Fine-Grained Mapping Between Daily Activity Features in Smart Homes

For daily activity recognition in smart homes, it is possible to reduce the effort required for labeling by transferring a trained model. This involves utilizing a labeled daily activity dataset from one smart home to recognize other activities in another. The foundation of this transfer lies in establishing a shared common feature space between the two smart homes, achieved through a feature mapping approach for daily activities. However, existing heuristic feature mapping methods are often coarse, resulting in only moderate recognition performance. In this paper, we propose a fine-grained daily activity feature mapping approach. Sensors are ranked by their significance using the PageRank algorithm, and a novel alignment algorithm is introduced for sensor mapping. Experiments conducted on the publicly available CASAS dataset demonstrate that the proposed method significantly outperforms existing daily activity feature mapping approaches.