基于YOLOv5的动态场景视觉SLAM研究  

Research on dynamic scene visual SLAM based on deep learning

作　　者：仉新 朱文辉 张旭阳 靳德利 左依林 ZHANG Xin;ZHU Wenhui;ZHANG Xuyang;JIN Deli;ZUO Yilin(School of Mechanical Engineering,Shenyang Ligong University,Shenyang 110158,China)

机构地区：[1]沈阳理工大学机械工程学院,辽宁沈阳110158

出　　处：《通信与信息技术》2024年第5期35-40,79,共7页Communication & Information Technology

基　　金：辽宁省教育厅面上青年人才项目(项目编号:LJKZ0258);2022年辽宁省科技厅博士科研启动基金计划项目(项目编号:2022-BS-187);沈阳理工大学高层次人才科研支持计划(项目编号:1010147001012)。

摘　　要：为解决动态环境下视觉SLAM系统准确性低和鲁棒性差的问题,提出基于深度学习的动态场景视觉SLAM方法。首先,利用改进的YOLOv5算法和光流约束算法剔除动态目标上的特征点,只保留不影响建图的静态特征点;然后,采用GMS-RANSAC融合算法剔除误匹配,准确估计相机位置和姿态;最后,通过融合词袋和YOLOv5目标检测算法,达到提高回环检测的效率和准确性的目的。通过使用TUM数据集中高动态的walking系列进行实验,实验结果表明,该方法能够在动态环境中保持稳定的鲁棒性和准确性。To address the issues of low accuracy and poor robustness of visual SLAM systems in dynamic environments,a deep learning based dynamic scene visual SLAM method was proposed.Firstly,using improved YOLOv5 algorithm and optical flow con⁃straint algorithm to remove feature points on dynamic targets.Only retain static feature points that do not affect the construction of the map.Then,the GMS-RANSAC fusion algorithm was used to eliminate mismatches and accurately estimate the camera pose.Finally,by integrating word bags and YOLOv5 object detection algorithm,the efficiency and accuracy of loop detection have been improved.By conducting experiments using the highly dynamic walking series in the TUM dataset,the experimental results show that the proposed method can maintain stable robustness and accuracy in dynamic environments.

关 键 词：SLAM 动态场景 YOLOv5 GMS-RANSAC 回环检测 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]