未知室内环境下基于RGB-D和微惯性系统的四旋翼飞行器自主导航方法研究

With the social development, the exploration of unknown environments for human, especially in small space, complex region, no GPS signal place and unreachable place, is one of the most popular and difficult areas for unmanned vehicle. In this project, a method of simultaneous localization and mapping based on RGB-D and MEMS-Inertial system is proposed for autonomous navigation of quadrotor under the unknown and indoor environment. Firstly, due to the complex environment, flight maneuverability and vast vision, a fast and robust feature extraction and matching method to ensure the whole system work robustly in real time is researched. Besides, a new location method based on two images is constructed to guarantee the precision of the vision odometry. Furthermore, since the 3D point cloud and vision are affected by quadrotor, the environment of 3D point clouds can be reconstructed by the distortion model and geometric model of the Kinect’s RGB-camera and depth camera. To update the map, the loop closure detection using keyframe is applied, and a pyramid scoring match scheme is presented to improve the accuracy and efficiency of the loop closure detection. After that, focusing on the errors of MEMS-Inertial and RGB-D, a tightly coupled integrated navigation method based on higher order non-linear optimization is proposed to ensure the drift free state estimation. Finally, the online autonomous navigation system is achieved by quadrotor.

随着社会发展，对未知环境特别是空间狭小、环境复杂、无GPS信号、人很难到达的环境的探索，成为小型无人飞行器领域研究的热点和难点。本课题针对未知室内环境下，拟采用基于RGB-D/惯性系统的位姿估计与三维地图重建方法进行自主导航。首先，针对工作环境复杂、飞行机动性大和视觉图像信息量大，研究一种快速鲁棒特征点跟踪及匹配方法保证系统的实时性和鲁棒性，并构建一种新的基于双帧图像的定位估计方法，以提高视觉位姿估计的精度。其次，针对三维点云和视觉受飞行环境影响，结合Kinect系统能同时提供RGB信息和深度信息的特点，建立摄像机畸变模型和几何模型，对环境实现三维点云重建；为实现对地图的更新，基于关键帧技术进行闭环检测，研究一种基于金字塔得分匹配的方法，提高闭环检测的效率和精度。然后，针对微惯性器件和RGB-D误差，研究一种非线性误差最优化的鲁棒高精度紧组合运动估计方法。最后实现四旋翼实时在线自主导航。

本研究提出了一种基于RGB-D与IMU紧耦合的无人机视觉里程计系统，且能够进行实时的三维重建。针对四旋翼在飞行过程中视觉图形信息量大，系统实时性要求高等特性，我们研究了一种基于光流法的快速鲁棒的特征点检测以及跟踪方法。同时本系统通过融合RGB-D和IMU，集成IMU和基于关键帧的视觉SLAM，构建非线性紧耦合优化的误差模型，通过最小化点特征的投影误差和IMU的测量残差组成损失函数来优化系统的状态向量，实现了高精度组合运动估计。最后我们利用octmap来降采样每一张深度图像来进行地图快速创建。最后，实验证明我们的系统能够取得很好的精度和鲁棒性，并且能够实时进行稠密地图创建。