Exploring the Impact of Locations and Activities in Person-Wise Data Mismatch in CSI-based HAR

Over the past decade, research has demonstrated the potential of Wi-Fi Channel State Information (CSI) in Human Activity Recognition (HAR). However, its real-world implementation is lacking due to the inability of CSI-based HAR to generalize across different domains (persons, locations, etc.). This inability of CSI can be attributed to the dynamic nature of CSI, leading to the issue of data mismatch. Therefore, in order to efficiently employ CSI-based HAR in real-world applications, a comprehensive understanding of the interplay between various data mismatch domains is essential. In that direction, the presented work aims to gain analytical insights into the impact of varying locations and activities in the personwise data mismatch in realistic scenarios. To understand the person-wise data mismatch, three different analysis types namely subject-specific, mixed-subject, and generic-subject were defined. To assess the impact of locations in person-wise data mismatch, two activity locations and four receiver locations were considered. Whereas to assess the impact of the type of activities, four different activity sets, including full-body activities, fine-grained hand and leg activities, only fine-grained hand activities, and a mix of all activities, were evaluated. F1 score degradation by 43% for full-body activities, 72% for only fine-grained hand activities, and 76% for a mix of all activities in the person-wise domain indicate that person-wise data mismatch has a significant impact on the performance of CSI-based HAR. Furthermore, the impact of receiver location and activity location varied based on the activity set but was found comparatively insignificant when observed individually.