ElasticLaneNet: An Efficient Geometry-Flexible Approach for Lane Detection

TL;DR

This work introduces ElasticLaneNet, a real-time lane-detection framework that leverages an implicit geometric representation called the Elastic Lane Map (ELM), where lanes are zero-level contours encoded across rows. A physics-inspired Elastic Interaction Energy (EIE) loss guides the evolution of the ELM by coupling predictions with ground truth, enabling robust handling of complex geometries such as dense, cross, and Y-shaped lanes. The architecture combines a ResNet34+FPN backbone with a Transformer Bottleneck and auxiliary refinements (AFR, CSN, RSN) to produce a flexible, end-to-end trainable pipeline, achieving state-of-the-art performance on the structurally diverse SDLane dataset (F1=89.51, Recall=87.50, Precision=91.61) with fast inference, and competitive results on TuSimple and CULane. Ablation studies confirm the benefits of the EIE loss and the implicit ELM representation, demonstrating improved generalization to challenging lanes while maintaining efficiency for real-time deployment.

Abstract

The task of lane detection involves identifying the boundaries of driving areas in real-time. Recognizing lanes with variable and complex geometric structures remains a challenge. In this paper, we explore a novel and flexible way of implicit lanes representation named \textit{Elastic Lane map (ELM)}, and introduce an efficient physics-informed end-to-end lane detection framework, namely, ElasticLaneNet (Elastic interaction energy-informed Lane detection Network). The approach considers predicted lanes as moving zero-contours on the flexibly shaped \textit{ELM} that are attracted to the ground truth guided by an elastic interaction energy-loss function (EIE loss). Our framework well integrates the global information and low-level features. The method performs well in complex lane scenarios, including those with large curvature, weak geometry features at intersections, complicated cross lanes, Y-shapes lanes, dense lanes, etc. We apply our approach on three datasets: SDLane, CULane, and TuSimple. The results demonstrate exceptional performance of our method, with the state-of-the-art results on the structurally diverse SDLane, achieving F1-score of 89.51, Recall rate of 87.50, and Precision of 91.61 with fast inference speed.

Figures (10)

• Figure 1: Illustrations of challenging cases for lane detection. (a) Weak features, including occlusion, no line or dashed lines, night/shadow/dazzle light conditions pan2018spatial. (b) Complex geometry lanes, such as Y-shaped (new branch, intersections, turns), dense, merging, winding driveway, etc jin2022eigenlanes.
• Figure 2: Overview of the proposed ElasticLaneNet architecture. The backbone is constructed as Encoder-FPN pipeline. Elastic Lane Map Module (ELMM) is the lane detection head. Classification and range sub-networks (CSN, RSN) are jointly trained and producted to obtain the final ELM. TB layer and AFR module are incorporated in the network. The EIE loss measures the distance between prediction and the ground truth.
• Figure 3: Elastic Lane Map, ELM of one lane line.
• Figure 4: Row-wise lane points in an image.
• Figure 5: Lanes moving informed by the EIE loss.