近地降落是无人机全任务周期中环境最复杂多变的阶段，研究无人机在复杂环境中的自主着陆能够有效降低事故率，极大提升其自主化与智能化水平，然而现有无人机自主着陆工作大多依赖于外部合作或只能进行有限且粗糙的环境感知。本文针对非合作环境中多旋翼无人机自主着陆感知问题，开展了系统平台设计、场景综合理解、威胁跟踪预测、实机演示验证等研究，主要工作及创新点包括：（1）针对非合作环境着陆场景复杂和适应性要求高的问题，构建了搭载高可靠全天候感知系统的多旋翼无人机平台；定义了各模块的坐标系，并对多传感器进行了内外参数的标定；基于所搭建的机载传感器与计算平台，在多种环境中收集了大量无人机飞行数据；通过自动化标注和简化的人工标注，构建了首个真实低空环境中涵盖深度图、语义分割图和目标检测框的综合航拍数据集。（2）针对无人机自主着陆时的安全着陆点选取问题，提出了综合考虑形态学和语义学特征的场景理解方法；构建了同时进行深度补全和语义分割的端到端网络模型，设计了多种监督和自监督式的损失函数实现稳定的多任务学习；为提升模型的鲁棒性和泛化能力，提出了基于双目视差重构原理的深度图自评估方法，实现了对激光雷达积累时间的动态选取；利用模型预测结果提出了基于阈值的安全着陆点选取策略；大量实验验证了模型和多种策略的有效性、必要性和可靠性。（3）为保障无人机着陆过程中环境与自身的安全，提出了回归式多目标跟踪模型以实现运动目标的跟踪与轨迹预测；在模型中引入重识别分支提升了目标的长期匹配能力；提出了基于数据增强和对比学习的模型训练方法以解决当前航拍场景多目标跟踪数据集不足的问题；进一步引入了域自适应的迁移学习策略提升模型在本文无人机平台上的性能；最后结合多分支结果提出了可靠的多目标跟踪策略，并融合深度信息实现了精准的三维轨迹预测。（4）针对本文多旋翼无人机平台在实际复杂环境中自主着陆的问题，提出了融合本文多种感知算法与系统性改进的鲁棒多阶段自主着陆策略；针对机载计算平台的有限算力开发了自主着陆软件系统，通过多节点设计实现对传感器数据的时空对齐与感知决策，同时还开发了手机端远程监控应用增加人在回路的安全保护；在多种真实环境中测试了无人机抗风悬停、感知测高和跟踪避障的基础功能，并开展了分步与完整的无人机自主着陆飞行实验。
Near-ground landing is the most complex and unpredictable task among the missions of Unmanned Aerial Vehicles (UAVs). Studying the autonomous landing of UAVs in complex environments can effectively reduce the accident rate and greatly improve its level of autonomy and intelligence. However, most of the existing autonomous landing works rely on external cooperation or can only perform limited and rough environmental perception. To realize the autonomous landing of multi-rotor UAVs in non-cooperative environments, this paper has carried out research on system platform design, comprehensive terrain understanding, threat tracking and prediction, and verification of real flight. The main work and innovations include:(1) To adapt to complex landing scenarios in non-cooperative environments, a multi-rotor UAV platform equipped with a highly reliable all-weather perception system is constructed. The coordinate system of each module is defined and the intrinsic and extrinsic parameters between them are calibrated. Based on the onboard sensors and computing platform, a large amount of flight data has been collected in a variety of environments. Through automated labeling and simplified manual labeling, the first real-scene low-altitude comprehensive aerial dataset is constructed, including depth maps, semantic segmentation maps, and object detection annotations.(2) To select a safe landing site, a terrain understanding method that considers both morphologic and semantic features of the ground is proposed. An end-to-end network model is constructed to simultaneously perform depth completion and semantic segmentation and a variety of supervised and self-supervised loss functions are proposed to achieve stable multi-task learning. To improve the robustness and generalization ability of the model, a self-evaluation method based on the binocular disparity reconstruction principle is proposed, realizing the dynamic selection of the LiDAR accumulation time. Using the model prediction results, a threshold-based safe landing site selection strategy is further proposed. Extensive experiments verify the effectiveness of the model and the reliability of proposed strategies.(3) To ensure the safety of the UAV and the surrounding environment during the landing process, a regression-based multi-object tracking model is proposed to track the moving targets and predict their trajectories. A re-identification branch is introduced into the model to improve its long-term matching ability. To address the problem of insufficient multi-object tracking datasets in the aerial scene, a model training strategy based on data augmentation and contrastive learning is proposed. A domain adaptation learning strategy is also introduced to further improve the performance of the model in the constructed UAV platform. By combining the results of multiple model branches, a reliable multi-object tracking strategy is proposed and the depth information is integrated to achieve accurate three-dimensional trajectory prediction.(4) To realize autonomous landing of our multi-rotor UAV in the real complex environment, a robust multi-stage autonomous landing strategy that integrates various perception algorithms and systematic improvements is proposed. Through ROS-based multi-node design, an onboard autonomous landing software system is developed to realize the spatio-temporal alignment of the sensor data and the perception of UAVs. A remote monitoring application is also developed on the phone to manually increase the safety level. The basic anti-wind hovering, altitude measurement, and obstacle avoidance functions of UAVs are evaluated. And the step-by-step and complete autonomous landing flight tests are carried out in a variety of real environments.
