Monocular Lane Detection Based on Deep Learning: A Survey

TL;DR

This survey tackles monocular lane detection for autonomous driving, covering both 2D and 3D lane estimation from a single front-view camera. It distills methods into four design axes: task paradigm, lane modeling, global context supplementation, and perspective elimination, and evaluates them on standard benchmarks under a unified setting. The analysis highlights BEV-based and BEV-free approaches, IPM limitations, and learnable view transformations, and it discusses extensions such as multi-task perception, video lane detection, and online HD map construction. The work provides a practical roadmap for researchers and practitioners to compare baselines, understand tradeoffs, and push toward end-to-end, robust monocular lane understanding.

Abstract

Lane detection plays an important role in autonomous driving perception systems. As deep learning algorithms gain popularity, monocular lane detection methods based on them have demonstrated superior performance and emerged as a key research direction in autonomous driving perception. The core designs of these algorithmic frameworks can be summarized as follows: (1) Task paradigm, focusing on lane instance-level discrimination; (2) Lane modeling, representing lanes as a set of learnable parameters in the neural network; (3) Global context supplementation, enhancing inference on the obscure lanes; (4) Perspective effect elimination, providing accurate 3D lanes for downstream applications. From these perspectives, this paper presents a comprehensive overview of existing methods, encompassing both the increasingly mature 2D lane detection approaches and the developing 3D lane detection works. Besides, this paper compares the performance of mainstream methods on different benchmarks and investigates their inference speed under a unified setting for fair comparison. Moreover, we present some extended works on lane detection, including multi-task perception, video lane detection, online high-definition map construction, and lane topology reasoning, to offer readers a comprehensive roadmap for the evolution of lane detection. Finally, we point out some potential future research directions in this field. We exhaustively collect the papers and codes of existing works at https://github.com/Core9724/Awesome-Lane-Detection and will keep tracing the research.

Figures (9)

• Figure 1: Two different technical routes to achieve complete lane detection. Different lane instances are visualized by different colors. The 2D lane detection results need to be projected to 3D space with the help of camera parameters and inverse perspective mapping (IPM). 3D lane detection methods directly incorporate the camera parameters into the network and achieve end-to-end prediction of 3D lanes.
• Figure 2: The structure of this paper. Different colors represent specific sections. The summary of existing methods covers four core designs of the lane detection algorithms mentioned in the introduction.
• Figure 3: The general pipelines for 2D lane detection: (a) Segmentation-based methods, which leverage mask, grids, or keypoints to model a lane, perform instance discrimination and lane localization sequentially. Each method only performs instance discrimination once. (b) Object detection-based methods, which leverage line anchor or parameter curve to model a lane, complete the instance discrimination and lane localization in parallel.
• Figure 4: The ways for discriminating different lane instances in segmentation-based methods: (a) Bottom-up faction represented by LaneNetLaneAFFOLOLaneGANet. (b) Top-down faction represented by SCNNSADRESAUFLDGANet, which predefines the maximum number of lanes and treats each lane as a category. (c) Top-down faction represented by CondLaneNetCondLSTR, which predicts dynamic kernels to generate instances. Among them, grids-based modeling is only applicable to the top-down approach.
• Figure 5: More details of different lane modeling. (a) A lane in FV image. (b) Mask-based modeling, which performs semantic segmentation. (c) Grids-based modeling, which executes the row/column-wise classification. (d) Keypoints-based modeling, which detects discrete points on the lane. (e) Line Anchor-based modeling, which learns the horizontal offset of equidistant points. (f) Curve-based modeling, which predicts the curve parameters.