Lightweight CNN-Based Radar System for Human Localization in Distribution Rooms

This paper proposes a lightweight convolutional neural network (CNN)-based radar system for human localization in distribution rooms, addressing the challenge of spatial safety management in complex power facilities. The core innovation lies in leveraging Ultra-Wideband (UWB) radar signals to capture location-specific multipath effects modulated by human presence in different electrical compartments. The proposed system adopts MobileNetV2 as the backbone, utilizing its inverted residual blocks with linear bottlenecks to achieve high feature extraction efficiency while maintaining a compact model size (<5 MB). By replacing standard convolutions with depthwise separable convolutions, MobileNetV2 significantly reduces computational complexity, enabling real-time inference on edge devices without compromising localization accuracy. Experimental results demonstrate robust performance in cluttered environments, outperforming conventional threshold-based UWB localization and video surveillance methods that suffer from blind spots and privacy risks. This work provides a cost-effective, privacy-preserving solution to mitigate operational errors caused by human misplacement in critical power infrastructure.