PBM_YOLO: An Enhanced YOLOv11 Architecture for Robust Real-Time Traffic Sign Detection
J. Zhang · IEEE Access · 2025
Real-time traffic sign detection is an integral component of environmental perception for modern intelligent driving systems. While one-stage detectors are favored for their high inference speeds, their practical performance is frequently hampered by real-world complexities, including small or distant objects, significant scale variations, and challenging lighting within cluttered backgrounds. Consequently, achieving robust detection of small-scale traffic signs under diverse conditions remains a persistent challenge. This paper introduces PBM_YOLO, a model that significantly enhances the YOLOv11n baseline to overcome these limitations. We integrate an enhanced Parallelized Patch-aware Attention (PPA) module into the backbone to strengthen feature extraction for small-scale targets. Within the neck, a weighted Bi-directional Feature Pyramid Network (BiFPN) is employed to facilitate more effective multi-scale feature fusion. The architecture is further improved by introducing an additional detection head tailored for small object recognition and adopting the Minimum Point Distance Intersection over Union (MPDIOU) loss function to refine the precision of bounding box regression. Evaluated on an augmented version of the Laboratory for Intelligent and Safe Automobiles (LISA) traffic sign dataset, PBM_YOLO demonstrates marked improvements over its baseline. The model achieves an increase of 9.3 percentage points in Precision, 4.6% in mean Average Precision (mAP) at an IoU threshold of 0.5, and 3.1% in [email protected]:0.95. A series of ablation studies confirms the individual and synergistic contributions of the proposed components. As a result, PBM_YOLO presents a robust and accurate framework for traffic sign detection, offering a promising advancement for the development of advanced driver-assistance systems (ADAS).