Description
Human Activity Recognition classifies a window of sensor data — here CSI or another wireless modality — into one of a set of body-scale activities (walking, sitting, falling, standing). It is the most-studied downstream task in WiFi sensing and the canonical benchmark for evaluating new feature representations and learning techniques. The thesis treats HAR as an upstream test: any CSI representation that fails on HAR is unlikely to support the harder crowd-dynamics tasks downstream.
When it's used
- Benchmark for new CSI feature pipelines
- In-home health monitoring (fall, immobility, activity-of-daily-living)
- Pretext task for
transfer-learning/self-supervised-learningevaluations
Limitations
- Activity vocabularies are dataset-specific and small
- Environment shift causes severe accuracy drops
- Overlap between activities (sit-to-stand, fall vs lie-down) is poorly handled