RT-DETRmg: a lightweight real-time detection model for small traffic signs

In intelligent transportation systems, real-time detection performance and accuracy are essential metrics. This paper proposes a lightweight real-time detection model, RT-DETRmg, to address the challenges of false and missed detections of small traffic signs and to improve the algorithm's real-time performance. RT-DETRmg enhances the multi-scale feature extraction capability of the RT-DETR backbone network by incorporating a Multiple Scale Sequence Fusion module, which effectively integrates global and local semantic information from different scales of images. Additionally, a cascaded group attention module is utilized within an efficient hybrid encoder to reduce computational complexity, thereby enhancing real-time performance. To further optimize small object detection, a small receptive field feature layer is introduced, while a large receptive field feature layer is removed. Experimental results on the TT100K and GTSDB datasets demonstrate the superiority of RT-DETRmg over existing models. On the TT100K dataset, RT-DETRmg achieves a 2.0% improvement in mean average precision and a 6.6% increase in frames per second compared to the baseline RT-DETR model, while reducing model parameters and computational complexity. On the GTSDB dataset, RT-DETRmg further demonstrates its strong generalization ability, achieving a 2.2% improvement in the F1 score and a 1.7% increase in mean average precision compared to the baseline network. These findings highlight the effectiveness of RT-DETRmg in enhancing both detection accuracy and real-time performance of small traffic signs in diverse scenarios.