一种动态场景下的双目视觉里程计算法  

A Stereo Visual Odometry Algorithm in Dynamic Scenes

作　　者：王子国 黄劲松[1] WANG Ziguo;HUANG Jinsong(School of Geodesy and Geomatics,Wuhan University,Wuhan 430079,China)

机构地区：[1]武汉大学测绘学院,湖北武汉430079

出　　处：《测绘地理信息》2023年第3期121-125,共5页Journal of Geomatics

基　　金：国家重点研发计划(2016YFB0501803)。

摘　　要：在同步定位与地图构建领域,视觉里程计因其传感器成本低廉、定位精度高、蕴含丰富环境信息等优势而得到广泛应用,但环境中的动态物体会显著降低其解算的精度与鲁棒性。针对此问题,提出了一种深度学习辅助的双目视觉里程计,基于几何约束与目标位置约束的方法,能够准确筛选出环境中的动态物体,以提高结果精度,同时能够恢复动态物体的运动轨迹,便于建立语义地图。在KITTI和TUM RGB-D的动态场景数据集上进行测试,得到的结果精度较高。In the field of simultaneous localization and map⁃ping,visual odometry is widely used because of its advantages of low cost,high positioning accuracy and rich environmental information.However,dynamic objects in the environment will significantly reduce the accuracy and robustness of the so⁃lution.Therefore,we propose a binocular visual odometry aided by deep learning.Based on geometric constraints and target location constraints,the method can accurately screen out the dynamic objects in the environment to improve the ac⁃curacy of the results.At the same time,it can restore the mo⁃tion trajectories of dynamic objects and facilitate the establish⁃ment of semantic map.The tests on the dynamic scene data set of KITTI and TUM RGB-D indicate high precision re⁃sults.

关 键 词：视觉里程计 实例分割 对极几何 动态目标检测 语义SLAM 

分 类 号：P20[天文地球—测绘科学与技术] P28[自动化与计算机技术—计算机应用技术] TP391[自动化与计算机技术—计算机科学与技术]