Abstract

An automatic surveillance system, based on event detection in the video image can be improved by implementing algorithms for audio analysis. Dangerous or illegal actions are often connected with distinctive sound events like screams or sudden bursts of energy. A method for detection and classification of alarming sound events is presented. Detection is based on the observation of sudden changes in sound level in distinctive sub-bands or in parameter values. Among the parameters, there are specially defined features connected with the energy ratio in certain sub-bands of the spectrum and the shape of the signal around transients. The parameter set is completed with MPEG-7 descriptors chosen on the basis of experiments and statistical analysis. For classification a Support Vector Machine classifier is implemented. The model is built using a test set of sounds recorded in real conditions. Separate classifiers are implemented for different classes of sound events. The length of the analysis frame and threshold values for event detection are set to fit the characteristics of a certain type of sound. The accuracy is then improved by the decision procedure. The final indication of the system is derived from decisions of all classifiers, compared in a number of adjacent analysis frames. The classifier yields high accuracy of detecting typical alarming sound events like gunshot, scream, explosion, broken glass or horn abuse. It also provides the possibility to retrain the model and add a new type of event to be recognized by the system. The described solution can be implemented in an automatic surveillance system together with image analysis. Processing both sound and image leads to a significant improvement of the event detection rate. The developed algorithms can be implemented to detect dangerous events in large public areas like stations, airports or stadiums.

Authors (3)