Knowledge Transfer in Activity Recognition Using Sensor Profile

Most previous work in activity recognition using transfer learning requires at least unlabelled target domain dataset, which is not practical for a smart home system to be deployed in practical use. A better solution to build an intelligent system for a smart home is to collect the dataset in a laboratory environment, and use transfer learning to reduce the effort of data collection. In this work, we represent a knowledge transfer method for activity recognition. Specifically, we define sensor profiles for sensors in the source and the target domain datasets using background knowledge about the sensor networks, and measure the similarity of features between two datasets using these profiles. Graph matching algorithms are adopted to automatically compute appropriate mappings of features based on the similarity measure. This method can be used in data preprocessing procedures, so it can be applied to an existing learning system without affecting the following procedures. The result of our experiment shows that it is possible to transfer knowledge between datasets in activity recognition using our method.