BARS A Benchmark for Airport Runway Segmentation Wenhui Chen1Zhijiang Zhang1 Liang Yu2Yichun Tai1

2025-05-06 0 0 8.7MB 18 页 10玖币

侵权投诉

BARS: A Benchmark for Airport Runway Segmentation

Wenhui Chen 1·Zhijiang Zhang 1, ·Liang Yu 2·Yichun

Tai 1

Abstract Airport runway segmentation can effectively reduce the accident rate during the landing phase, which

has the largest risk of ﬂight accidents. With the rapid development of deep learning (DL), related methods achieve

good performance on segmentation tasks and can be well adapted to complex scenes. However, the lack of large-

scale, publicly available datasets in this ﬁeld makes the development of methods based on DL difﬁcult. Therefore,

we propose a benchmark for airport runway segmentation, named BARS. Additionally, a semiautomatic annota-

tion pipeline is designed to reduce the annotation workload. BARS has the largest dataset with the richest cate-

gories and the only instance annotation in the ﬁeld. The dataset, which was collected using the X-Plane simulation

platform, contains 10,256 images and 30,201 instances with three categories. We evaluate eleven representative

instance segmentation methods on BARS and analyze their performance. Based on the characteristic of an air-

port runway with a regular shape, we propose a plug-and-play smoothing postprocessing module (SPM) and a

contour point constraint loss (CPCL) function to smooth segmentation results for mask-based and contour-based

methods, respectively. Furthermore, a novel evaluation metric named average smoothness (AS) is developed to

measure smoothness. The experiments show that existing instance segmentation methods can achieve prediction

results with good performance on BARS. SPM and CPCL can effectively enhance the AS metric while modestly

improving accuracy. Our work will be available at https://github.com/c-wenhui/BARS.

Keywords Airport runway benchmark ·Synthetic airport runway dataset ·Instance segmentation ·Boundary

smoothing

Wenhui Chen

E-mail: wenhuichen@shu.edu.cn

Zhijiang Zhang

E-mail: zjzhang@staff.shu.edu.cn

Liang Yu

E-mail: yuliang1@comac.cc

Yichun Tai

E-mail: taiyc@shu.edu.cn

1 School of Communication and Information Engineering, Shanghai University, PR China

2 Shanghai Aircraft Design and Research Institute, PR China

arXiv:2210.12922v3 [cs.CV] 17 Apr 2023

2 Wenhui Chen 1et al.

1 Introduction

Aircraft ﬂight phases include departure, cruising, and landing. Compared to other phases, the landing phase has

the largest risk of ﬂight accidents, as it is the most difﬁcult phase to operate in. A successful landing requires

maintaining the proper glide angle and descent speed, as well as ensuring that the aircraft’s ﬂight path is aligned

with the runway centerline and crosses the intended landing point on the runway. Therefore, reducing pilot work-

load and improving safety during the landing phase are vital goals for the aviation industry. Existing landing

phase navigation systems include the instrument landing system and ground-based augmentation system. How-

ever, the deployment cost of such systems is high. In recent years, the visual navigation system has emerged as

a new development in this ﬁeld due to its low cost. Many studies [1–6] are aimed at achieving automatic land-

ing with the help of computer version technology. As an important part of the visual navigation system, airport

runway segmentation classiﬁes the runway markings at the pixel level, resulting in segmentation results that can

indicate whether the aircraft is aligned with the runway centerline and whether the current glide angle is reason-

able, which helps pilots better perceive the runway position, enabling automatic landing and improving safety

during the landing phase.

Existing solutions to airport runway segmentation are mainly implemented by identifying runway character-

istics such as textures [7] and line segments [8, 9]. However, such traditional image processing methods provide

limited categories and cannot distinguish instances in the same category. Moreover, these methods are difﬁcult

to adapt to complex scenes, as some unrelated objects with similar shapes or structures may decrease accuracy.

Segmentation methods [10–13] based on deep learning (DL) provide good generalization and performance but

rely on related datasets. A few studies [14–17] have proposed datasets for airport runway segmentation. However,

the datasets in [14–16] are remote sensing image datasets taken from the Earth view, which are not applicable to

aircraft landing phase scenes. Additionally, they all have a small quantity of data and are not publicly available.

Due to the lack of relevant large-scale, publicly available datasets, existing segmentation methods based on DL

are difﬁcult to apply to this ﬁeld. In addition, those methods cannot be perfectly applicable because they mainly

target irregular objects, whereas airport runway segmentation is for objects with more regular shapes.

To address the two issues raised above, namely, the lack of large-scale, publicly available datasets and the in-

applicability of existing methods, we propose a benchmark for airport runway segmentation (BARS), along with

a smoothing postprocessing module (SPM) and a contour point constraint loss (CPCL) function. Furthermore,

the average smoothness (AS) is designed to measure the smoothness of the segmentation results. BARS contains

10,256 airport runway images from the aircraft view, with images captured from the X-Plane simulation platform

1. There are 30,201 instances with three categories in BARS. The LabelMe toolbox 2is used to complete the

annotation, and a semiautomatic annotation pipeline is designed to reduce the annotation workload. Compared

with other datasets [14–17], the proposed BARS 1) has the largest number of images, 2) contains the fullest

categories and instance annotations, 3) involves a variety of scenes, and 4) holds large variations as the images

are obtained in different weather and at different times. Some examples from BARS and other datasets are shown

in Fig. 1. We employ instance segmentation methods to simultaneously segment different runway markings and

to distinguish instances, such as multiple runways within an image. Eleven representative instance segmentation

methods, which include mask-based [18–20] and contour-based methods [21,22], are evaluated on BARS. SPM

and CPCL are proposed based on the regular shape airport runway characteristic. SPM is a plug-and-play mod-

ule that is designed for the inference phase of mask-based instance segmentation methods. SPM employs coarse

to ﬁne smoothing operations to alleviate the problem of rough segmentation boundaries. CPCL is proposed for

contour-based methods, which can smooth boundaries and speed up the convergence of the model by introducing

prior knowledge to restrict the contour points.

1X-Plane, https://www.x-plane.com/

2Labelme, http://labelme.csail.mit.edu/Release3.0/

BARS: A Benchmark for Airport Runway Segmentation 3

(a) The dataset in [15] (b) The dataset in [16] (c) RunwayDataset [17]

(d) BARS

Fig. 1 Some examples from (a) dataset in [15], (b) dataset in [16], (c) RunwayDataset [17], and (d) BARS (ours)

The main contributions of this paper are as follows.

(1) We propose a publicly available benchmark for airport runway segmentation, named BARS. BARS has

the largest dataset with the richest categories and the only instance annotation in the ﬁeld. We also create a

semiautomatic annotation pipeline. Eleven representative instance segmentation methods are evaluated on BARS.

(2) Based on the regular shape airport runway characteristic, we propose SPM, CPCL, and an evaluation

metric named AS.

(3) Extensive experiments demonstrate that existing instance segmentation methods can provide prediction

results with good performance on BARS. SPM and CPCL can effectively enhance the AS metric while modestly

improving accuracy.

4 Wenhui Chen 1et al.

2 Related work

In this section, we provide a brief overview of existing airport runway segmentation methods and the current

progress of instance segmentation.

2.1 Airport runway segmentation

Airport runway segmentation methods include traditional image processing-based methods and machine

learning-based methods, with traditional image processing-based methods being more common.

2.1.1 Traditional image processing-based methods

Traditional image processing-based methods for airport runway segmentation rely on line segments [9] and

saliency features [23]. In [24], the runway was extracted using the line segment detector (LSD) algorithm, but

it was insufﬁcient to eliminate boundary information and spurious line segments. Ajith et al. [25] detected the

runway boundaries by selecting the appropriate Hough lines using runway characteristics. Abu-Jbara et al. [26]

proposed a method that combines segmentation and minimization of the energy function.

Traditional image processing-based methods employ artiﬁcial features that heavily depend on prior knowl-

edge. Therefore, these methods are effective in some situations, but their application scenarios are quite limited.

Unrelated objects having similar shapes or structures, such as rivers, roads, and coastlines, may degrade the

performance.

2.1.2 Machine learning-based methods

Aytekin et al. [14] proposed a texture-based method that used the AdaBoost algorithm to segment runways, and

a dataset consisting of 57 large satellite images was utilized for the experiment. In [15], 700 remotely sensed

images provided by NWPU-RESISC45 [27] were annotated, and Mask R-CNN was used to accomplish the

segmentation. Men et al. [16] collected 1,300 remote sensing images from ”Google Earth” and used DeepLab [28]

to complete runway segmentation. Wang et al. [17] proposed the RunwayDataset with 2,000 images for semantic

segmentation.

Images from existing runway segmentation datasets can be classiﬁed as remote sensing (Earth view) [14–16]

or natural scene (aircraft view) [17]. Remote sensing images are taken from the Earth view, which cannot meet the

needs of the aircraft landing phase and are not applicable to the visual navigation system. The current datasets for

airport runway segmentation are limited in scale and scene, with only one category and no instance annotations.

Moreover, these datasets are not publicly available.

2.2 Instance segmentation

Instance segmentation is a challenging task because it requires segmenting different categories while distinguish-

ing instances of the same category.

2.2.1 Mask-based instance segmentation methods

Mask-based methods segment instances by classifying each pixel. Classical mask-based two-stage instance seg-

mentation methods include bounding box extraction and pixel-level segmentation, such as Mask R-CNN [18]

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：

BARS:ABenchmarkforAirportRunwaySegmentationWenhuiChen1ZhijiangZhang1,LiangYu2YichunTai1AbstractAirportrunwaysegmentationcaneffectivelyreducetheaccidentrateduringthelandingphase,whichhasthelargestriskofightaccidents.Withtherapiddevelopmentofdeeplearning(DL),relatedmethodsachievegoodperformanceons...

展开>> 收起<<

BARS A Benchmark for Airport Runway Segmentation Wenhui Chen1Zhijiang Zhang1 Liang Yu2Yichun Tai1.pdf

共18页,预览4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！

BARS A Benchmark for Airport Runway Segmentation Wenhui Chen1Zhijiang Zhang1 Liang Yu2Yichun Tai1

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: