| Title | : | Resource-Efficient RF Sensing for Scalable IoT Applications |
| Speaker | : | Yamini Shankar (IITM) |
| Details | : | Wed, 8 Oct, 2025 11:00 AM @ SSB 334 |
| Abstract: | : | Radio Frequency or RF-based sensing offers a privacy-preserving and non-intrusive modality compared to vision based alternatives, enabling perception tasks such as indoor localization, activity recognition and surveillance. RF-sensing is increasingly pervasive, with applications spanning everyday environments. To be practical at scale, such sensing must rely on portable IoT devices that operate under tight hardware and energy constraints. These devices extract raw PHY-layer estimates, for instance, WiFi channel state information (CSI), and run learning-based algorithms to carry out perception tasks. However, two key challenges arise. First, state-of-the-art learning models are resource-heavy, and offloading them to the edge not only increases energy usage but also introduces network contention in streaming sensing data continuously. Second, indiscriminate sensing on the device without task awareness produces redundant measurements, resulting in battery drain. Our work addresses these challenges with the goal of reducing both energy and network overheads in RF sensing. First, we design AutoCompress, a system that reduces the volume of sensing data (CSI spectrograms) transmitted from IoT devices to the edge. As the number of sensing devices scales, network contention grows non-trivially, degrading the performance of regular data traffic/QoE. AutoCompress tackles this by employing model-aware compression, inspired by compressive-sensing primitives, to select informative time-frequency blocks from CSI spectrograms that maximizes inference accuracy. We evaluate the system on both our in-house testbed and public datasets. AutoCompress reduces sensing data transmissions by up to 91%, improves throughput by 35.5% and lowers latency by 27.5%, all without sacrificing inference accuracy at the edge. Second, we develop LowFi, a system for energy-efficient WiFi localization through context-aware scanning. Conventional approaches trigger a fixed number of channel scans, disregarding the environment. While more scans can improve data fidelity, they also drain battery and increase delay. Crucially, the number of measurements required for reliable localization is context-dependent: multipath-rich environments may demand more scans, whereas line-of-sight regions or locations close to access points often need far fewer. LowFi dynamically orchestrates WiFi scans and access point selection to achieve energy-efficient localization. Across multiple IoT platforms and diverse indoor environments, LowFi achieves sub-meter accuracy while cutting energy cost by up to 65%. |
