LanePtrNet: Revisiting Lane Detection as Point Voting and Grouping on Curves

TL;DR

This paper proposes a novel approach, LanePtrNet, which treats lane detection as a process of point voting and grouping on ordered sets, and takes backbone features as input and predicts a curve-aware centerness, which represents each lane as a point and assigns the most probable center point to it.

Abstract

Lane detection plays a critical role in the field of autonomous driving. Prevailing methods generally adopt basic concepts (anchors, key points, etc.) from object detection and segmentation tasks, while these approaches require manual adjustments for curved objects, involve exhaustive searches on predefined anchors, require complex post-processing steps, and may lack flexibility when applied to real-world scenarios.In this paper, we propose a novel approach, LanePtrNet, which treats lane detection as a process of point voting and grouping on ordered sets: Our method takes backbone features as input and predicts a curve-aware centerness, which represents each lane as a point and assigns the most probable center point to it. A novel point sampling method is proposed to generate a set of candidate points based on the votes received. By leveraging features from local neighborhoods, and cross-instance attention score, we design a grouping module that further performs lane-wise clustering between neighboring and seeding points. Furthermore, our method can accommodate a point-based framework, (PointNet++ series, etc.) as an alternative to the backbone. This flexibility enables effortless extension to 3D lane detection tasks. We conduct comprehensive experiments to validate the effectiveness of our proposed approach, demonstrating its superior performance.

Figures (5)

• Figure 1: Proposed method introduces the concept of centerness as a fundamental scoring metric for candidate proposals in lane detection. Points located along the lane are represented as a bell-shaped distribution, with the curve centers ignited at the peak denoted by yellow dots. This formulation enables the reformation of lane detection as a grouping task based on the provided seeds.
• Figure 2: The overall architecture of LanePtrNet. Given an input image from the front view, the backbone encoder is employed to extract point-wise features. In the decoding phase, the centerness head is employed to predict the centered distribution of lane lines, generating a confidence map. Subsequently, the C-FPS algorithm is applied to identify $K$ seed points for the lanes. The grouping head performs clustering based on the positions of these seed central points, producing seed masks. Following this, the cross-instance attention module determines $N$ (where $N\leq{K}$) lane lines as the final prediction results. Additionally, the semantic head provides global semantic information about the lane lines to guide the generation of seed masks.
• Figure 3: Computing the centerness heatmaps of curves where points in yellow depict peak value. (a) Using \ref{['eq:fcos']}. (b) Using the path integral method \ref{['eq:changeSi']}. (c) Corner cases where lane markings approximate extreme aspect ratio may lead to unstable box centerness (e.g. a straight line).
• Figure 4: The process of generating lane groups. Blue blocks represent the grouping map while the yellow cubes are seed points aggregating their corresponding features. Through tensor broadcasting, these selected seeds are duplicated and concatenated to match the size of the spatial size, obtaining the final seed mask after several convolutional blocks.
• Figure 5: (a) shows predicted seed points ordered in numbers from 0 to 9. Green lines are ground-truth results for this case. (b) denotes attention scores across lane groups given centers. Red regions indicate highly correlated groups while blue blocks show unrelated prediction. (c)(d)(e) indicate the 1st,3rd and 4th lane results for reference.