Deep Human Activity Recognition
Human Activity Recognition (HAR) identifies human movement, such as standing and talking, through sensor data.
Omni-directional antennas can sense human movements by changing signals from multipaths.
- Analyzes the patterns of changes in wireless signals, extracts unique features, and classifies the user's behavior.
We are aiming to separate multipaths at the receiver side to extract more information from the CSI. When a user changes his/her position and orientation, the multi-paths from Tx to Rx alter, which affects the changing pattern in wireless signals. Acquiring each signal from multiple propagation paths (i.e., multipath) means that we can track multiple moving points of human activity.
Domain Adaptation for Robustness
The multipaths completely change in new environments. It is impossible to collect data every time to account for every change in the surrounding. To mitigate the problem, we are developing DL models with domain adaptation that work with a relatively small target dataset.