The fundamental problem of the existing Activity Recognition (AR) systems is that these are not generalpurpose. An AR system trained in an environment would only be applicable to that environment. Such a system would not be able to recognize the new activities of interest. In this paper we propose a General-Purpose Activity Recognition System (GPARS) using simple and ubiquitous sensors. It would be applicable to almost any environment and would have the ability to handle growing amounts of activities and sensors in a graceful manner (Scalable). Given a set of activities to monitor, object names (with embedded sensors) and their corresponding locations, the GPARS first mines activity knowledge from the web, and then uses them as the basis of AR. The novelty of our system, compared to the existing general-purpose systems, lies in: (1) it uses more robust activity models, (2) it significantly reduces the mining time. We have tested our system with three real world datasets. It is observed that the accuracy of activity recognition using our system is more than 80%. Our proposed mechanism yields significant improvement (more than 30%) in comparison with its counterpart.

No abstract

No abstract