煤矿井下移动机器人是煤矿机器人的主力军，煤矿井下移动机器人的自主导航是其研究的难点和热点。 

目前，煤矿井下移动机器人自主导航所必须的三维环境数据库尚未形成，尤其是制作高分辨率、多信息融合的煤矿井下高精度地图还处于研究阶段。 为了有效解决煤矿井下移动机器人自主导航问题，构建了基于深度相机的机器视觉系统，提出了一种基于深度视觉的导航方法，自主导航过程分为地图创建与自主运行两个阶段。 在地图创建中:

① 对深度视觉数据进行特征提取与匹配，利用 10 组煤矿井下真实视频截图，对比测试 5 种特征提取与匹配组合算法，结果表明SURF+SURF+FLANN 与 GFTT+BRIEF+BF 算法能够在煤矿井下获得良好匹配结果;

② 建立煤矿井下移动机器人深度视觉定位与建图问题的捆集调整迭代最近点图模型(Iterative Closest Points Bundle Adjustment，ICP BA);③ 通过图优化方式估计当前观测下的最优位姿与环境路标点坐标。 在实验室场景中利用提出的 ICP-BA 图优化算法，建立了包含关键位姿与三维环境点的原始点云地图。

在自主运行阶段，① 通过八叉树数据结构，将点云地图转化为移动机器人运动规划可使用的 Octomap 导航地图，实验结果表明，Octomap 导航地图分辨率可调、系统资源占用低、索引效率高;② 使用三维到二维映射的视觉图匹配 PNP(Perspective N Points)方法进行实时在线重定位;③ 基于图搜索的 A∗(A Star)路径规划作为轨迹规划初值，自定义最小化能量损失泛函为最小化加加速度的变化率(Minimum-Snap)求解 2 次规划问题，生成用于煤矿井下移动机器人运动执行的轨迹。

 在Matlab 开发环境中设计随机导航地图，生成时间分配、位置、速度、加速度、加加速度的最优轨迹规划结果，验证了运动规划算法的正确性。 通过理论分析和实验验证，表明本文提出的煤矿井下移动机器人深度视觉自主导航方法的有效性。

Underground mobile robot is the main force of coal mine robot. Autonomous navigation is the difficulty and the hotspot task in research. Currently three-dimensional environment database which is necessary for au-tonomous navigation of mobile robots in coal mines has not been fully developed. In particularly,the production of high-resolu- tion,multi-information fused,high-precision maps of underground coal mine is still under investigation. In order to solve the problem of autonomous navigation of mobile robot in underground coal mine,a machine vision system based on depth camera was built,and a navigation method based on depth vision was proposed. The autonomous navigation process have two stages: map creation and autonomous operation. In the stage of map creation,① depth vision data was used for feature extracting and matching. Five depth visual feature extraction and matching algorithms were com- pared and tested in ten groups underground coal mine images. Result shows that the algorithm SURF+SURF+FLANN and GFTT+BRIEF+BF have better performance. ② An Iterative Closest Points Bundle Adjustment model for depth vi- sion based localization and mapping problem of mobile robot in underground coal mine was established. ③ The optimal camera poses and landmarks under current observation were estimated by graph optimization. A laboratory scene origi- nal point cloud map containing key poses was established by using the proposed ICP-BA algorithm. In the stage of au- tonomous operation,① the point cloud map was transformed into an octree data structure Octomap which can be used for mobile robot motion planning. Compared with the original point cloud map,Octomap had adjustable resolution,low system resource occupancy and high indexing efficiency. ② The PNP method of 3 d to 2 d projecting was used for re- al-time online relocation. ③ On these basis,A∗ (A Star) path planning based on graph search was taken as the ini- tial value of trajectory planning,and the customized minimum-energy loss functional ( minimum-snap) was used to solve the quadratic programming problem to generate the trajectory for motion controller. Random navigation map was designed in Matlab development environment,the optimal trajectory planning results of time allocations,positions,ve- locities,accelerations and jerks were generated,which verified the correctness of the proposed motion planning algo- rithm. Through the above theoretical analysis and experimental verification,the effectiveness of the proposed depth vi- sion autonomous navigation method for underground coal mine mobile robot was proved.