Domain adaptation for cross-domain alignment in human activity recognition using device-free sensing

Over the past years, Wireless-Fidelity (Wi-Fi) sensor-oriented system networks have gained significant attention for conveying information via radio data transmission due to their ease of deployment on a large scale and high throughput yield. Recently, researchers have explored using Wi-Fi signals to track human motion in laboratory environments. These signals fluctuate in response to movement, creating identifiable activity patterns through channel state information (CSI). Such a sensing methodology is gaining popularity among traditional approaches as it is a contactless sensing strategy without requiring wearable sensors and no need to take the personal identity of the user, thus preserving privacy. Also, there are no constraints for lighting around the area or line of sight (LOS) requirements. This thesis explores both traditional and modern advancements in device-free human activity recognition (HAR), analyzing their strengths and limitations. A significant challenge in this field is the domain-shifting issue, where models trained in one environment fail to adapt to different conditions due to variations in users, surroundings, and atmospheric factors. Existing approaches often lack generalization since models are trained and tested on data collected under similar settings. To address this, our first solution evaluates inter-domain and intra-domain alignment methods to identify which one is better at robustly matching the source to the target domain. The second approach introduces a supervised domain alignment technique, requiring minimal labeled data to improve recognition accuracy in a new environment. Lastly, we present an unsupervised multi-source domain adaptation method, which aligns diverse data distributions with a target domain affected by new users and environmental variations. We validate the proposed methods using three HAR CSI datasets, demonstrating significant improvements in transfer learning tasks compared to state-of-the-art baselines. These findings contribute to making Wi-Fi-based HAR more robust and applicable to real-world scenarios.