"PanelWrist"

"Power distribution panels are densely populated with switches, meters, and control components, where improper operations may lead to serious grid safety risks. Continuous monitoring of workers' panel interactions is therefore important for preventing unsafe actions and improving operational reliability. Existing vision-based and RF-based solutions achieve strong performance under line-of-sight (LOS) conditions, but their effectiveness drops significantly in the non-line-of-sight (NLOS) situations commonly encountered in practical power environments. To address this problem, we present PanelWrist, a device-free and contactless system that uses a single commercial millimeter-wave radar to enable real-time fine-grained wrist tracking in power distribution rooms. Accurate wrist tracking in NLOS environments is difficult because weak wrist reflections are heavily corrupted by multipath propagation, body interference, and background clutter. To tackle these issues, we first develop a Least Squares Estimation-based Body-Background Signals Removal (LSE-BBSR) algorithm, which suppresses interference from non-target body parts and environmental clutter while preserving wrist-related motion information. We then design a Convolutional Embedding Global Cross Vision Transformer (CE-GCrossViT) to extract and fuse wrist multipath features from radar images. The model adopts two branches with different receptive fields to capture complementary multipath characteristics, employs multi-scale convolutional embedding to learn hierarchical local representations, and introduces a global cross-attention mechanism to strengthen feature interaction across branches. Experiments conducted in real power distribution room environments show that PanelWrist achieves an average tracking error of approximately 2.0 cm under LOS conditions and 3.2 cm under NLOS conditions, significantly outperforming existing wrist-tracking approaches. These results demonstrate that PanelWrist enables accurate fine-grained wrist tracking in occluded power-distribution environments and suggest the feasibility of mmWave sensing for operation-aware safety monitoring."