Real-time skeleton-based indoor activity recognition

Indoor activity recognition is very important for assistance living applications for the handicapped and the elder people, where real-time response is of primary concern. Conventional approaches rely on pattern comparisons over a period of time-frames often result in delayed decision, which is especially true when more comparison patterns are involved. This paper proposes a state diagram modeling of activities that an occupant can assume in an indoor environment. Each state models one particular kind of activity, which is best characterized by the relative positions and movements of skeleton joints. In particular, six most robust skeleton joints provided by RGB-D cameras like Kinect are employed to detect transitions from one state (activity) to another. The process boils down to a decision tree with branch conditions on relative positions among these six observed skeleton joint data. Our experiments show that the proposed skeleton-based indoor activity recognition approach is capable of recognizing five main indoor environment in real-time with satisfactory results, making the proposed approach promising candidates for context aware services and ambient assistant living applications.