LR-DETR: a lightweight real-time traffic sign detection model based on improved RT-DETR

Real-time detection of traffic signs is crucial for safe autonomous driving and intelligent transportation systems. The current key challenge in the field of traffic sign detection is to achieve model lightweighting and improve real-time performance while maintaining effectiveness. To address these challenges, this paper proposes a lightweight real-time traffic sign detection model called LR-DETR. LR-DETR is based on the end-to-end object detection model of RT-DETR. By redesigning the core modules in RT-DETR, LR-DETR significantly improves the lightweighting level of the model while maintaining high detection accuracy. A PCIR-Block module based on Partial Convolution and Inverted Residual Structure is proposed to more fully extract multi-scale features in the backbone network. A Context-Guided Feature Fusion Module (CGFFM) is proposed to utilize contextual information between features of different scales to enhance the effectiveness of feature representation and subsequently improve the fusion performance of multi-scale features. In addition, with the help of dilated convolution and reparameterization techniques, LR-DETR designed a DRBC3 module for feature re-extraction to further enhance the model’s ability to capture features at different scales, while effectively reducing the number of parameters and floating-point operations. The experimental results on the CCTSDB 2021 dataset show that compared with the state-of-the-art baseline models, LR-DETR performs better in increasing the value of mAP@0.5 by 0.5%, decreasing the value of FLOPs by 28.1%, decreasing the value of Params by 23.7%, and decreasing the value of Latency by 11.6%. The experimental results on the TT100K traffic sign dataset show that the precision and recall of the LR-DETR model are 87.1% and 81.6%, respectively, outperforming other baseline models. LR-DETR significantly reduces the number of parameters and floating-point operations while maintaining high detection performance, improving the detection speed of the model. This will provide a constructive contribution to achieving real-time detection of traffic signs.