摘要
针对基于特征点法的视觉里程计实时性和鲁棒性较差的问题,提出一种基于半直接法的双目视觉惯性里程计算法.在跟踪线程中将惯性测量数据作为先验,并使用逆光流法跟踪均匀化的特征关键点,以提高特征匹配的速度和鲁棒性,构建高精度的初始化地图,为后续的运动估计提供良好的初值.使用简化的双目视觉模型构造重投影误差,结合IMU误差项构建联合优化模型,并在滑动窗口中进行非线性优化求解.实验结果显示,该算法在数据集上的定位精度达到主流算法的水平,与VINS-Fusion算法相比,此算法拥有更低的CPU负载和更高的运行帧率.
A stereo visual inertial odometry based on semi-direct method has been proposed to improve the poor real-time performance and robustness of visual odometry based on feature-based method.The inertial measurement data is used as a priori in the tracking thread,and the reverse optical flow method is used to track the homogenized feature key points to improve the speed and robustness of feature matching.A high-precision initialization map is constructed to provide a accurate initial value for the subsequent motion estimation.The joint optimization model,which is constructed by combining the reprojection error which is constructed by a simplified stereovision model and IMU error,is solved by nonlinear-optimization in the sliding window.The experimental results show that positioning accuracy of the proposed algorithm reaches the level of the mainstream algorithm.Compared with the VINS-Fusion,our algorithm in this paper has lower CPU load and higher running frequency.
作者
种一帆
冀杰
宫铭钱
陈琼红
CHONG Yi-fan;JI Jie;GONG Ming-qian;CHEN Qiong-hong(College of Engineering and Technology,Southwest University,Chongqing 400715,China)
出处
《西南师范大学学报(自然科学版)》
CAS
2021年第2期112-120,共9页
Journal of Southwest China Normal University(Natural Science Edition)
基金
国家自然科学基金(61304189)
中央高校基本业务费专项资金重点项目(XDJK2019B053)
汽车主动安全测试技术重庆市工业和信息化重点实验室2019年度开放课题(19AKC8).
关键词
实时性
鲁棒性
简化的双目视觉模型
初始化地图
非线性优化
real-time performance
robustness
simplified stereo visual model
initialization map
nonlinear-optimization