Monitoring daily activities of a person has many potential benefits in pervasive computing. These include providing proactive support for the elderly and monitoring anomalous behaviors. A typical approach in existing research on activity detection is to construct sequence-based models of low-level activity features based on the order of object usage. However, these models have poor accuracy, require many parameters to estimate, and demand excessive computational effort. Many other supervised learning approaches have been proposed but they all suffer from poor scalability due to the manual labeling involved in the training process. In this paper, we simplify the activity modeling process by relying on the relevance weights of objects as the basis of activity discrimination rather than on sequence information. For each activity, we mine the web to extract the most relevant objects according to their normalized usage frequency. We develop a KeyExtract algorithm for activity recognition and two algorithms, MaxGap and MaxGain, for activity segmentation with linear time complexities. Simulation results indicate that our proposed algorithms achieve high accuracy in the presence of different noise levels indicating their good potential in real-world deployment.
- Activity recognition and segmentation
- Web mining
- Knowledge engineering
- Relative term weighting