Joint detection and state recognition of traffic lights and signs through a lightweight hybrid CNN–transformer framework

Accurate detection of traffic lights and signs is a critical task for safe autonomous vehicle systems (AVS). Despite advances in deep learning, most methods treat traffic light detection, traffic sign detection, and state recognition as separate tasks, while CNN-based models lack global context modeling and transformer-based approaches incur high computational cost. To address these limitations, this study presents a unified and lightweight YOLOv11–EfficientViT framework that jointly performs traffic light detection, traffic sign detection, and traffic light state recognition within a single end-to-end architecture. The proposed method integrates convolutional block attention modules (CBAM) into the YOLOv11 backbone for enhanced local feature refinement and incorporates an efficient vision transformer (EfficientViT) module to capture global contextual dependencies with reduced computational cost. A multi-task loss formulation is adopted to simultaneously optimize localization and semantic state recognition. Experiments were conducted on curated, task-focused subset of BDD100K dataset for traffic light detection and state recognition and TT100K dataset for traffic sign detection. On BDD100K, the proposed model achieved 99.4% mean average precision (mAP@0.5), 92.2% mAP@[0.5:0.95], and 97.7% traffic light state recognition accuracy under the defined experimental setting while maintaining inference at 3.8 ms per frame. On TT100K, the framework attained 96.5% mAP@0.5 and 81.2% mAP@[0.5:0.95] with only 2.5M parameters. Cross-dataset evaluation on subsets of LISA and Bosch datasets further indicates promising cross-dataset transferability with limited performance variation under domain shift. The results indicate that the hybrid integration of attention-enhanced convolutional features and lightweight transformer-based global reasoning enhances detection performance and semantic consistency without compromising computational efficiency.