thesis are as follows:1.Implementthe abnormal audio detection system,the system consists of four subsystems: quantization and construction of dictionary,retrieval system,construct model of audio event,classification and update.The experimental results show that the system has a miss probability of 6.74% and false alarm probability of 20.91% in public data and real scene data.2.Since extensive work needs to do for label the audio event,an audio retrieval method is proposed.Based on fuzzy quantization,we proposed a significant Gaussian atomic.By introducing the TF-IDF weight of significant Gaussian atomics for each audio event,the index is constructed.And the retrieval system outperforms inverted-index-based method(TSI).The results show that with query duration of 10 seconds,the proposed method achieves the precision of 94.10%.3.Using significant Gaussian atomics to construct the model of audio event,based on theory of Incremental Estimation of GMMs,we proposed a method for updating the audio event's model and dictionary.4.Since the traditional audio endpoint detection technology is only suitable for stationary noise condition,we use the entropy-based method to segment the audio.