Row-wise LiDAR Lane Detection Network with Lane Correlation Reﬁnement Dong-Hee Paek1 Kevin Tirta Wijaya2 and Seung-Hyun Kong1

2025-05-03 0 0 2.89MB 7 页 10玖币

侵权投诉

Row-wise LiDAR Lane Detection Network

with Lane Correlation Reﬁnement

Dong-Hee Paek1, Kevin Tirta Wijaya2, and Seung-Hyun Kong∗1

Abstract— Lane detection is one of the most important func-

tions for autonomous driving. In recent years, deep learning-

based lane detection networks with RGB camera images have

shown promising performance. However, camera-based meth-

ods are inherently vulnerable to adverse lighting conditions

such as poor or dazzling lighting. Unlike camera, LiDAR

sensor is robust to the lighting conditions. In this work, we

propose a novel two-stage LiDAR lane detection network with

row-wise detection approach. The ﬁrst-stage network produces

lane proposals through a global feature correlator backbone

and a row-wise detection head. Meanwhile, the second-stage

network reﬁnes the feature map of the ﬁrst-stage network via

attention-based mechanism between the local features around

the lane proposals, and outputs a set of new lane proposals.

Experimental results on the K-Lane dataset show that the

proposed network advances the state-of-the-art in terms of F1-

score with 30% less GFLOPs. In addition, the second-stage

network is found to be especially robust to lane occlusions,

thus, demonstrating the robustness of the proposed network

for driving in crowded environments.

I. INTRODUCTION

To be able to navigate from a point to another, an au-

tonomous driving agent needs to plan a safe and efﬁcient

route according to the environmental conditions. Therefore,

the ability of perceiving the environment through raw sensor

measurements is crucial for the autonomous driving. One

of the perception tasks for autonomous driving is the lane

detection task, where the autonomous driving agent needs to

detect the location of lane lines on the roads.

Extensive studies have been conducted on the lane de-

tection task, particularly with the RGB camera sensors. In

the earlier days, rule-based and heuristic systems have been

developed to provide lane detection capability in limited

predeﬁned environments [1][2][3]. Recently, data-driven ap-

proaches become popular owing to the advancements of

deep learning. Various neural networks for camera-based

lane detection have been developed [4][5][6], with promising

accuracy in most of driving conditions.

However, RGB camera has an inherent weakness towards

harsh lighting conditions such as low or dazzling light. This

is evident in the widely-used CULane benchmark [7], where

the performance degradation of various camera-based lane

detection networks occurs in the dazzling light and dark. As

∗corresponding author

1Dong-Hee Paek and Seung-Hyun Kong are with the CCS Grad-

uate School of Mobility, Korea Advanced Institute of Science and

Technology, 193, Munji-ro, Yuseong-gu, Daejeon, Republic of Korea

{donghee.paek, skong}@kaist.ac.kr

2Kevin Tirta Wijaya is with the Robotics Program, Korea Advanced

Institute of Science and Technology, 291, Daehak-ro, Yuseong-gu, Daejeon,

Republic of Korea kevin.tirta@kaist.ac.kr

an autonomous driving agent needs to be robust in various

driving conditions, the vulnerability of the existing camera-

based lane detection methods towards adverse lighting con-

ditions need to be resolved.

One possible solution to the problem is to use Light

Detection and Ranging (LiDAR) sensor. Since the LiDAR

sensor emits infrared signals, which is hardly interfered by

the visible light, adverse lighting conditions do not affect

its measurement capability signiﬁcantly. Moreover, unlike

camera images, the LiDAR point cloud does not require a

bird’s eye view (BEV) projection for motion planning, which

often causes lane line distortions.

Despite the several advantage of the LiDAR sensor, only

a handful of studies have proposed deep learning-based

LiDAR lane detection methods. This is largely due to the

absence of publicly-available LiDAR lane detection datasets.

As seen in the deep camera-based lane detection ﬁeld, the

majority of the lane detection networks are developed after

the publication of open lane detection datasets such as [7][8].

Recently, [9] opens a large-scale LiDAR lane detection

dataset, K-Lane, to the public, along with a segmentation-

based LiDAR lane detection network (LLDN) baseline. The

baseline consists of a projection network, a global feature

correlator (GFC), and a segmentation head, which generates

a feature map in the BEV format, extracting features via

global correlation, and predicting lanes per grid, respectively.

While a segmentation-based lane detection network is ca-

pable of producing lane detection of various shapes, it is

computationally expensive due to the need of processing

each grid of the ﬁnal feature map through shared multi-layer

perceptron (MLP).

In this work, we propose a novel LiDAR lane detec-

tion network that is computationally efﬁcient and especially

effective to the severe occlusion cases. We set the lane

detection problem as a row-wise prediction task instead of

a semantic segmentation task, so that the detection head of

the network performs row-wise MLP operations instead of

grid-wise MLP operations. The row-wise formulation leads

to a signiﬁcantly less computational cost than the prior work

[9], with about 30% less GFLOPs.

Furthermore, we design an additional second-stage net-

work that reﬁnes the output feature map of the ﬁrst-stage

network via correlation of features around the lane pro-

posals. The correlation process is implemented with the

dot-product attention mechanism, which allows the network

to exchange information between the features of the lane

proposals globally. As the lane lines often have some degree

of regularity in terms of shapes and distances between each

arXiv:2210.08745v1 [cs.CV] 17 Oct 2022

other, such global correlation process should be advanta-

geous, for example, in detecting lane lines that are occluded

by neighboring vehicles. The row-wise two-stage approach

enables our proposed network to achieve the state-of-the-art

(SOTA) performance with less computational cost compared

to the existing baselines.

In a summary, our contributions are as follows:

•We propose a new technique in LiDAR lane detection

through row-wise lane prediction. This technique is

more computationally efﬁcient compared to existing

segmentation-based LiDAR lane detection techniques.

•We design a two-stage LiDAR lane detection network

that ﬁrst predicts lane proposals, and then reﬁnes the

ﬁrst-stage feature map via attention-based mechanism

between the lane proposals. The reﬁned feature map is

then used to predict the ﬁnal lane detection proposals.

•We demonstrate the excellent performance of the pro-

posed network; the proposed network achieves SOTA

performance in K-Lane with an overall F1-score of

82.74% reducing the GFLOPs by about 30%. While

the proposed network achieves a slight overall F1-

score improvement (i.e., 0.62%) over the prior SOTA

network, the proposed network largely improves the F1-

score (i.e., 3.24%) for the severe occlusion cases. Thus,

the proposed network enables the robust lane detection

required for safe autonomous driving in congested traf-

ﬁcs where performance degradation has previously been

signiﬁcant.

The rest of this paper is organized as follows: Section II

introduces existing works related to our study, Section III

details our proposed LiDAR lane detection network, Section

IV discusses the experimental results on the K-Lane dataset,

and Section V draws conclusion this study.

II. RELATED WORKS

In this section, we discuss existing works that are related to

our study. We start with a general review of the more matured

camera-based lane detection. Then we discuss further on

the row-wise lane detection methods. Finally, we cover the

existing LiDAR lane detection methods.

A. Camera-based Lane Detection

Traditional camera-based lane detection methods heavily

rely on rule-based systems that require various predeﬁned

variables such as intensity thresholding [1][2][3]. As the

deep learning ﬁeld becomes more matured, various camera

lane detection networks emerge, with promising accuracy in

various driving conditions. Most deep learning-based camera

lane detection networks utilize convolutional neural network

(CNN) as their backbone feature extractor, and a task-speciﬁc

detection head.

In [10], the detection head is designed to perform anchor-

based lane predictions and coordinate offset predictions. In

[4], the detection head produce two afﬁnity-ﬁeld maps that

represent the locations of lane lines. In [6], the detection

head is conditioned to predict row-wise lane proposals before

being further processed by a post-processing algorithm.

Compared to other approaches, row-wise lane detection often

perform faster while maintaining good accuracy. As such, we

design our LiDAR lane detection network with a row-wise

paradigm.

B. Row-wise Lane Detection

Row-wise lane detection is proposed by [11], in which

lane lines are detected through predicting the lane location

probability of each row. That is, for each row, the location

of the lane line is determined as the column of which the

lane probability is the highest. Unlike segmentation-based

lane detection, row-wise lane detection is based on geometric

prior which picks only one location per each lane, so that this

method may robust to false alarm. Furthermore, [4] modify

the argmax operation into taking the sums of each index of

the column weighted by its lane probability to enable the

gradients to ﬂow through the lane structure loss.

C. LiDAR Lane Detection

In the early LiDAR lane detection methods, lane lines

are detected through an intensity thresholding operation

with additional heuristics. These methods often incorporate

additional algorithms such as Kalman Filter [12], polar

coordinates operation [13], or clustering with DBSCAN [14].

However, heuristic methods are not adaptive towards diverse

driving conditions due to the need for predeﬁning numerous

thershold variables.

More recently, several studies are starting to incorporate

deep learning to their LiDAR lane detection methods. In

[15], the front view camera images are combined with the

2D BEV images from the LiDAR point cloud to improve

the lane detection performance. In [16], a CNN backbone

is utilized to detect the ego lane lines in the highways. In

[9], a segmentation-based neural network with global feature

correlator backbone is used to perform LiDAR lane detection

under various environments in the K-Lane dataset.

III. METHOD

In this section, we describe the proposed network in

details. Firstly, we provide an overview of the structure of

the proposed network. Then, we explain in details about the

components of the network: the feature extractor, the row-

wise detection head, and the reﬁnement head. Finally, we

express the loss function that is used to train the network.

A. Two-stage Row-wise Lane Detection Network

As shown in Fig. 1, we design the proposed lane detection

network in two-stage detection network with a row-wise

approach. First, a feature extractor backbone takes the raw

point cloud data as an input. The feature extractor then

encodes the raw point cloud into a pseudo bird-eye-view

(BEV) image, which is further processed by a global feature

correlator to produce the output feature map. This output

feature map is utilized by the ﬁrst stage row-wise detection

head to predict two values: row-wise existence and row-wise

probability. Through the row-wise existence and location

probability predictions, we can obtain the ﬁrst-stage lane

detection proposals.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Row-wiseLiDARLaneDetectionNetworkwithLaneCorrelationRenementDong-HeePaek1,KevinTirtaWijaya2,andSeung-HyunKong1AbstractLanedetectionisoneofthemostimportantfunc-tionsforautonomousdriving.Inrecentyears,deeplearning-basedlanedetectionnetworkswithRGBcameraimageshaveshownpromisingperformance.However,ca...

展开>> 收起<<

Row-wise LiDAR Lane Detection Network with Lane Correlation Reﬁnement Dong-Hee Paek1 Kevin Tirta Wijaya2 and Seung-Hyun Kong1.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Row-wise LiDAR Lane Detection Network with Lane Correlation Reﬁnement Dong-Hee Paek1 Kevin Tirta Wijaya2 and Seung-Hyun Kong1

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: