RDL-YOLO: an efficient algorithm for detecting road surface defects.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Accurate road surface defect detection is crucial for autonomous driving safety and efficient road maintenance.
• Existing lightweight models struggle with complex conditions like varied illumination and multi-scale defects.
• Deployment on edge devices requires models with reduced computational complexity.

Purpose of the Study:

• To propose RDL-YOLO, an efficient and lightweight road defect detection framework.
• To enhance feature extraction, robustness in low-light, and detection accuracy for potholes and protrusions.
• To achieve real-time inference on resource-constrained platforms for intelligent transportation systems.

Main Methods:

• Built upon the YOLO11n architecture, integrating C3k2_RVB and C3k2_REVB modules for multi-scale feature extraction.
• Introduced a shared detection head LSCD_LQE for joint optimization of classification and localization accuracy.
• Incorporated the DySample module for adaptive feature upsampling and DWConv to reduce computational overhead.

Main Results:

• RDL-YOLO demonstrated performance gains: 0.63% precision, 4.24% recall, 1.82% F1-score, and 1.76% mAP@0.5 improvement over YOLO11n.
• Significantly reduced computational complexity: 44% decrease in GFLOPs and 25% reduction in parameter count.
• Showcased robust cross-dataset generalization capabilities in diverse road conditions.

Conclusions:

• RDL-YOLO offers an effective and lightweight solution for real-time road defect detection.
• The model enhances safety for autonomous vehicles and supports preventive road maintenance strategies.
• The framework is suitable for deployment on edge devices in intelligent transportation systems.