| Title | : | IoTzing Wireless Sensing Applications |
| Speaker | : | Manoj Kumar Lenka (IITM) |
| Details | : | Mon, 25 Nov, 2024 3:00 PM @ SSB 334 Turing Hall |
| Abstract: | : | In this work, we focus on making wireless sensing systems more energy and bandwidth efficient – essentially IoTzing them. We explore various wireless sensing modalities in current literature, focusing on their energy consumption, bandwidth requirements, and sensing performance. We compare edge-based and on-device sensing, emphasizing the need for optimization. To this end, we address two major topics: on-device Wi-Fi CSI-based sensing and energy and bandwidth aware video compression using mmWave. In the first contribution, we introduce a framework called Wisdom, designed to provide an optimal model based on user requirements and system constraints. We formulate a utility maximization problem that minimizes cost (in terms of memory and energy) while maximizing performance (in terms of accuracy and inference time). The framework chooses the optimal architecture, size, and compression technique for a neural network to meet these goals. Although our primary focus is on Wi-Fi CSI-based applications, the framework is extendable to other sensing modalities. The second contribution is EcoVis, a system aimed at compressing surveillance footage in an energy and network aware manner using mmWave radar. This system provides three key solutions. First, it uses mmWave to detect objects of interest and control the camera’s sleep cycle, turning it on or off as needed, which significantly reduces energy consumption. Second, it identifies regions of interest (RoIs) in video frames based on mmWave range-azimuth (RA) maps, transforming these coordinates into the camera’s field of view to focus on relevant areas for video compression. Tiling and applying quantization parameters based on these RoIs further integrates with the H.265 encoder for optimized video compression. Finally, for applications such as vehicle or pedestrian detection, where visual input might not be necessary, the system can rely solely on mmWave data. Since RA maps are much smaller than video data, this can be done on-device, resulting in considerable savings in inference time and network bandwidth. |
