【作者】 李旭；

【导师】 丁萌；

【作者基本信息】 南京航空航天大学 ， 交通信息工程及控制， 2021， 硕士

【摘要】 随着计算机与自动化技术的发展,以无人机(Unmanned Aerial Vehicle,UAV)和无人车为代表的自主无人系统逐渐走入军事和民用领域。但是无论在军用还是民用领域,无人系统的制导、导航与控制(Guidance/Navigation and Control,GNC)能力都是其自主化能力的重要标志。其中,无人系统能自主估算前视场景深度信息与并进行障碍规避是其GNC能力的重要体现。因为具有能耗低、性价比高以及视场大的优点,通过视觉进行深度估计是目前最有研究价值的方法之一。因此,本文以无人系统为研究对象,以其前视场景的深度感知与障碍距离估计为研究目标,采用单目视觉结合深度学习的方法,同时展开基于可见光和红外图像的深度估计研究,重点研究了如何通过视觉提高无人系统全天候GNC能力,并且进行了全闭环仿真验证。本文的研究内容如下:首先,考虑无人系统前视场景特征、平台尺寸限制及数据来源问题,确定了单目自监督学习的理论框架。结合三维视觉知识推导出自监督信号的来源,并设计合适的卷积神经网络进行图像的深度、位姿特征提取,再结合自监督信号来源与图像处理知识构建了网络的损失函数。其次,考虑无人系统的全天候使用场景,提出了基于上述神经网络框架的可见光训练模式与红外光训练模式,并模拟无人机白天飞行与无人车夜间行驶的场景进行训练。针对数据集的特性采用了平均/最小重投影误差、多尺度上采样等一系列优化策略。然后,本文着重研究了夜间环境下的深度感知,结合红外图像相对于可见光图像的单通道、低纹理特征进行了迁移学习,使得带有“先天缺陷”的红外图像同样能很好的训练。在两种训练模式结束后,还对深度估计结果进行了定性定量评估。最后,为了验证本文方法对于视觉前视深度估计任务的有效性,搭建了一个全闭环无人系统仿真平台,通过嵌入本文算法,实现了无人机通过单目视觉感知前视场景深度,在不依靠雷达的情况下于山区完成过峰避障与地形跟随的机动效果,进一步证实了本文方法对于提高无人系统GNC能力的意义与贡献。更多还原

【Abstract】 With the development of computer and automation technology,unmanned system represented by unmanned aerial vehicle(UAV)and unmanned vehicle has gradually entered the military and civil fields.However,no matter in the military or civil field,the Guidance/Navigation and Control(GNC)capability of unmanned systems is a key research object.Among them,the ability of unmanned systems to autonomously estimate the depth information of forward-looking scenes and to circumvent obstacles is an important embodiment of their GNC capability.Due to its advantages of easy deployment and low energy consumption,depth estimation through vision is one of the most valuable research methods at present.Combined with the end-to-end deep learning theory,it is easier to bring this method into full play.However,some unmanned systems are limited in payload,such as light and small unmanned aerial vehicles,which leads to the limitation of sensor carrying.Moreover,in some applications,the sensors also need to work all day and all night,which also brings difficulties to solve the depth estimation problem of forward-looking scenes.Therefore,this paper takes the unmanned system as the research object,takes the depth perception and obstacle distance estimation of its forward-looking scenes as the research target,and adopts the method of monocular vision combined with deep learning to focus on how to improve the all-weather GNC capability of unmanned system through vision.The research contents of this paper are as follows.Firstly,the theoretical framework of monocular self-supervised learning is determined by considering the forward-looking scene features,platform size limitations and data sources of the unmanned system.The source of the supervised signal was derived by combining the 3D vision knowledge,and the appropriate convolutional neural network was designed to extract the depth and pose features of the image.Then the loss function of the network was constructed by combining the source of the self-supervised signal and the image processing knowledge.Secondly,considering the all-weather use scenario of the unmanned system,the visible light training mode and infrared light training mode based on the above neural network framework were proposed,and the training scenes of UAV flying in the daytime and unmanned vehicle driving at night were simulated.According to the characteristics of the data set,a series of optimization strategies,such as average/minimum reprojection error and multi-scale up-sampling,are adopted.Then,this paper focuses on the depth perception in the nighttime environment.Combined with the single channel and low texture features of the infrared image compared with the visible image,the transfer learning is carried out,so that the infrared image with "congenital defects" can also be well trained.At the end of the two training modes,the depth estimation results were evaluated qualitatively and quantitatively.Finally,in order to validate the method for the validity of the visual front depth estimation tasks,set up a complete closed loop unmanned system simulation platform,by embedding algorithm,this paper realizes the UAV by monocular visual perception forward-looking scene depth,in the case of not rely on radar completed mountain peak of obstacle avoidance and terrain following motor effect,further confirmed the significance of the method to improve the ability of unmanned systems GNC and contribution.更多还原