LAR-TSDETR: a lightweight adaptive robust traffic sign detection transformer with multi-scale fusion for real-time recognition under challenging conditions

Abstract Traffic sign detection plays a pivotal role in autonomous driving and intelligent transportation systems, where complex driving scenarios pose significant challenges to detection accuracy and model robustness. To address these issues, this paper proposes the lightweight adaptive robust traffic sign detection transformer (LAR-TSDETR), a novel transformer-based object detection framework that is both lightweight and highly resilient to environmental disturbances. Built upon the RT-DETR architecture, LAR-TSDETR integrates three key modules: an adaptive multi-scale feature enhancement module dynamically expands receptive fields and aggregates global context, enhances feature representation capability; an efficient multi-scale fusion enables efficient cross-scale feature interaction by fusing semantic and spatial information across multiple resolutions; a multi-scale hierarchical fusion block employs a multi-scale feature perception mechanism to adaptively focus on critical regions and enhance small object detection performance. Experimental evaluation on the CCTSDB dataset demonstrates that LAR-TSDETR achieves 79.2% mAP@50 and 50.1% mAP@50–95, representing improvements of 2% over RT-DETR in both metrics. Through architectural optimization, the proposed method reduces model parameters by 2.77 M and computational complexity by 7.5 GFLOPs while achieving real-time performance at 69 FPS. Furthermore, generalization experiments validate the model’s robustness across diverse datasets, where LAR-TSDETR attains 82.8% and 95.2% mAP@50 on TTK100 and GTSDB datasets respectively, demonstrating improvements of 1.1% and 1.6% over the baseline and confirming excellent generalization capabilities. These results demonstrate the model’s optimal equilibrium between detection accuracy, computational efficiency, and cross-domain adaptability, establishing its practical viability for real-world autonomous driving deployment.