Activity Recognition of Multiple Subjects for Homecare

Abstract

With consideration of the aging society, the homecare associated with seniors becomes a critical issue in which there is a need of activity recognition of multiple subjects at home. This work adopts a portable panoramic camera locating at the living room to record daily activities which are analyzed, processed, and identified by the background subtraction, generation of Binary Large Object (BLOB), Kalman filter, subject extraction, Gaussian Mixture Model (GMM), majority voting, Finite State Machine (FSM), and so on. Based on the identified individual subjects, each one is extracted by six parameters for GMM to attain the initial activity estimation per picture. The majority voting is fulfilled by a picture segment of 10 pictures, and then the FSM is employed to further confirm the recognized activity and to get rid of the infeasible state transition. Particularly, when the subjects are overlapped, the previous activity state associated with each subject is kept. Until subjects being separated, the activity recognition of each subject is not activated. Additionally, the states of the overlapping interval are fairly estimated based on the states of the boundary picture segments neighboring to the beginning and end of the overlapping interval. In order to increase the accuracy of activity estimations on the boundary picture segments, the incremental majority voting approach is used where its probability model is established and illustrated. The simulation results reveal that the average accuracy of four activities, standing, walking, sitting, and falling, can reach 93.0% during multiple subjects at the room. Compared to the conventional work, our system shows fairly good performance, portability, and convenience for homecare applications.