Advancements in 3D Lane Detection Using LiDAR Point Clouds: From Data Collection to Model Development

TL;DR

The paper tackles the scarcity of large-scale, surround-view LiDAR datasets for 3D lane detection and the sparsity of manual lane annotations by introducing LiSV-3DLane and an automatic labeling pipeline. It proposes LiLaDet, a two-pathway LiDAR-based detector that combines BEV-based segmentation with a 3D spatial refinement module and cross-modal fusion via BVAT, reinforced by SFWA. Experimental results on LiSV-3DLane and K-Lane show LiLaDet outperforms existing camera- and LiDAR-based approaches in 3D lane detection and spatial accuracy, with ablations confirming the contribution of each component. This work enables more accurate and dense 3D lane perception in diverse driving scenarios and sets the stage for future multi-modality fusion to further reduce computational demands.

Abstract

Advanced Driver-Assistance Systems (ADAS) have successfully integrated learning-based techniques into vehicle perception and decision-making. However, their application in 3D lane detection for effective driving environment perception is hindered by the lack of comprehensive LiDAR datasets. The sparse nature of LiDAR point cloud data prevents an efficient manual annotation process. To solve this problem, we present LiSV-3DLane, a large-scale 3D lane dataset that comprises 20k frames of surround-view LiDAR point clouds with enriched semantic annotation. Unlike existing datasets confined to a frontal perspective, LiSV-3DLane provides a full 360-degree spatial panorama around the ego vehicle, capturing complex lane patterns in both urban and highway environments. We leverage the geometric traits of lane lines and the intrinsic spatial attributes of LiDAR data to design a simple yet effective automatic annotation pipeline for generating finer lane labels. To propel future research, we propose a novel LiDAR-based 3D lane detection model, LiLaDet, incorporating the spatial geometry learning of the LiDAR point cloud into Bird's Eye View (BEV) based lane identification. Experimental results indicate that LiLaDet outperforms existing camera- and LiDAR-based approaches in the 3D lane detection task on the K-Lane dataset and our LiSV-3DLane.

Figures (6)

• Figure 1: Our work aims at extracting 3D lane lines from a surround-view LiDAR point cloud(a). These detected lanes are visualized in BEV (b) and 3D coordinate system (c).
• Figure 2: Current LiDAR-based 3D lane detection methods project LiDAR points into a BEV grid map. The voxelization process leads to the loss of spatial details due to the low resolution of discrete grid cells.
• Figure 3: Automatic Lane Annotation Pipeline for generating finer lane annotation.(a) Raw manual lane annotation; (b) Lane skeleton representations and lane skeletal points equidistantly sampled along links; (c) Unlabeled lane points selected by ball-query searching; (d) Smooth lane points sampled from the interpolated cubic curve function.
• Figure 4: Dataset Statistics Analysis. We analyze the coordinates, height, curvature, and slope of lanes to illustrate the diversity of lane geometry in LiSV-3DLane.
• Figure 5: Overview of our proposed LiLaDet framework. Given a LiDAR point cloud as input, our model first identifies the lane segments from the projected BEV space to generate 3D lane point proposals at the BEV Pathway. Then, we design a Spatial Pathway to refine the lane proposal points through geometric regression and confidence prediction.