融合双目视觉与惯导信息的高效视觉里程计算法  

Efficient visual odometry algorithm combining stereo vision and inertial navigation information

作　　者：潘林豪 田福庆[1] 应文健[1] 佘博[1] Pan Linhao;Tian Fuqing;Ying Wenjian;She Bo(College of Weapon,Naval University of Engineering,Wuhan 430033,China)

机构地区：[1]海军工程大学兵器院,武汉430033

出　　处：《计算机应用研究》2021年第6期1739-1743,1769,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目;海军工程大学自主立项基金资助项目。

摘　　要：为提高视觉里程计(VO)在大尺度环境下运行的实时性,提出一种融合双目视觉与惯导信息的视觉里程计算法,主要由前端位姿跟踪和后端局部地图优化两个线程组成。位姿跟踪线程首先使用惯导信息辅助光流法进行帧间特征点跟踪并估计相机初始位姿;接着通过最小化图像光度误差获取当前帧像素点与局部地图点的对应关系;而后最小化当前帧上局部地图点的重投影误差和惯性测量单元(IMU)预积分误差,得到当前帧准确的位姿估计。后端局部地图优化线程对滑动窗口内的关键帧提取特征点并三角化新地图点,使用光束平差法(BA)对逆深度参数化表示的地图点位置、关键帧位姿、速度以及陀螺仪和加速度计零偏进行滑窗优化,为前端提供更加精确的局部地图相机位姿和环境信息。在EuRoC数据集上的实验表明,相比于ORB-SLAM2、ORB-SLAM3算法,该融合双目视觉与惯导信息的视觉里程计算法的定位精度略有下降,但可以较大程度地提高位姿跟踪的实时性。In order to improve the real-time performance of VO(visual odometry)in large-scale environment,this paper proposed a VO algorithm based on stereo vision and inertial navigation information.The algorithm consisted of two threads:the front-end tracking thread and the back-end local mapping thread.The tracking thread firstly used the inertial navigation information assisted optical flow method to track the features between frames and estimate the initial pose of the camera.Then,it obtained the corresponding relationship between the pixels of the current frame and the local map points by minimizing the photometric error,and obtained the accurate pose estimation of the current frame by minimizing the re-projection error of the local map points on the current frame and pre-integration error of the IMU(inertial measurement unit).The back-end local mapping thread extracted features from the keyframes in the sliding window and triangulated new map points.It used BA(bundle adjustment)to optimize the map point positions,keyframe poses,velocities and gyroscope and accelerometer biases in the sliding window,which could provide more accurate local pose trajectory and environment information for the front-end.The experiment on the EuRoC datasets shows that,the positioning accuracy of the proposed algorithm is slightly reduced,however it can greatly improve the real-time performance of pose tracking comparing with the ORB-SLAM2 algorithm and the ORB-SLAM3 algorithm.

关 键 词：同时定位与地图构建 特征点法 直接法 传感器融合 光束平差法 

分 类 号：TP242.62[自动化与计算机技术—检测技术与自动化装置]