Assessment of Lane Marking Retroreflectivity Using Vision-Based ADAS Data: A Double Machine Learning Approach

Abstract Continuous monitoring of pavement marking quality is essential to ensure lane visibility and guidance for both human drivers and advanced driver assistance systems (ADAS), while standard photometric surveys are costly and infrequent. In such a framework, this study assessed whether the availability of vision-based ADAS lane keeping data can support scalable, maintenance-oriented marking monitoring when calibrated against standard measurements. A multi-source dataset was developed by spatially fusing camera lane keeping data, certified section-level retroreflectivity statistics from the mobile photometric survey, roadway geometry features, and lighting and weather labels across 476 km of two-lane rural roads. The conservative segment-level indicator R L 3 was subsequently calculated as average retroreflectivity (R L avg) minus three within-segment standard deviations. The analysis combined double machine learning (DML) to identify causal relationships between ADAS lane detection scores (Qmin) and R L 3, and double-binary CatBoost to classify R L 3 in three classes defined by thresholds of 90 and 195 mcd/m²/lx. Results showed condition-dependent performance and classification accuracy (F1 score of 0.83) when repeated measures were taken under different lighting and weather conditions.