垂直起降固定翼无人机兼具固定翼飞机飞行速度快、续航能力久与直升机起降灵活等优势，具有广泛的应用前景。其中，尾座式无人机不需要额外作动装置便可实现多种飞行模态的转换，是近年来垂直起降固定翼无人机的研究热点。本文针对现有尾座式无人机悬停控制能力不足、全状态飞行控制器切换繁杂、起降过程细粒度环境感知困难等问题开展研究。主要研究工作包括： (1) 针对现有单旋翼尾座式无人机在悬停状态单纯依靠舵面控制能力不足的问题，提出了一种矢量控制与气动舵面控制相结合的复合操控方法。对于该种矢量单旋翼尾座式无人机，首先建立其矢量单旋翼模型，在此基础上构建采用四元数姿态表征的六自由度非线性动力学模型，并设计了反步法悬停控制器；通过在研制的THU-800矢量单旋翼尾座式无人机上开展悬停飞行试验，验证了所提操控方法与悬停控制器的有效性。 (2) 针对尾推式双旋翼尾座式无人机推力作用点在重心下方而导致的系统自身稳定性差的问题，提出了一种基于反步法的非线性悬停控制方法。首先针对尾推式双旋翼尾座式无人机建立了六自由度非线性动力学模型，然后在简化模型的基础上利用李雅普诺夫稳定性理论设计了基于反步法的非线性悬停控制器；在研制的THU-Albatross尾座式无人机上开展的悬停飞行试验结果显示，即使在受到外界干扰时，所提控制方法依然能够实现尾推式双旋翼尾座式无人机的稳定悬停。 (3) 针对现有尾座式无人机全状态飞行控制器切换繁杂的问题，提出了一种尾座式无人机全状态飞行自适应姿态控制方法。首先建立了尾座式无人机全状态飞行动力学模型，然后基于反步法设计了全状态飞行姿态控制器。该控制器通过带有遗忘因子的最小二乘法对模型未知参数进行在线辨识，并根据飞行状态自适应调整控制器参数；在THU-800矢量单旋翼尾座式无人机上开展的全状态飞行试验结果显示，所提控制方法能够实现尾座式无人机的高品质全状态飞行姿态控制。 (4) 针对尾座式无人机起降过程细粒度环境感知困难的问题，提出了一种细粒度单目视觉环境感知方法。首先采用空洞卷积空间金字塔池化模型和多尺度中间层特征融合设计了适用于单目深度估计和语义分割的神经网络模型；接着提出了一种基于概率图梯度的语义分割损失函数，通过约束局部区域概率值间的相对关系促进模型对细节特征的学习；在多个数据集上开展的实验结果显示，所提方法能够有效学习场景细节特征，实现复杂环境下的细粒度单目深度估计和语义分割。
The vertical take-off and landing （VTOL） fixed-wing unmanned aerial vehicle （UAV） possesses the high cruise speed and long endurance advantages of fixed-wing UAVs as well as the flexible take-off and landing ability of unmanned helicopters, and has broad application prospects. Among various types of VTOL UAV configurations, the tail-sitter UAV can achieve the transition between different flight modes without requiring additional actuation mechanisms and has become a research hotspot of VTOL fixed-wing UAVs in recent years. However, existing tail-sitter UAVs face many problems including insufficient control capability during hover flight, cumbersome controller switches among different flight modes as well as inability of achieving fine-grained environment perception during take-off and landing processes. This paper focuses on addressing above-mentioned problems and the main work includes: （1） A hybrid control method leveraging both vector control and control surfaces is proposed to address the problem of insufficient hovering control capability in existing single-propeller tail-sitters, which solely rely on control surfaces to effect control. Aiming at the single-propeller tail-sitter that adopts this control method, the model of the vectored propeller is first established, base on which the six-freedom nonlinear dynamics model of the tail-sitter is developed using the quaternion attitude description. A backstepping controller is then derived to achieve its hovering control. Hovering experiment conducted on the developed THU-800 tail-sitter with single vectored propeller verifies the effectiveness of the proposed control method and the designed hovering controller. （2） To address the low stability of the tail-propulsion twin-rotor tail-sitter due to the fact that its center of propulsion locates below the centre of gravity, a nonlinear backstepping hovering control method is proposed. }Aiming at this type of tail-sitter configuration, the six-freedom nonlinear dynamics model is first built using the quaternion attitude description, based on which a nonlinear backstepping hovering controller is derived using the Lyapunov stability theory. The hovering experiment conducted on the developed tail-propulsion twin-rotor tail-sitter THU-Albatross suggests that the proposed hovering control method is able to stabilize the aircraft even under external disturbances. （3） An adaptive attitude control method for the full-state flight of the tail-sitter is proposed to avoid the cumbersome switches among different flight controllers in existing tail-sitter UAVs. The full flight regime dynamics model of the tail-sitter is first established based on which the adaptive attitude controller for the full-state flight is designed using the backstepping method. The designed controller identifies unknown parameters online with the forgetting factor recursive least square method and adjusts controller parameters according to the tail-sitter‘s flight state, avoiding the incontinuity incurred by controller switches. Full-state flight experiment conducted on the developed THU-800 tail-sitter with single vectored propeller suggests that the proposed control method can achieve full-state flight attitude control with high quality. （4） A fine-grained environment perception method based on monocular vision is proposed to enable accurate environment perception of the tail-sitter UAV during the take-off and landing processes. By leveraging the atrous spatial pyramid pooling （ASPP） module and fusing multi-scale features from intermediate layers, a deep convolution neural network （CNN） that is suitable for monocular depth estimation and semantic segmentation is designed. Furthermore, a probability gradient loss is proposed for the semantic segmentation task, which can promote the learning of fine-details by regulating the relative relationship of nearby probability values. Experiments conducted on various datasets demonstrate that the proposed method can effectively learn scene details and achieving fine-grained monocular depth estimation and semantic segmentation of complex environment.
