A Five-Era Taxonomy and Benchmark Framework for Lane Detection: From Classical Heuristics to Vision Foundation Models in Autonomous Driving

Volume 17, Number 2

A Five-Era Taxonomy and Benchmark Framework for Lane Detection: From Classical Heuristics to Vision Foundation Models in Autonomous Driving

Authors

Nitin Vishnoi, ESGCI, Paris

Abstract

This survey paper traces the technological history of lane detection systems over the last quarter century, discussing paradigm shifts from classical computer vision methods to modern foundation models. The evolution is divided into five eras: Classical Vision-based (2000–2010), Feature + Geometry (2006–2014), CNN Segmentation (2015–2019), Anchor/Curve-based and Transformer Methods (2020–2022), and the Foundation Model generation (2023–present). Each phase is discussed based on methodological developments, pivotal contributions, performance attributes, and shortcomings. The survey synthesizes the original literature, showing how machine learning, deep learning, and scale-based pre-training have tackled robustness, generalization and real-time issues. We identify research gaps in edge cases, system integration, and suggest future directions towards cohesive perception models achieving optimal accuracy, efficiency, and interpretability.

Keywords

Lane Detection, Computer Vision, Autonomous Driving, Deep Learning, Foundation Models, Survey, Intelligent Transportation System