摘要
视觉SLAM系统在相机快速旋转或光照频繁变化时,极易跟踪丢失。为此,提出一种基于直接法和共视图优化的紧耦合视觉惯性SLAM系统,融合IMU信息提高系统的鲁棒性,采用直接法前端提高系统的实时性,共视图优化后端提高系统的定位精度。该系统由前端和后端以及回环检测三个模块组成。跟踪线程利用IMU信息和基于稀疏图像对齐的直接法进行初始位姿估计;后端采用共视图的方法,以当前帧的二级相邻共视关键帧范围为局部优化窗口,利用光束平差法(Bundle Adjustment,BA)对系统状态变量进行优化;另外,仅对关键帧提取ORB特征点,并计算描述子信息供回环检测使用。在TUM VI数据集上的实验证明,与ORB-SLAM3和VINS-mono相比,该算法提高了系统的定位精度,且位姿估计速度提高了50%以上,在一帧完整跟踪任务中,比VINS-mono实时性提高了26%。
Visual SLAM system is easy to track and lose when the camera rotates rapidly or the light changes frequently.Therefore,a tightly coupled visual inertial SLAM system based on direct method and covisibility graph optimization is proposed.The robustness of the system is improved by fusing IMU information,the real-time performance of the system is improved by using the direct method front end,and the positioning accuracy of the system is improved by using the covisibility graph optimization back end.The system consists of front-end,back-end and loop detection.The tracking thread uses IMU information and direct method based on sparse image alignment to estimate the initial pose.The back-end uses the covisibility graph method to optimize the state of the camera and IMU by using the Bundle Adjustment(BA)with the range of the two adjacent covisibility graph key frames of the current frame as the local optimization window.In addition,only ORB feature points are extracted for key frames,and the description sub-information is calculated for loop detection.Experiments on TUM VI dataset show that the proposed algorithm improves the positioning accuracy of the system compared with ORB-SLAM3 and VINS-mono,and the pose estimation speed is more than doubled.The real-time performance is 26%higher than that of VINS-mono in a complete tracking task.
作者
张有全
祁宇明
邓三鹏
孙建康
王帅
ZHANG Youquan;QI Yuming;DEND Sanpeng;SUN Jiankang;WANG Shuai(Institute of Robotics and Intelligent Equipment,Tianjin University of Technology and Education,Tianjin 300222,China;Tianjin Key Laboratory of Intelligent Robot Technology and Application,Tianjin 300222,China)
出处
《自动化与仪器仪表》
2022年第5期197-203,共7页
Automation & Instrumentation
基金
天津市科技军民融合重大专项(18ZXJMTG00160)
全国职业院校教师教学创新团队建设体系化课题研究项目(TX20200104)。
关键词
视觉惯性SLAM系统
紧耦合
直接法前端
共视图
visual inertial SLAM System
tightly coupled
direct method front end
covisibility graph
