Interaction of road, traffic and environment in the operational design domain of lane support systems: hybrid factorial–observational design and causal analysis

The Operational Design Domain (ODD) defines the conditions under which automated driving and driver-assistance systems are expected to operate. This study evaluates the ODD of a camera-based Lane Support System (LSS) using direct Mobileye 6.0 lane-detection quality outputs. A large-scale hybrid factorial–observational field design covered 6 different Light × Weather combinations across 9,351 road sections on two-lane rural roads with wide variability in lane-marking retroreflectivity (RL) and road horizontal alignment and cross section characteristics. Statistical and machine-learning classification models were calibrated and compared to analyze the relationships between lane-marking quality and environmental, road, and traffic features. An AutoML-LightGBM pipeline with SMOTE-based class-imbalance treatment achieved the best accuracy of 0.81. SHAP analysis identified low RL, rain, night conditions, narrow lanes, and high curvature as contributors to critical detection conditions. Because standard ML is optimized for prediction rather than causal inference, Double Machine Learning was added to estimate adjusted effects from observational data. Higher RL, higher speed, and wider lanes were associated with better expected Mobileye quality scores, whereas rain, night conditions, and higher curvature were associated with lower detection quality. SHAP dependence and conditional SHAP analyses supported the identification of maintenance-mitigable infrastructure constraints and harder environmental/geometric ODD limits. One practical result is that RL transitions from low-quality detection mainly occur within 120–150 mcd/(m 2 ·lx) across the majority of environmental and physical conditions, although this transition is less evident under sharp curvature or rain.