SIA OpenIR  > 海洋信息技术装备中心
水下特定目标的探测与位姿估计方法研究
Alternative TitleResearch on Detection and Pose Estimation for Underwater Specific Targets
刘爽
Department海洋信息技术装备中心
Thesis Advisor林扬
Keyword水下目标探测 姿态估计 水下对接
Pages131页
Degree Discipline机械电子工程
Degree Name博士
2019-11-28
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract本文重点研究基于视觉的水下特定目标探测与位姿估计方法。水下特定目标探测与位姿估计是基于视觉的水下机器人水下自主对接的关键技术。水下自主对接技术通过赋予自主水下机器人自主进入水下基站的能力达到为水下机器人补充能源和回收水下机器人的目的,是提高自主水下机器人续航能力、无人化管理水下机器人的核心关键技术,具有重大的实际应用需求。本文面向自主水下机器人水下自主对接与水下自主回收任务,以水下对接基站为特定目标,研究针对水下基站特定目标的探测与位姿估计,具体研究内容包括:(1) 水下特定目标的探测方法研究。针对水下特定目标探测过程探测准确率不高的问题,提出了基于区域提取的水下特定目标探测网络与基于回归的水下特定目标探测网络。在基于区域提取的水下特定目标探测网络内容中,分析了各类数据扩充方法对于探测性能的提升效果。在基于回归的水下特定目标探测网络内容中,通过实验验证了该网络在水下基站探测中的优异性能,并分析了其在诸如模糊、颜色偏移、对比度偏移、镜像等常见水下图像特性中的鲁棒性。实验结果表明,所提出的基于回归的水下特定目标探测网络在水下基站的探测中综合性能优于基准方法。(2) 水下特定目标探测的域适应方法研究。在水下特定目标探测中,训练模型所使用的训练集数据与实际应用场景中的数据往往不满足独立同分布假设,具有较大的域差异性。这种域差异性将导致使用训练集训练得到的模型域适应能力弱,在实际应用中性能显著降低。为了解决上述问题,提出了一种新型的对抗神经网络T2FGAN,能够生成给定水质、光源属性、位姿和配置下的水下图像。实验结果表明利用所生成的图像训练探测模型,显著地提高了的探测模型域适应能力。(3) 水下特定目标特征标识提取方法研究。提取水下特定目标的特征标识是在探测得到的结果基础上以图像分割的形式提取出目标的特征标识,为下一步的三维空间中的位姿估计做准备。针对浅水区域特征标识提取出现的复杂环境光和非均匀扩散问题,提出了基于拉普拉斯高斯滤波器的水下特征标识提取方法。实验通过定性与定量的分析表明所提出的方法优于基准方法。(4) 三维空间中水下特定目标的位姿估计方法。三维空间中水下特定目标的位姿估计方法计算水下机器人相对目标的位置和姿态。针对实际应用中可能出现的特征标识失效或被部分遮挡的问题,提出了全观测下与部分观测下三维空间中的位姿估计方法。分别在数据集、陆地实验和室内水池实验验证了所提出的方法准确有效,能够满足实际应用需求。(5) 湖上试验验证。将所提出的方法应用于静态基站AUV水下资助对接能源补充和移动基站水下自主回收AUV两项实际任务中。分别通过二者的湖试试验验证了所提出的方法有效,满足自主水下机器人在这两项任务中的实际需求。
Other AbstractThe 21st century is the century of oceans. The ocean plays more and more important roles in the world. Oceans are unknown to human beings. Human beings yearn to the resources in the oceans. All of above motivate human beings to explore the oceans. Underwater robotics can be divided into three kinds; Human Occupied Vehicle (HOV), Remotely Operated Vehicle (ROV) and Autonomous Underwater Vehicle (AUV). AUV offers broad operation range and efficient operating type. It has become the trend of underwater robotics. Detection and pose estimation of underwater specific targets is a very important component of the sensing of underwater robotics. It is also of great importance and of high requirements to the sensing of underwater robotics. The detection of underwater specific target refers to that a specific underwater target can be found and its location can be located in the view of the sensor once it appears in the view of the sensor. The location of underwater specific target refers to computing the location and orientation between the underwater robotics and the target. Among sensors used in the detection and pose estimation of underwater targets, cameras are widely adopted due to its relatively low price and high resolution. Detection and pose estimation of underwater specific targets is a key technique to vision-based autonomous underwater docking. AUVs is able to refill energy and being recovered automatically with the help of autonomous underwater docking. In this paper, we aim at autonomous underwater docking and automated underwater recovery. Taking the underwater docking station as the specific underwater target, we study the detection and pose estimation of the underwater docking station. Our research includes: (1) Research on the detection of underwater specific target. In order to solve poor performance of existing methods for detection of underwater specific target, a region-based underwater specific target detection network and a regression-based underwater specific target network were proposed. The contribution of the different data augmentation method for improving detection performance are analyzed in the content of region-based underwater specific target detection network. In the chapter of a regression-based underwater specific target network, the good performance of the proposed region-based underwater specific target network is validated by experiments. We also analyzed its robustness in blurring, color shift, contrast shift and mirror underwater images. Experiments show that the proposed regression-based underwater specific target network achieved better performance than baseline models. (2) Research on domain adaptation in underwater specific target detection. In underwater specific target detection, the training data and the data in practice is not independent identically distributed. There is a domain shift between them. The domain shift results in poor performance of the trained model in practice, which is weak in domain adaptation. In order to solve problems above, a novel Generative Adversarial Network T2FGAN was proposed. The network is able to generate underwater images with given water properties, illumination, pose and configurations. Experiments show that the performance of the model trained on the generated images is improved obviously. (3) Research on extraction of feature points of the underwater specific target. Extraction of feature points of underwater specific target bases on the detection, laying foundation for pose estimation of targets in 3D space. In order to solve the problem of complex ambient light and non-uniform spreading, a Laplacian of Gaussian filter based feature extraction method was proposed. Experiments show that the proposed method is better than the baseline method both qualitatively and quantitatively. (4) Research on pose estimation of underwater specific target in 3D space. Pose estimation of underwater specific target in 3D space computes the location and orientation between the target and the underwater robotics. In some situations, feature points are missing. In order to estimate the location and the orientation in these situations, pose estimation methods in full and partial observation were proposed. The proposed method was validated by the dataset, land test and indoor underwater experiments. Experiments show that the proposed method is accurate and effective, meeting the real requirements. (5) Field experiments validating. The proposed methods were applied to autonomous underwter docking for AUV energy refilling using the static docking station and autonomous underwater docking for AUV recovery using mobile docking station. The proposed methods were validated by field experiments of these two tasks. Experiments show that the proposed methods can meet the requirements in practice.
Language中文
Contribution Rank1
Document Type学位论文
Identifierhttp://ir.sia.cn/handle/173321/25947
Collection海洋信息技术装备中心
Recommended Citation
GB/T 7714
刘爽. 水下特定目标的探测与位姿估计方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
水下特定目标的探测与位姿估计方法研究.p(10978KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[刘爽]'s Articles
Baidu academic
Similar articles in Baidu academic
[刘爽]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[刘爽]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.