Towards improving feature extraction and classification for activity recognition on streaming data

An Activity Recognition system on streaming data must analyze the drift in the sensing values and, at any significant change detected, decide if there is a change in the activity performed by the person. The performances of such system depend on both the Feature Extraction (FE) and the classification stages in the context of streaming data. In the context of streaming and high imbalanced data, this paper proposes and evaluates three FE methods in conjunction with five classification techniques. Our results on public smart home streaming data show better performances for our proposed methods comparing to the state-of-the-art baseline techniques in terms of classification accuracy, F-measure and computational time. Test on Aruba Database show an improvement in term of accuracy and computation time of the results when using the proposed method, using a KNN-based classifier (both around 87% of correct classification but with a largely higher computing time for SVM).