【作者】 葛勇；

【导师】 杨志华；

【作者基本信息】 哈尔滨工业大学 ， 电子信息（专业学位）， 2022， 硕士

【摘要】 随着计算机视觉与无人机技术的快速发展,利用无人机对大规模环境进行探索、加工和反馈感知信息的技术在目标侦察、测绘、野外救援等场景中有着越来越重要的应用。上述应用场景中,为了获取场景的视觉信息,人类需要手动控制无人机飞行并拍摄具有重叠视角的图像数据,该方法在人失去与无人机的联系时失效,且对人类操控无人机的技术要求较高。为将人从复杂、冗长的图像数据采集过程中解放出来,同时使得无人机在未知环境且完全无人工干预的情况下仍能进行环境的探索与恢复场景视觉信息,提出了一种新的自主探索与重建系统,利用RGB-D数据的无人机以当前已探索的区域为边界对未探索的区域进行评估,为无人机生成目标点并规划路径,最终,对已探索的区域输出占位地图与场景重建模型。具体来说:针对无人机在未知环境缺乏地图先验信息的情况下仍要求能自主且快速地进行探索的问题,从经典的快速扩展随机树路径规划算法出发,对其过于随机的采样过程做出改进,考虑未探索体素带来的信息增益,提出基于边界探索价值改进的快速扩展随机树算法,每次采样过程中都贪婪地选取具有最大探索价值的节点加入任务树,利用哈希表进行边界探索价值的存储与查询,从而节省时间消耗,实现场景的快速探索。针对仅通过单次观测就确定环境体素状态对系统带来不健壮的缺陷,提出了基于概率更新的占位地图生成方法以保证体素状态的高置信度。生成的占位地图不仅用于确定未知环境探索的进度,还用于提供无人机进行无碰撞导航的地图。Gazebo仿真结果表明基于边界探索价值改进的快速扩展随机树算法性能优于经典快速扩展随机树算法,探索完成度最大能提升18.9个百分点,单次规划耗时最多减少了13.4%。针对占位地图在视觉上的可理解性与生成速度的矛盾,对点云地图三维表面重建,以达到两者性能的折中。针对泊松表面重建算法对输入点云地图具有水密性的要求,提出了基于泊松重建质量评价的补充采样策略,为待重建点云的低密度处生成置信图,以指导无人机规划用于捕获新点云数据的路径,直到扫描对象的点云是水密的。最后,通过泊松表面重建算法将完整密集对象的点云数据转换为三角形网格表面。实验结果表明,经补充采样策略的泊松重建能保证大规模复杂环境的高精度完整重建,大大降低了单一精细探索的耗时,提供了比单一粗略探索更易理解的场景视觉信息,重建完整度提高7.8个百分点。更多还原

【Abstract】 With the rapid development of computer vision and UAV technology,the technology of using UAV to explore,process and feedback perceptual information in large-scale environment has become more and more important in target reconnaissance,mapping,field rescue and other scenes.In the above application scenarios,in order to obtain the visual information of the scene,humans need to manually control the UAV flight and shoot image data with overlapping visual angles.This method is invalid when humans lose contact with the UAV,and has high technical requirements for human control of the UAV.In order to free people from the complicated and lengthy image data acquisition process,and enable the UAV to explore the environment and restore the scene visual information in an unknown environment without human intervention,this paper proposes a new autonomous exploration and reconstruction system.The UAV with RGB-D camera evaluates the unexplored area with the currently explored area as the boundary,generates target points and plans the path for the UAV,Finally,the occupied map and scene reconstruction model are output for the explored area.Specifically:Aiming at the problem that UAVs still need to be able to explore independently and quickly in the unknown environment without prior map information,this paper starts from the classic fast expanding random tree path planning algorithm,improves its overly random sampling process,considers the information gain brought by unexplored voxels,and proposes an improved fast expanding random tree algorithm based on the value of boundary exploration,Each sampling process greedily selects the node with the greatest exploration value to join the task tree,and uses the hash table to store and query the boundary exploration value,so as to save time and realize the rapid exploration of the scene.In order to solve the problem that only one observation can determine the state of environmental voxels,which is not robust to the system,a method of generating occupancy map based on probability update is proposed to ensure high confidence of voxel state.The generated occupancy map is not only used to determine the progress of location environment exploration,but also used to provide a map for collision free navigation of UAVs.Gazebo simulation results show that the performance of the improved fast expanding random tree algorithm based on the value of boundary exploration is better than that of the classical fast expanding random tree algorithm.The exploration completion can be improved by 18.9 percentage points at most,and the single planning time can be reduced by 13.4% at most.In view of the contradiction between the visual understandability of the occupancy map and the generation speed index,this paper reconstructs the three-dimensional surface of the point cloud map to achieve a compromise between the visual information understandability and the generation speed index.In view of the requirements of Poisson surface reconstruction algorithm for the watertightness of the input point cloud map,this paper proposes a supplementary sampling strategy based on the quality evaluation of Poisson reconstruction,which generates a confidence map for the low density of the point cloud to be reconstructed,to guide the UAV to plan the path for capturing new point cloud data until the point cloud of the scanning object is watertightness.Finally,the point cloud data of complete dense objects are converted into triangular meshes by Poisson surface reconstruction algorithm.The experimental results show that the Poisson reconstruction with supplementary sampling strategy can ensure the high-precision and complete reconstruction of large-scale complex environment,greatly reduce the time consumption of a single fine exploration,provide more understandable scene visual information than a single rough exploration,and improve the reconstruction integrity by 7.8 percentage points.更多还原