Lightweight RT-DETR with Dynamic Optimization and Multi-Scale Attention for Real-Time Traffic Object Detection

To accurately detect multi-scale targets in complex traffic scenarios, traditional Transformer-based object detection models require excessive computational resource consumption due to a large number of parameters, while multi-scale target features pose challenges to the adaptability of traditional Transformer-based target detection models. To address the above issues, this work proposes a detection model integrating a lightweight architecture and dynamic optimization strategies based on the Real-Time Detection Transformer (RT-DETR). Specifically, we design a new backbone network of RT-DETR. To address the problem of large numbers of parameters, standard convolution operations are replaced with Partial Convolution (PConv) to significantly reduce redundant computations through selective channel processing. Then, a FasterBlock module is designed based on a T-shaped receptive field to optimize feature representation efficiency. To further enhance the model's dynamic adaptability, Structural Reparameterization Convolution (RepConv) is incorporated to reduce computational complexity during inference. Simultaneously, an Efficient Multi-scale Attention mechanism (EMA) is developed to strengthen the perception of multi-scale targets in complex scenes through cross-spatial feature interaction. Experiments on a self-built traffic scenario dataset demonstrated the effectiveness of the proposed model. The results show that, compared to the original RT-DETR-R18, the proposed model achieves stable detection accuracy with 15 % fewer parameters and improves accuracy by 2 % in dense-target scenarios while increasing inference speed by 14 %. This research provides a high-precision, resource-efficient solution for realtime object detection tasks in traffic scenarios, advancing the lightweight deployment of Transformer models in practical traffic applications.