Hybrid Lane Detection and Turn Prediction Framework using U-Net-based Lane Marking Visibility and Geometric Curve Analysis

Autonomous driving systems rely heavily on accurate lane detection and turn prediction for safe and reliable navigation. However, faded, occluded, or inconsistent lane markings present significant challenges, especially under varying road conditions. To address these issues, this paper introduces a hybrid framework that integrates deep learning with classical vision techniques to enhance lane perception and directional understanding. U-Net, a convolutional neural network architecture, is employed to perform semantic segmentation for lane marking visibility, ensuring robust detection even in degraded scenarios. For straight lanes, a region of interest is extracted, and the Hough Transform is applied to identify solid and dashed lanes based on line continuity and slope filtering. Curved lane detection is achieved by generating a bird’s-eye view using homography, followed by binary and HSV thresholding to isolate white and yellow lanes. Polynomial fitting models the lane curvature, and the radius of curvature is computed accordingly. Turn direction is predicted by evaluating the sign of the highest-degree polynomial coefficient. Experimental validation demonstrates that the proposed framework improves both lane boundary detection accuracy and turn prediction reliability. This method provides a balanced trade-off between deep learning precision and classical algorithm efficiency, making it suitable for real-world autonomous vehicle applications.