Lane markings are critical infrastructure for road safety and autonomous driving, yet they deteriorate over time and require systematic monitoring. this thesis proposes a deep learning based framework for automated lane marking damage analysis using d...
Lane markings are critical infrastructure for road safety and autonomous driving, yet they deteriorate over time and require systematic monitoring. this thesis proposes a deep learning based framework for automated lane marking damage analysis using drone captured Bird's Eye View (BEV) images. The framework integrates YOLOv11 for detecting straight and curved lane markings, YOLOv11-seg for pixel level segmentation of curved lanes, and Real-ESRGAN for super resolution restoration to enhance degraded inputs. For damage quantification, binary mask analysis was applied to straight lanes, while a polynomial curve fitting based Region of Interest (ROI) method was introduced for curved lanes to overcome the limitations of sliding window approaches. Experimental results show that YOLOv11 achieved a mean Average Precision at IoU threshold 0.5 (mAP0.5) of 91.0%, and YOLOv11-seg reached a mean Intersection over Union (mIoU) of 95.63%. Restored images improved visual clarity with a Peak Signal to Noise Ratio (PSNR) 24.05 dB and Structural Similarity Index Measure (SSIM) of 0.462. These findings demonstrated that the proposed framework enables reliable and consistent lane marking damage assessment, supporting efficient road maintenance and enhancing the reliability of autonomous driving systems.