Abstract: RGB-T object detection enhances accuracy and robustness in complex environments by fusing complementary information from visible and thermal infrared images. For practical UAV applications, two core challenges arise: ① intra-modality: visible images captured at night or in bad weather suffer severe degradation and detail loss; ② inter-modality: due to varying perspectives, small targets, and complex backgrounds, target information is hard to align during cross-modal fusion, leading to high noise and difficulty in detecting small targets. To address these, a prior-guided improved scheme for UAVs was proposed. To address the intra-modality problem, a pre-trained low-light enhancement prior was used to enhance low-light RGB images in the spatial domain, restoring details. To address the inter-modality problem, a human attention prior was introduced to design a lightweight foreground discrimination branch, which helped the model focus on target regions via multi-task learning, reducing background noise. Experimental results show that the framework achieves robust detection in complex scenarios with varying illumination and multi-scale targets, providing reliable multi-modal detection support for low-altitude intelligent perception.