Purpose: Reliable and quantitative assessment of road marking degradation is paramount for traffic safety and the operational robustness of autonomous vehicle (AV) systems, which rely heavily on visual contrast. Traditional inspection methods are slow, subjective, and fail to provide the high-resolution, continuous data required for modern maintenance planning. This study addresses this gap by proposing a novel deep learning framework for the precise quantification of road marking wear from Mobile Laser Scanning (MLS) reflectance imagery.

Methods: We introduce the Percentage of Residual Marking (PRM)-Enhanced Detector (PRMED), an end-to-end deep learning model based on an EfficientNet backbone integrated with a Feature Pyramid Network. Crucially, the architecture incorporates a dedicated PRM Regression Head that directly predicts the continuous wear percentage (0.0 to 1.0) for each detected marking instance, bypassing the computational complexity and error propagation of a sequential segmentation-then-calculation pipeline. The model was trained and validated on a synthesized dataset derived from MLS data, which accurately represents a full spectrum of real-world degradation states.

Results: The PRMED model achieved a high detection accuracy, registering an $mAP@0.5$ of 0.94 and significantly outperforming a two-stage segmentation baseline in quantitative wear assessment. Specifically, the model demonstrated a Mean Absolute Error (MAE) for PRM prediction of only 1.85%, which is critical for establishing objective maintenance thresholds. Inference speed was confirmed to be suitable for real-time mobile deployment.

Conclusion: The proposed multi-scale, end-to-end deep learning framework provides a robust, efficient, and objective solution for road marking wear assessment. The continuous PRM metric offers a crucial data point for infrastructure managers to optimize maintenance schedules and, more importantly, to ensure the consistent functional integrity of perception systems in autonomous vehicles.