Nighttime monitoring in power facilities is challenging due to low illumination and occlusions. Visible detectors often degrade, while thermal imaging remains stable. We introduce LUCID-RTDETR, a lightweight early fusion detector combining pixel-aligned RGB and thermal inputs. The model enhances RT-DETR-R18 with an upsampling block designed to alleviate upsampling-induced artifacts (EUCB-SF) and a content-adaptive cross-scale fusion module (GDPCF). These modules improve small-object recall and edge fidelity without heavy attention mechanisms. On LLVIP, LUCID-RTDETR improves over the RT-DETR baseline by + 3.2 \({\text{AP}}\) and + 1.5 \({\text{AP}}_{50}\) , and repeated-run evaluation shows that this gain is stable across independent random seeds. In zero-shot evaluation on FLIR, the proposed model yields a modest but consistent improvement (+ 0.5 \({\text{AP}}\) ), which should be interpreted as improved robustness under domain shift rather than complete domain invariance. Additional efficiency analysis shows that the proposed model preserves the same parameter count as the baseline, with slightly lower GFLOPs, nearly unchanged latency/FPS, and reduced peak memory usage. Controlled perturbation tests under RGB–thermal misalignment further show slower performance degradation, especially under slight and moderate offsets. Overall, our approach offers a practical accuracy–efficiency trade-off for deployment-oriented safety surveillance. To promote reproducibility and further research, we have made our source code and pre-trained models publicly available at https://github.com/Heisenberg-u/LUCID-RTDETR.
