LUGA: Lightweight Uncertainty-Guided Sensing Resolution Adaptation for Energy Efficient Radar Processing

Le résumé fourni par la source

Static Frequency Modulated Continuous Wave (FMCW) radar sensing treats all scenes uniformly, resulting in inefficient resource usage given that naturally occurring scenes vary in complexity. Prior adaptive schemes depend on compute heavy online learning methods, leading to high compute cost and latency, which limits edge deployment. Instead, we pair the uncertainty of a learned detector with a lightweight hysteretic policy that adjusts the chirp and virtual antenna counts on the fly, creating an adaptive perception-sensing feedback loop. Evaluated on a public object detection dataset, our approach reduces the inference compute energy by 63% and, when mapped onto a representative TI IWR1443 radar front-end, simulation under the adaptive schedule projects a 23% reduction in sensing energy. This yields an overall 41% reduction in the total energy budget while maintaining competitive accuracy, allowing high-resolution sensing on demand while conserving resources in simpler scenes.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.