1

Abstract—In this paper, we propose a new activity analysis framework to facilitate the independence of older adults living in the community, reduce risks, and enhance the quality of life at home by recognizing activities of daily living (ADLs) by using RGB-D cameras. Comparing to the traditional RGB cameras, the depth information implicitly has advantages of handling illumination changes and protecting privacy. Our contributions include three aspects. First, to detect abnormal activities which are dangerous for elderly people, we recognize 5 activities related to fall including standing, fall from standing, fall from sitting, sit on chair, and sit on floor. Second, to recognize finer activities of daily living, we propose a discriminative representation of structure-motion features based on skeleton joints. Third, to continually track same person when there are multiple people appear in the same camera view, we further develop a binary classification based person identification method by combining appearance and depth information. The proposed framework is evaluated on a dataset we collected under different lighting conditions for fall detection and a benchmark dataset for daily living activity recognition. Experiment results demonstrate the effectiveness of the proposed framework and outperform the state-of-the-art method