毕业生知识库

空中加油是提高飞行器续航时间最为有效的方法。传统的空中加油过程对于 飞行员的操作有很高的要求,受油机飞行员在执行空中加油的过程中无法及时反 应意外情况,而空中无人加油技术的发展将解决这些问题。此外,军事无人机的 迅速发展对空中加油技术有了更高的要求。针对空中无人加油技术的对接过程, 本课题将开展基于单目视觉的空中加油目标的关键点检测和位姿测量方法研究, 主要采用计算机视觉及机器学习技术实现对空中加油目标的关键点检测,在此基 础上结合单目视觉的测量方法,获得空中加油目标的位姿。测量得到的目标锥套 位姿可以用于指导空中加油飞行器完成受油管与输油管的自动对接。本文的主要 工作和创新性成果包括:

(1)基于地面模拟系统建立了锥套图像数据库。通过搭建地面的双机器人 模拟系统,在此系统上采集锥套图像,人工标注锥套位置、锥套关键点坐标、锥 套位姿,前二者用于验证本研究中提出的锥套关键点检测算法的有效性,标注位 姿用于验证本研究提出的位姿测量框架的有效性。

(2)提出了一种结合锥套关键点邻域特征及关键点之间几何约束的关键点 检测方法。传统的基于随机森林的关键点检测算法是在提取每个关键点邻域特征 的基础上回归各个关键点,本研究在此基础上,融入了锥套所有关键点之间的约 束关系,通过实验验证了本算法在关键点定位精度上的提升。此外,通过引入图 像预处理及关键点选择策略,提升了锥套轮廓定位的精度。

(3)提出了一个基于单目视觉的锥套位姿测量框架。整个锥套位姿测量框 架包括锥套关键点检测、锥套轮廓定位、锥套位姿测量三个部分,在得到锥套轮 廓的基础上,利用简化的位姿测量算法或基于三维立体几何的精确位姿测量算法 获得锥套在相机坐标系下的位姿,在实验部分利用标注数据验证了位姿测量系统 的有效性。

本文提出的测量框架在 Intel i5-4570 CPU 平台上,输入图像分辨率为 320×240 像素时,锥套位姿测量耗时为 38.8ms,在大部分位姿上,测量误差能 保持在 10cm 以内,对基于视觉的空中自主加油技术的研究具有重要指导意义。 

Aerial refueling is the most effective approach to enhance duration of flight of aerocraft. It takes a lot of energy for pilots to accomplish aerial refueling in traditional aerial refueling methods, and it is hard for pilots to react to emergency in this process, autonomous aerial refueling have solved this challenge. Besides, the development of unmanned aerial vehicle also get a higher requirement on this technology.To contrib- ute to the development of autonomous aerial refueling, we focus on the research of pose measurement of drogue via monocular vision, which can teach the machine to accomplish aerial refueling autonomously. The basic technologies of this research are landmark detection and pose estimation based on monocular vision. Main work and contributions are as following:

(1) Building the database of drogue based on a simulation system. The simula- tion system is consisted of two robots, one of which plays the role of flying tanker, another plays the role of receiver aircraft. We mark each sample in this dataset with labels, which include the position of drogue in this image, cordinates of landmarks and the pose of drogue, the first two are used to validate the accuracy of our landmard detection algorithm, the last one is used to validate the practicability of the pose measurement method we proposed.

(2)Combining local feature and global feature of landmarks to detect landmarks. Traditional random forest regression uses local feature of landmarks to regress each landmarks, based on these research, we introdece a fusing strategy to get global fea- ture involved, which has be validated to be helpful to improve the accuracy in our experiment. Besides, we propose another two strategies to improve the accuracy of locating drogue.

(3)Proposing an effective framework of pose measurement based on monocular vision. This framework is consisted of landmarks detection, locating outline of drogue in image, locating drogue in camera cordinate system, these three parts are executed step by step, so as to get the final result. We validate the effectiveness of this frame- work in our experiment, which will be described detailedly in this paper.

We proposed an effective pose measurement framework for autonomous aerial refueling, this framework is consisted of landmark detection, position locating, pose measurement. On the plateform based on Inter i5-4570 CPU, we can measure pose of target in the rate of 38.8ms per frame, when resolution of input image is 320 × 240. Besides, the measuring error is smaller than 10cm at most time. This research has given great contribution to aerial aotomatical refueling.