Abstract:To address the problem of reduced traffic sign detection accuracy in intelligent driving systems under nighttime low-light conditions due to strong noise interference and the loss of critical features, this paper proposes a low-light traffic sign detection and recognition algorithm based on an improved RT-DETR framework. In the feature extraction stage, a lightweight RepViT network is introduced to reconstruct the backbone network, thereby enhancing the feature learning capability for low-resolution and blurred targets. In addition, a Low-light Small-target Enhancement Module (LSEM) is designed, which combines channel-level feature recalibration with spatial multi-scale perception to effectively emphasize high-response channels while suppressing noise in dark regions. Furthermore, an Asymmetric Directional Feature Enhancement Module (ADFE) is incorporated at the decoder input stage. By employing large-kernel strip convolutions to expand the directional receptive field, the module reduces the interference of local dark noise on edge features and improves the perception of geometric boundaries preserved by reflective materials. Experimental results on the CNTSSS nighttime traffic sign dataset demonstrate that the proposed model reduces the number of parameters and computational cost by 30.7% and 27.8%, respectively, while improving the mAP50 by 2.1 percentage points to 90.2%. The proposed method achieves favorable detection performance in nighttime traffic sign detection tasks.