Research On Audio Event Detection For Audio Surveillance

Traditional safety monitoring system is based on video, and mostly the video is used to be reviewed afterward. In recent years, audio surveillance has been researched by many researchers. Because of its nice real-time performance and the auxiliary effects to video monitoring, audio surveillance has important research and application value. Audio surveillance is based on audio event detection. There are some difficult problems in this area, such as it’s very hard to find effective features, different audio events may be overlapped, training data may be unbalanced, and so on. In this paper, we intend to detect four audio event categories, including footsteps, sound of breaking glasses, sound of opening or closing door and speech. Given the important effect of feature in a pattern recognition system, we have done some researches to solve the difficult feature problem. These researches are mainly in three aspects below:1. Set up and optimized the audio event detection system based on SVMIn this paper, we used support vector machine as the classifier and built the baseline system based on MFCC feature. Then we chose two kinds of smoothing algorithms to improve the system performance, but the result was not well. We designed another smoothing algorithm which reduced a lot of output fragments. The overall accuracy increased to51.8%from35.0%, and the recall rate increased to86.1%from82.7%. We noticed that the accuracy and recall rate got rather different between categories, then found that this was because of the unbalance of the training data. We performed some random under-sampling experiments and determined the training data ratio.2. Researched a large number of audio features, proposed a new audio feature called Amplitude Interval RatioWe researched a lot of audio features, and summarized methods to design features. Inspired by sub-band energy ratio feature, we propose a new feature called amplitude interval ratio, and we used these ratio values to calculate the information entropy. We performed some experiments to test the effectiveness of the new feature. The result was not good when using the new feature alone. But when we combined the new feature and MFCC feature, the system accuracy improved. So these new features still had certain effects.3. Researched PCA and LDA feature transformation algorithms, did some improvements to LDA algorithmUsing PCA and LDA feature transformation algorithms, we converted original features of high dimensions to new feature spaces of low dimensions. We expected that we could separate different categories more effectively in new feature space. Results of experiments showed that the performance of LDA was better than PCA. Then we did several improved LDA algorithm. Original LDA algorithm could neglect the distance of two different categories, and using category sample count as sum factor could result in that the category with much more samples dominated the result. So we did three methods to improve that. The accuracy and the recall rate got higher in the improved algorithms. And also AED-ER got smaller.