Space-time flexible kernel for recognizing activities from wearable cameras

Recognizing activities of daily living is useful for ambient assisted living. In this regard, the use of wearable cameras is a promising technology. In this paper, we propose a novel approach for recognizing activities of daily living using egocentric viewpoint video clips. First, in every frame, the appearing objects are detected and labelled depending if they are being used or not by the subject. Later, the video clip is divided into spatiotemporal bins created with an object-centric cut. Finally, a support vector machine classifier is computed using a spatiotemporal flexible kernel between video clips. The validity of the proposed method has been proved by conducting experiments in the ADL dataset. Results confirm the suitability of using the space-time location of objects as information for the classification of activities using an egocentric viewpoint.