晨光
暗夜
晨光
极光
Bilingual Paper Reading · 中英对照精读

智能视觉自主舰载降落:无 GPS 环境下 VTOL 无人机在舰船上的自主跟踪与降落

准大一 · 轮机工程 × 舰载航空保障 × AI+控制 —— 无人机自主舰载降落精读材料
原文:arXiv:2202.13005 2022年2月25日发布 arXiv 预印本(cs.RO) 视觉降落 × 非线性控制 × 无 GPS 附英文摘要朗读音频

一、论文档案

英文标题Intelligent Vision-based Autonomous Ship Landing of VTOL UAVs
中文标题基于智能视觉的垂直起降(VTOL)无人机自主舰载降落
作者李甫灿, 维什努·萨吉, 莫布尔·贝内迪克特, 迪利普·卡拉蒂尔(机构未在素材中标注)
发布时间2022年2月25日(v1)|分类:cs.RO(机器人学)
一句话概括不靠 GPS,只用一台机载相机:机器学习检测远距离舰船、经典视觉识别「地平线参考条」估计位姿、非线性控制器完成垂直降落——在模拟 6 自由度船体运动的移动甲板上成功完成 100+ 次降落试验。
💡 为什么选这篇给你:① 直升机舰载起降是海军航空保障的核心难题,也是轮机工程「舰船-航空」交叉方向的经典问题;② 思路巧妙——不追踪甲板运动(这正是海军飞行员训练的反直觉结论),而是参考陀螺稳定的「地平线条」,把专业领域知识直接变成算法设计;③ 数字扎实:0.25 km 检测距离、等效 17.3 km 探测、100+ 次真实飞行降落试验,故事完整、可复现。

二、核心术语表(先扫一遍再读正文)

英文术语中文大白话解释
VTOL (Vertical Take-Off and Landing)垂直起降能像直升机一样垂直起飞、垂直降落的飞行器,本文指四旋翼无人机。
UAV (Unmanned Aerial Vehicle)无人机无人驾驶的航空器;本文的验证平台是带机载相机的四旋翼(quad-rotor)无人机。
autonomous ship landing自主舰载降落无人机不依赖人工,自动完成在舰船甲板上的跟踪与降落全过程。
GPS-denied / spoofed environmentGPS 拒止 / 欺骗环境GPS 信号不可用或被伪造的环境,是作战与电子对抗中的常见威胁。
horizon bar地平线参考条陀螺稳定、始终指示真实水平线的视觉基准装置,多数现代海军舰船都装有它。
deck motion甲板运动舰船在风浪中的六自由度(6 DOF)运动:三个平移 + 三个旋转(横摇 roll、纵摇 pitch 等)。
machine learning object detection机器学习目标检测用深度神经网络(如 YOLOv3)在图像中找出目标位置并给出边界框。
YOLOv3YOLOv3单阶段目标检测算法,检测速度比 Faster R-CNN、SSD 更快,本文用它实时检测舰船与地平线条。
classical computer vision经典计算机视觉不用深度学习的传统图像处理方法,本文用于近距离的相对位置与姿态估计。
nonlinear controller非线性控制器控制律与状态呈非线性关系的控制器,适合大范围工况与强扰动场景。
nonlinear gain variation非线性增益变化控制增益随误差大小变化:小误差时低增益、大误差时高增益,用于应对时间延迟。
probabilistic approach概率化方法用概率判断估计值是否「物理上可能」,拒绝虚假估计,避免控制器乱动。
Kalman estimator卡尔曼估计器融合多源传感器数据、在噪声中估计系统状态的经典滤波算法。
sea state海况等级按浪高与风浪剧烈程度划分的海面状态等级(如海况 6 级属于剧烈海况)。

三、摘要中英对照(精读核心)

🎧 音频在文末,可先听一遍原文再读;每个英文句都配了逐句翻译。

摘要 Abstract

EN · 原文
The paper discusses an intelligent vision-based control solution for autonomous tracking and landing of Vertical Take-Off and Landing (VTOL) capable Unmanned Aerial Vehicles (UAVs) on ships without utilizing GPS signal.
CN · 翻译
本文讨论一种基于智能视觉的控制方案,用于垂直起降(VTOL)无人机在不使用 GPS 信号的情况下自主跟踪并降落在舰船上。
EN · 原文
The central idea involves automating the Navy helicopter ship landing procedure where the pilot utilizes the ship as the visual reference for long-range tracking; however, refers to a standardized visual cue installed on most Navy ships called the "horizon bar" for the final approach and landing phases.
CN · 翻译
核心思想是把海军直升机舰载降落流程自动化:远距离跟踪阶段飞行员以舰船本身为视觉参照;而最后进近与降落阶段则参照大多数海军舰船都安装的标准化视觉标识——「地平线条」(horizon bar)
EN · 原文
This idea is implemented using a uniquely designed nonlinear controller integrated with machine vision.
CN · 翻译
该思想通过一个与机器视觉集成的独特非线性控制器来实现。
EN · 原文
The vision system utilizes machine learning-based object detection for long-range ship tracking and classical computer vision for the estimation of aircraft relative position and orientation utilizing the horizon bar during the final approach and landing phases.
CN · 翻译
视觉系统采用基于机器学习的物体检测进行远距离舰船跟踪,并用经典计算机视觉在最后进近与降落阶段利用地平线条估计飞机的相对位置与姿态。
EN · 原文
The nonlinear controller operates based on the information estimated by the vision system and has demonstrated robust tracking performance even in the presence of uncertainties.
CN · 翻译
非线性控制器基于视觉系统估计的信息工作,即使在存在不确定性的情况下也展现出鲁棒的跟踪性能
EN · 原文
The developed autonomous ship landing system was implemented on a quad-rotor UAV equipped with an onboard camera, and approach and landing were successfully demonstrated on a moving deck, which imitates realistic ship deck motions.
CN · 翻译
所开发的自主舰载降落系统在搭载机载相机的四旋翼无人机上实现,并在模拟真实舰船甲板运动的移动甲板上成功演示了进近与降落。
EN · 原文
Extensive simulations and flight tests were conducted to demonstrate vertical landing safety, tracking capability, and landing accuracy.
CN · 翻译
开展了大量仿真与飞行试验,以验证垂直降落安全性、跟踪能力与降落精度

关键词 Keywords:Vision-based Control 视觉控制 | Ship Landing 舰载降落 | VTOL UAV 垂直起降无人机 | YOLOv3 目标检测 | Nonlinear Control 非线性控制

四、引言精选(为什么这个问题重要)

① 问题本质:小甲板 + 六自由度运动 + 视野受限

EN · 原文
Landing a helicopter on a small ship at rough sea states is an extremely challenging task even for human pilots due to the small landing space, six degrees of freedom ship deck motions, limited visual references for pilots, and lack of alternative landing spots. There have been many studies in the past that focused on automating helicopter ship landing by utilizing a wide array of sensors such as GPS, vision sensors, motion sensors, LIDAR, etc. This paper investigates a novel solution that falls under the category of vision-based control system that does not use GPS signals and thus ensures its functionality in GPS-denied/spoofed environments.
CN · 翻译
在剧烈海况下把直升机降落到小船上,即使对经验丰富的飞行员也是极具挑战的任务——降落空间小、甲板有六自由度运动、飞行员视觉参照有限、且没有备选降落点。过去有许多研究利用 GPS、视觉、运动传感器、激光雷达等各类传感器实现直升机自主舰载降落。本文研究一种不依赖 GPS 信号的视觉控制系统,从而保证其在 GPS 拒止/欺骗环境中的可用性。

② 反直觉的第一课:海军飞行员被训练成「不跟随甲板运动」

EN · 原文
Contrary to intuition, Navy pilots are trained not to follow ship deck motions for two main reasons. First, spatial disorientation can occur when a pilot has no fixed, visible horizon to refer to, which is a critical element for maintaining a proper sense of helicopter attitude independent of ship motions. The key visual aid that helps pilots to land safely is a ”horizon reference bar” shown in Fig. 1, which is gyro-stabilized to indicate a perfect horizon regardless of ship motions and is widely used in most modern Navies [15, 16].
CN · 翻译
与直觉相反,海军飞行员被训练为不要跟随甲板运动,原因有二。第一,当飞行员没有固定、可见的地平线可参照时,会发生空间定向障碍——而地平线是独立于船体运动维持直升机姿态感的关键要素。帮助飞行员安全降落的关键视觉辅助装置就是图 1 所示的「地平线参考条」,它经陀螺稳定,无论船体如何运动都指示出完美的水平线,被大多数现代海军广泛使用 [15, 16]。
EN · 原文
Second, constantly changing the helicopter attitude to match ship deck motions can trigger unstable helicopter dynamics introducing serious potential hazards. Hence, a pilot tries to control the helicopter in a stable manner independent of the ship's roll and pitch motions, and then lands vertically.
CN · 翻译
第二,不断改变直升机姿态去匹配甲板运动会激发不稳定的直升机动力学,带来严重安全隐患。因此飞行员会独立于船的横摇与纵摇运动、以稳定的方式控制直升机,然后垂直降落

③ 混合视觉系统:远距离机器学习 + 近距离经典视觉,检测器选 YOLOv3

EN · 原文
The vision system is hybrid in nature with two different methods, a machine learning object detection and a classical computer vision method, each of which is designed to operate depending on the relative distance to the landing pad. In the long-distance, the machine learning object detection method is applied to identify the landing platform (the ship), and an image-based control is utilized in the autonomous flight control system.
CN · 翻译
视觉系统本质上是混合的,包含两种方法:机器学习目标检测与经典计算机视觉,二者按与降落平台的距离分工。远距离时用机器学习目标检测识别降落平台(舰船),并在自主飞行控制系统中采用基于图像的控制。
EN · 原文
According to the studies that compared the state-of-the-art algorithms, YOLOv3 demonstrated faster detection performance than Faster R-CNN and SSD [23, 24]. Hence, the YOLOv3 algorithm is selected to train an object detector that is able to detect a ship and a horizon bar in real-time.
CN · 翻译
根据对比最先进算法的研究,YOLOv3 的检测速度优于 Faster R-CNN 和 SSD [23, 24]。因此选择 YOLOv3 训练目标检测器,使其能够实时检测舰船与地平线条
💡 这是全文最有味道的一句"Contrary to intuition, Navy pilots are trained not to follow ship deck motions for two main reasons."——专业领域里「反直觉」的成熟经验,往往就是算法设计的金钥匙:与其把甲板运动建模得越来越准,不如像飞行员一样干脆不去跟随它。

五、论文贡献(3 个要点)

EN · 原文
1. A GPS-free vision-based landing framework. This paper investigates a novel solution that falls under the category of vision-based control system that does not use GPS signals and thus ensures its functionality in GPS-denied/spoofed environments.
CN · 翻译
1. 无 GPS 的视觉降落框架。提出一种不依赖 GPS 信号的视觉控制系统方案,保证系统在 GPS 拒止/欺骗环境中仍能正常工作——贴近真实作战需求。
EN · 原文
2. A hybrid vision system. The vision system utilizes machine learning-based object detection for long-range ship tracking and classical computer vision for the estimation of aircraft relative position and orientation utilizing the horizon bar during the final approach and landing phases.
CN · 翻译
2. 混合视觉系统。远距离用机器学习目标检测跟踪舰船,近距离用经典计算机视觉结合地平线条估计飞机相对位置与姿态——两段式分工,兼顾速度与精度。
EN · 原文
3. A situation-adaptive nonlinear controller. The control system is developed to generate situation-adaptive control inputs by introducing the idea of nonlinear gain variation and a probabilistic approach to limit the impact of incorrect estimations.
CN · 翻译
3. 情境自适应的非线性控制器。通过非线性增益变化概率化方法生成情境自适应控制输入,把错误估计的影响限制住——即使视觉系统给出错误估计,控制器也绝不产生突变的巨大控制量。

六、结论中英对照

EN · 原文
The goal of this study is to develop an autonomous ship landing solution for VTOL UAVs by closely following the Navy helicopter ship approach and landing procedure. This automation has been achieved by using a single onboard camera without using GPS. The ship landing system consists of a machine vision system and a nonlinear control system. The machine vision system is developed while taking advantage of state-of-the-art machine learning and classical computer vision techniques. The control system is developed to generate situation-adaptive control inputs by introducing the idea of nonlinear gain variation and a probabilistic approach to limit the impact of incorrect estimations.
CN · 翻译
本研究的目标是严格遵循海军直升机舰船进近与降落流程,为 VTOL 无人机开发自主舰载降落方案。该自动化仅用一台机载相机、不使用 GPS 即实现。系统由机器视觉系统与非线性控制系统组成:视觉系统融合了最先进的机器学习与经典计算机视觉技术;控制系统通过非线性增益变化与概率化方法生成情境自适应控制输入,限制错误估计的影响。
EN · 原文
Multiple flight tests are systematically conducted to verify the safety of vertical landing maneuver, long-range detection and robust tracking capability, and landing accuracy. The UAV lands vertically independent of ship motions in the same manner as a Navy helicopter lands on a ship. More than 100 landing tests are successfully conducted on a moving deck, which mimicked realistic and challenging 6 DOF ship motions.
CN · 翻译
系统性地开展了多次飞行试验,验证垂直降落动作的安全性、远距离检测与鲁棒跟踪能力以及降落精度。无人机像海军直升机一样独立于船体运动垂直降落。在模拟真实而苛刻的 6 自由度船体运动的移动甲板上,成功完成 100 多次降落试验
EN · 原文
The machine learning object detector trained by the YOLOv3 algorithm begins identifying the 1.8 x 1.8 meters (6 x 6 feet) ship platform from 0.25 kilometers away, which means the range in the case of a real ship landing can be 17.3 kilometers (9.3 nautical miles). This is estimated assuming that the rear-side of a typical small ship occupies an area of 15 x 15 meters (50 x 50 feet). The unique nonlinear control system demonstrates robust tracking capability during a wide range of realistic scenarios such as random initial positions, complicated ship motions, communication latency, sensor noise, and in the presence of winds up to 9 m/s (17.5 knots).
CN · 翻译
由 YOLOv3 训练的机器学习目标检测器从0.25 千米外就能识别 1.8 x 1.8 米(6 x 6 英尺)的舰船平台——按典型小型舰船尾部占据 15 x 15 米(50 x 50 英尺)面积推算,真实舰船降落时探测距离可达 17.3 千米(9.3 海里)。独特的非线性控制系统在随机初始位置、复杂船体运动、通信延迟、传感器噪声以及高达 9 m/s(17.5 节)风等广泛真实场景下都展现出鲁棒跟踪能力。
EN · 原文
The proposed vertical landing maneuver for VTOL UAVs, which involves not following the ship deck motions, was verified as a safe way of landing. This was achieved through multiple landing tests on the deck mimicking challenging ship motions (Oliver Hazard Perry frigate at sea state 6) and NATOPS helicopter ship landing operational limits (roll: ±8, pitch: ±3).
CN · 翻译
所提出的「不跟随甲板运动」的垂直降落动作被验证为安全的降落方式。验证过程包括在模拟严苛船体运动(海况 6 级下的奥利弗·哈泽德·佩里级护卫舰)以及 NATOPS 直升机舰载降落操作极限(横摇 ±8°、纵摇 ±3°)的甲板上进行多次降落试验。
EN · 原文
The long-range vision system developed based on the state-of-the-art machine learning algorithm YOLOv3 demonstrated a 10 times greater detection range than the classical computer vision systems. The control system successfully utilized the detected object position and relative size as states for long-range tracking.
CN · 翻译
基于 YOLOv3 的远距离视觉系统展示了比经典计算机视觉系统大 10 倍的检测距离。控制系统成功地把检测到的目标位置与相对尺寸作为状态用于远距离跟踪。
EN · 原文
The biggest challenge to implement the machine learning based object detection on the real-time autonomous flight was the time delay. To cope with the time delay issue, a long-range controller was constructed that responded less sensitively to errors around the setpoint and aggressively to large errors, using an exponential variation of feedback gain with error.
CN · 翻译
把基于机器学习的目标检测用于实时自主飞行的最大挑战是时间延迟。为应对延迟,远距离控制器被设计成:设定点附近的误差响应不敏感、大误差时响应激进——通过反馈增益随误差的指数变化实现,使无人机在接近舰船平台时保持在正确航线上。
EN · 原文
Even after going through the configured Kalman estimator, large/false estimation error can still occur from time to time. To prevent responding to such non-physical estimations, the probabilistic nonlinear controller was developed. It probabilistically perceives if the estimation is physically possible or not, based on the normal distribution curve and known UAV characteristics. Multiplying the estimation value by its probability can effectively reject responding to false estimations. By this approach, the controller never generated abrupt large control inputs even when the vision system provided inaccurate estimations.
CN · 翻译
即使经过卡尔曼估计器,偶尔仍会出现大误差/虚假估计。为阻止控制器响应这类非物理估计,开发了概率化非线性控制器:它基于正态分布曲线与已知的无人机特性,概率化地判断估计是否物理可行,用估计值乘以其概率,从而有效拒绝虚假估计。通过这一方法,即使视觉系统给出不准确的估计,控制器也从未产生突变的巨大控制输入。

七、编者解读:这篇论文到底讲了什么(大白话版)

  1. 问题:在摇晃的小型舰船上降落直升机/无人机极其危险——甲板六自由度乱动、参照物少、没有备选降落点。传统做法是用传感器「盯住甲板、跟着甲板动」,但这样既难又危险。
  2. 反直觉的解法:海军飞行员几十年实战总结的经验是「别跟着甲板动」——盯着陀螺稳定的「地平线参考条」,保持自身姿态稳定,然后垂直落下去。作者把这条行业经验原封不动变成了算法:不追踪甲板运动,只做视觉定位 + 垂直降落。
  3. 两段式视觉:远距离(还在几公里外)用 YOLOv3 深度学习检测舰船;近距离用经典视觉识别地平线条、精确估计相对位置和姿态——「深」与「浅」的方法各管一段,既快又准。
  4. 结果:0.25 km 外就能认出 1.8 x 1.8 米的模拟舰船平台(等效真实舰船 17.3 km);在模拟海况 6 级船体运动的甲板上成功降落 100+ 次;抗 9 m/s 风速、通信延迟和传感器噪声。
  5. 工程智慧:视觉估计偶尔会出错,控制器用「概率化判断」拒绝非物理估计,绝不因错误输入猛打舵——这种「宁可保守、绝不失控」的容错思路,和船舶动力系统的故障安全设计一脉相承。
🎯 对保研的启示:这篇论文示范了「把行业成熟经验翻译成算法设计」的能力——先理解领域专家几十年实践出来的规则(不跟随甲板运动),再为其设计机器视觉与控制系统。复试时若能讲出「我从领域经验里提炼了什么约束、又用算法兑现了它」,会非常打动人。

八、给准大一的阅读路线图 & 延伸方向

📖 怎么读这篇论文(三遍法)

  1. 第一遍(10 分钟):只读摘要和术语表,回答三个问题——问题是什么(舰载降落)?方法是什么(视觉 + 非线性控制)?结果是什么(100+ 次降落成功)?
  2. 第二遍(20 分钟):读引言 + 结论,重点体会「为什么飞行员不跟随甲板运动」以及「为什么远距离用深度学习、近距离用经典视觉」。
  3. 第三遍(30 分钟):读引言文字部分,跳过所有公式、引用编号和图表,只追踪主线:视觉如何分工 → 控制如何抗延迟 → 控制器如何拒绝虚假估计。

🚀 这个方向你能延伸做什么

九、英文摘要朗读(练听力用)

先盲听一遍→再看对照稿→再听一遍。目标是听出每个数字(0.25 km、17.3 km、100+ 次)和术语(horizon bar、VTOL、GPS-denied)。