Vision-Based Lane Detection Using Machine Learning

Lane detection is one of the most fundamental components of intelligent transportation systems, particularly in autonomous vehicles and modern Advanced Driver Assistance Systems (ADAS). Accurate lane perception enables safe navigation, stable lane keeping, and informed decision-making. Traditional lane detection approaches—such as Canny edge detection, Hough Transform, and color-based thresholding—show reasonable performance in controlled environments but fail under challenging real-world conditions involving low visibility, shadows, faded lane markings, and abrupt illumination changes. With the rise of deep learning, particularly Convolutional Neural Networks (CNNs), models have gained the ability to learn robust lane features directly from data. However, real-world driving requires more than lane detection; it also demands an understanding of drivable areas and the presence of objects such as vehicles or pedestrians. This paper presents a comprehensive review of classical and modern lane detection techniques, with a focus on multi-task deep learning architectures, such as YOLOP (You Only Look Once for Panoptic Driving Perception). We also implement YOLOP on real-world Indian road videos and enhance its performance on nighttime scenes using custom brightness, contrast, and gamma preprocessing. The integration of night enhancement improved lane IoU from 0.72 to 0.84 and pixel accuracy from 0.88 to 0.93. The review highlights major advancements, limitations, research gaps, and future opportunities in machine-learning-based lane detection. The findings emphasize that multi-task learning, domain adaptation, and lightweight models are essential steps toward practical and reliable autonomous vehicle perception systems.