【作者】 刘悦；

【导师】 李化义；

【作者基本信息】 哈尔滨工业大学 ， 航空宇航科学与技术， 2025， 博士

【摘要】 近些年,航天发射任务日益频繁,尤其是大规模卫星星座的部署,使在轨航天器和空间碎片的数量急剧上升,轨道空间资源有限,这也使航天器在轨运行的风险急剧上升。各国通过建立空间态势感知系统来实现在轨目标监测和风险规避。针对空间非合作目标的在轨自主视觉感知技术是实现态势感知的前提,在未来复杂的太空任务中,将成为在轨航天器必备的功能之一,有重要的研究价值。本文针对空间非合作目标,利用基于视觉的方法,对在轨自主感知过程中航天器接近测量阶段的目标检测、多目标跟踪和相对状态估计问题展开深入研究,具体内容如下:针对航天的复杂场景、非合作目标以及有限的训练数据集,本文提出一种用于少样本高效训练网络的辅助模块,称为因素忽视模块,并提供了一套该模块使用的完整方案。首先,对航天场景下的域间距进行分析与建模,然后在此基础上进行基于因素忽视模块的航天器专用数据集的制作,最后通过改变训练网络的损失函数、采用独特的训练方法来实现网络参数的调整。通过在已公开的非合作航天器数据集上进行的测试表明,忽视地球背景能够提升目标检测的精度,忽视色彩和曝光因素会导致回召率的提升,增加因素数据集的丰度和调整超参数可以实现效果的进一步提升。此外,在像素级检测任务语义分割中的应用也表明了模块的可迁移性。针对漏检、遮挡、相同外观、轨迹变化等难点,本文基于Deep SORT框架,提出了一种专门的在线多航天器跟踪方法Sat SORT。由于目前尚未有公开的多航天器跟踪数据集,为了研究的顺利进行,本文首先基于MOT16框架自制了多航天器跟踪数据集,然后对在线跟踪算法Deep SORT进行了测试与分析,最后在此基础上对匹配损失方程、数据关联流程和滤波器预测更新策略等进行了全方面的升级改进。大量的测试结果表明,Sat SORT在跟踪效果上远远超过Deep SORT,并在处理效率上提升了一个数量级,在漏检、遮挡、外观相同等场景中均有更佳的表现。此外,通过与其他经典的地面多目标跟踪框架对比,本文的方法更适用于航天中训练样本少、外观相同的场景。针对近距离非合作目标状态估计问题,本文采用捕获-跟踪框架。对于状态捕获阶段,本文针对该阶段快速鲁棒估计的要求,采用To F相机的灰度与深度图像(点云)深度融合的方式,研究了一种基于圆特征的非合作目标位姿估计方法。首先,本文对圆特征与相对位姿进行关联建模。然后,分别研究了基于点云和灰度图像的圆参数求解方法。针对点云数据庞大耗时长的特点,提出了一种基于灰度图像辅助点云的分割方法;针对圆特征阴影和立体结构导致椭圆检测不佳的问题,提出多椭圆筛选策略;针对圆特征不规则不完整的难点,提出了一种点云圆参数拟合方法。最后,对两种数据融合的相关影响因素进行了分析。通过模型实拍数据与仿真实验表明,该捕获方法能够大大提升点云求解的速度,在不同程度的噪声影响下仍然能够稳定求解,并且总姿态角误差保持在0.1度以内,位移误差保持在毫米级以下。对于状态跟踪阶段,本文针对该阶段高精度高动态的要求,采用双目相机与陀螺仪传感器,进行相对状态连续估计的非线性高斯滤波公式推导与相关方法的性能分析,为实现在轨在线高动态估计提供参考。更多还原

【Abstract】 In recent years,space launch missions have become increasingly frequent,especially with the deployment of large-scale satellite constellations.This has led to a sharp rise in the number of spacecraft and space debris in orbit.Given the limited orbital space resources,the risks associated with in-orbit spacecraft operations have significantly in-creased.To address this,various countries have been developing Space Situational Aware-ness(SSA)systems to monitor on-orbit targets and avoid potential risks.Autonomous visual perception technologies for non-cooperative targets in space are fundamental for achieving SSA.In future complex space missions,such technologies will become essential capabilities for on-orbit spacecraft,making them highly valuable for research.This paper focuses on non-cooperative targets in space and explores,through visual-based methods,the problems of target detection,multi-object tracking,and relative state estimation dur-ing the autonomous perception process of spacecraft in the proximity measurement phase.The specific contributions are as follows:Given the challenges posed by complex aerospace environments,non-cooperative targets,and limited training datasets,the paper proposes an auxiliary module for efficient few-shot training,called the Factor Ignoring Module.A complete implementation scheme is presented.First,inter-domain distances in aerospace scenarios are analyzed and mod-eled.Based on this,a dedicated spacecraft dataset is constructed using the factor ignoring module.Then,by modifying the loss function of the training network and applying a unique training method,network parameters are optimized.Tests on a publicly avail-able non-cooperative spacecraft dataset demonstrate that ignoring the Earth background improves detection accuracy,while ignoring color and exposure factors boosts recall.In-creasing the diversity of the factor dataset and tuning hyperparameters further enhances performance.Moreover,applying the module to pixel-level semantic segmentation tasks indicates strong transferability.To address challenges such as missed detections,occlusion,identical appearances,and trajectory changes,this paper builds upon the Deep SORT framework and proposes Sat SORT,a online multi-spacecraft tracking method specifically designed.Since no pub-lic multi-spacecraft tracking dataset currently exists,the paper first constructs a custom dataset based on the MOT16 framework.After evaluating and analyzing the online track-ing algorithm Deep SORT,comprehensive improvements are made to the matching loss function,data association process,and filter prediction-update strategies.Extensive test results show that Sat SORT significantly outperforms Deep SORT in tracking performance and achieves an order-of-magnitude improvement in processing efficiency.It also per-forms better under challenging conditions such as occlusion,missed detections,and tar-gets with similar appearances.Compared to other classic ground-based multi-object track-ing frameworks,this method proves more suitable for space applications with limited training samples and visually similar targets.For the problem of estimating the state of close-range non-cooperative targets,this paper adopts a capture-tracking framework.During the capture phase,to meet the demand for fast and robust estimation,a fusion method combining grayscale and depth images(point clouds)from a Time-of-Flight(To F)camera is used.A pose estimation method based on circular features for non-cooperative targets is proposed.First,circular features are correlated with relative poses.Then,methods for solving circular parameters from both point clouds and grayscale images are explored.To address the issues of large point cloud data and long processing time,a segmentation method that uses grayscale images to assist point cloud processing is proposed.To counteract poor ellipse detection caused by shadowing and the 3D structure of circular features,a multi-ellipse filtering strategy is introduced.For irregular and incomplete circular features,a point cloud circle param-eter fitting method is developed.Finally,the influence of data fusion factors is analyzed.Real-world model data and simulation experiments demonstrate that this capture method significantly improves the speed of point cloud processing.It remains stable under vari-ous noise levels,with total attitude angle errors within 0.1 degrees and displacement errors within millimeter-level accuracy.During the tracking phase,to meet the requirements of high precision and high dynamics,the paper uses stereo cameras and gyroscope sensors to derive nonlinear Gaussian filtering equations for continuous relative state estimation,and analyzes the performance of relevant methods.This provides a reference for achieving real-time,high-dynamic state estimation on orbit.更多还原