Acoustic sensor based activity recognition using ensemble of one-class classifiers

In this paper we address the problem of human activity recognition based only on acoustic modality. The ultimate goal is continuous acoustic monitoring of public places like parks and bus stops for detecting littering activities so that the people involved in such acts can be prompted to bin appropriately. We exploit the fact that when human interacts with objects, a characteristic sound is produced, and this sound can be used to recognize the activity. We propose a method based on perceptual features and ensemble of fuzzy rule-based one-class classifiers for activity recognition. The method is validated using real data and compared with support vector machine classifier. The results show that the classifier has very low false alarm rate and potentially well suited for incremental learning.