Virtual Lane Generation for Unmarked Roads via Adaptive Segmentation and MPC Steering Control

Abstract

Lane Keeping Assist (LKA) systems are rendered unreliable on rural, degraded, or unmarked roads where traditional lane detection pipelines fail. This paper proposes a simulation-based LKA framework that operates without physical lane markings by integrating DeepLab v3+ semantic segmentation with a novel Adaptive Morphological Feedback Loop (AMFL) and Model Predictive Control (MPC). The AMFL dynamically adjusts post-processing morphological kernel size in response to BFS reachability failures, reducing overall path generation failure from 11.1% (static baseline) to 1.5% across diverse scene conditions. The segmentation network, trained on a curated 312-frame annotated subset of approximately 10,000 collected rural driving frames over 585 iterations, achieves a test-set mean IoU of 74.37%, a road-class F1 score of 80.54%, and a training pixel accuracy of 89.24% at convergence. Virtual lane boundaries derived via BFS pathfinding and cubic spline interpolation feed an MPC lateral controller grounded in a linear bicycle model. Closed-loop simulation on the winding P7229 road near Ifrane, Morocco demonstrates a 12.2% reduction in mean lateral deviation (0.422 m vs. 0.481 m) and a 40.9% reduction in mean relative yaw error (6.87° vs. 11.63°) compared to unassisted manual driving, and a 22.8% reduction in mean lateral deviation compared to a Stanley geometric baseline controller, with lane departure events reduced by 83%. These results confirm the viability of perception-driven lane keeping on infrastructure-poor unstructured roads.