随着机器视觉和图像处理技术的不断发展，视觉目标检测、跟踪及测量技术得到了广泛应用。在航空航天领域中空中目标的视觉检测、跟踪及测量日益受到重视，特别是在空中复杂条件下，如何实现目标的检测、跟踪及位姿测量成为一个重要的课题。基于视觉的自主空中加油技术能够有效地减少飞行员在空中加油过程中的操作难度进而减少飞行训练成本，同时也能够有效地提高无人机装备的搭载水平和续航能力。基于视觉的鲁棒自主导航系统需要准确快速的目标检测与跟踪算法来保证控制系统输入信息的准确性与时效性。本文针对复杂条件下软管式空中加油的目标视觉检测与跟踪策略开展研究。主要的工作和贡献有：
（1）提出了一种基于目标形状特征的视觉检测与跟踪方法。软管式空中加油的锥套目标主要由三部分组成：内部黑心，伞骨和伞套。其中黑心部分的成像对光线变化不敏感，具有显著的轮廓特征。本文基于锥套黑心部分的形状先验知识制定图像处理规则，提出了基于目标形状特征的视觉检测算法，实现了对目标轮廓的有效检测；提出了基于目标形状特征的粒子滤波跟踪算法，实现了对目标轮廓的有效跟踪。提出了视觉检测与跟踪的切换策略，在无遮挡条件下，实现了对目标轮廓有效的检测和跟踪。
（2）提出了基于离线有监督学习的鲁棒锥套目标视觉检测方法。在空中加油过程中，会存在光照变化导致的目标图像亮度不均匀或局部饱和，空中气流扰动导致的锥套姿态变化，相对运动导致的目标尺度的变化及受油杆对目标的部分或全部遮挡等复杂条件。本文提出了两种基于离线有监督学习的锥套目标视觉检测算法，实现复杂条件下对锥套目标的有效检测。一种采用基于特定图像特征的支持向量机，利用低维归一化鲁棒局部二值特征，在不损失检测精度的基础上有效地减少了目标检测的时间。另一种通过深度卷积神经网络直接预测目标的类别和目标在图像的位置信息，通过图形处理器加速实现了对锥套目标的快速鲁棒视觉检测。
（3）提出了一种在线基于状态的并融合增量主成分分析的结构输出支持向量机目标跟踪算法。因光线变化、部分遮挡、自身尺度及姿态变化等因素的影响，目标的表面模型会不断的发生变化，进而导致跟踪失败等问题。本文通过融合生成模型和判别模型来设计鲁棒跟踪器，其中生成模型采用在线增量主成分分析实现对目标表面模型的不断更新，判别模型采用在线结构输出支持向量机实现对目标与背景的判别区分。将目标在图像空间中进行状态化表示，在目标跟踪的过程中直接预测目标在图像空间的状态。利用目标的虚拟状态，实现了判别模型和生成模型的效融合。
（4）提出了一种鲁棒检测—学习—跟踪框架，实现了检测器和跟踪器的有效融合。检测器的优势是全局判断目标的存在与否，进而实现目标的定位。检测器的劣势是需要离线学习，检测速度相对较慢。跟踪器的优势在于拥有局部的搜索域，拥有集中少量的训练样本，可以实现快速的在线学习，速度相对较快。在严重遮挡，光线剧烈变化等复杂条件下，目标跟踪器会逐渐产生漂移现象进而导致跟踪失败。本文通过检测器对跟踪器中判别模型的正支持向量进行置信度评价，寻找出不可靠的正支持向量。通过在线数据挖掘算法从观测的历史数据中挖掘出可靠的正支持向量替换不可靠的正支持向量，实现检测器对跟踪器的在线修正，从而减少跟踪器的漂移现象。通过检测器对当前跟踪的目标进行置信度评价，判断跟踪器是否产生严重漂移，进而决定是否进行全局检测。
（5）搭建了空中加油地面模拟平台。提出了基于单目视觉的空中加油目标的简化视觉测量模型，并给出了相机标定方法。将本文提出的视觉检测、跟踪及测量方法在地面模拟平台上进行了验证，利用基于位置的视觉控制实现了在笛卡尔空间对锥套的跟踪，模拟了空中加油的自主对接过程。

   With the development of machine vision and image processing techniques, visual object detection, tracking and measurement methods have been widely applied. In the field of aerospace, detection, tracking and measuring for aerial targets are increasingly valued. Specially, how to accomplish object detection, tracking and measurement under aerial complex conditions becomes an important research topic. Autonomous aerial refueling based on vision can make the operation simple for the pilot during the process of aerial refueling and reduce the cost of the flight training. At the same time, the carrying capacity and flight endurance of unmanned aerial vehicles can be improved effectively with the aid of autonomous aerial refueling. It is required to develop accurate and fast object detection and tracking methods to make sure the accurate and fast inputs for a robust autonomous navigation system based on vision. This paper mainly focuses on visual detection and tracking strategies for autonomous aerial refueling under complex conditions. Main work and contributions are as follows:
(1) A method of object detection and tracking based on the object’s shape is proposed. The drogue of probe-and-drogue refueling system is composed of three parts: inner dark part, umbrella ribs and umbrella fabric. The imaging of the inner dark part is not sensitive to light changes and the contour of the inner dark part is salient in the image. The image processing rules are designed according to the shape prior knowledge of the inner dark part. A detection algorithm based on the object’s shape is proposed to efficiently detect the contour of the inner dark part and a particle filter tracking algorithm based on the object’s shape is proposed to efficiently track the contour of the inner dark part.
(2) Robust drogue detection methods based on offline supervised learning are proposed. There are some complex conditions during aerial refueling such as uneven brightness and partial saturation of the object image caused by light changes, the changes of drogue’s position and posture caused by airflow disturbance, the scale changes caused by the relative motion and the drogue partially or fully covered by the probe. Two visual detection methods based on offline supervised learning are proposed to detect the drogue effectively under complex conditions. In the first method, a SVM classifier with certain image feature is trained to detect the drogue and a kind of low-dimensional normalized robust local binary pattern feature is proposed to reduce the detection time without losing accuracy. In the second method, a deep convolutional neural network is used to predict the category and position of the object directly. The deep convolutional neural network can detect the object fast and robustly by the aid of GPU parallel computing.
(3) A robust visual tracking method based on online state-based structured SVM combined with incremental PCA is proposed. Object’s appearance changes during tracking because of light changes, occlusions, scale changes and posture changes and so on, which leads to failure in tracking. A robust tracker is designed by combining the discriminative model with the generative model. The incremental PCA is used as the generative model to update the object appearance model and the state-based structured SVM is used as the discriminative model to distinguish the object and the background. A state is used to describe the object in the image space and the object’s state is predicted directly during tracking. The object’s virtual state is proposed to combine the discriminative mode with the generative model effectively.
(4) A Robust visual detection-learning-tracking framework is proposed to combine the detector with the tracker effectively. The detector’s advantage is that it can judge globally whether there is an object in an image or not and then locate the object’s position. The detector’s disadvantage is that it needs to learn the classifier offline and the detection time is relatively long. The tracker’s advantage is that it can accomplish fast online learning and location prediction because it owns a small number of representative training samples and local search area. The tracker will fail to track the object because of the drift problem due to the incorrect update of the appearance model and the classifier under complex conditions such as serious occlusions and light changes. In this paper, the detector is used to evaluate the positive support vectors in the discriminative model of the tracker and then the unreliable positive support vectors are found in the T-classifier. An online data mining method is adopted to mine some reliable positive examples from the observation history in the learning to replace unreliable positive support vectors in the tracker. The tracker is rectified during the process of replacement and the drift problem is reduced. The tracked object is evaluated by the detector to determine whether the drift problem is so serious that the object is detected globally or not.
(5) The ground simulation platform for aerial refueling is built. A simplified measurement model based on monocular vision is proposed to measure the position of the drogue target in Cartesian space and the camera calibration method is given. The proposed visual detection, tracking and measurement method is verified on the ground simulation platform for aerial refueling. Position based visual control method is used to track the drogue target in Cartesian space and the autonomous docking process of aerial refueling is simulated.