License Plate Recognition for Smart Construction Sites Based on GMH-YOLO

With the rapid urbanization, vehicle management at construction sites has become crucial for safety and logistics efficiency. However, License Plate Detection in such environments faces unique challenges, such as simultaneous detection of dual plates and body plates, strong light reflections, and muddy interferences. This paper constructs a license plate dataset, CSLPD, specifically for construction sites, containing 1495 images and 2301 license plate instances, with double and body license plates accounting for 27.2% and 25.4%, respectively. To enhance detection performance in complex environments, this paper proposes the GMH-YOLO model, which integrates an innovative Gated Multi-Head Attention mechanism into the C3k2 module of YOLO11. This lightweight gating unit adaptively allocates feature channel resources, effectively enhancing key information while suppressing background interference, making it particularly suitable for detecting multiple license plate types and partially occluded plates in complex construction site environments. Experimental results show that GMH-YOLO achieves 93.3% mAP@50 on the CSLPD dataset, outperforming YOLO11 by 1.4%. For the challenging body license plate task, detection accuracy improves from 81.3% to 87.2%, a 5.9% increase. The model maintains high real-time performance due to the optimized gating mechanism. Comparative experiments with six attention mechanism integration schemes confirm that the gated mechanism provides the best balance between feature extraction and computational efficiency, offering a high-precision, efficient solution for intelligent license plate recognition at construction sites.