基于语义分割和光流加速的动态场景ORB-SLAM算法
Dynamic Scene ORB-SLAM Algorithm Based on Semantic Segmentation and Optical Flow Acceleration
赵健 1李虹1
作者信息
- 1. 太原科技大学电子信息工程学院,太原 030024
- 折叠
摘要
针对同步定位与地图建立(simultaneous localization and mapping,SLAM)算法在动态环境下存在位姿估计和地图构建误差较大的问题,提出一种光流语义分割方法用于增加动态场景下的可运行性.将ORB-SLAM2系统与YOLOv5模型结合,对传入图像提取特征点的同时将YOLOv5网络模型语义分割后的物体分为高、中、低动态物体.利用运动一致性检测算法,对三种检测物体动态阈值判断,辨别其是否需要剔除特征点,增加ORB-SLAM2算法在动态环境下的运行精度.为加快系统运行速度,用LK光流法加快普通帧与普通帧之间的匹配,其原理为使用LK光流匹配特征点代替ORB特征点匹配,大大的缩小运行时间,同时运行误差变化不大.实验在TUM数据集下测试,平均每一帧提取2000个特征点,在增加LK光流后缩短0.01 s以上,若在900帧数据集下,可缩短9 s.其绝对轨迹误差对比于ORB-SLAM2和DS-SLAM平均提升在95％与30％以上,证明了算法在动态场景下良好的运行精度与鲁棒性.
Abstract
Simultaneous localization and mapping(SLAM)algorithm has the problem of large errors in pose esti-mation and map construction in dynamic environment.This paper proposes an optical flow semantic segmentation method to increase the operability in dynamic scene.Combining orb-slam2 system with YOLOv5 model,the detec-ted objects are divided into high,medium and low dynamic objects by semantic segmentation of YOLOv5 network model while extracting feature points from the incoming image.The motion consistency detection algorithm is used to judge the dynamic threshold of three detected objects,whether it needs to eliminate feature points,on the tum data set,and the LK optical flow method is used to speed up the matching between ordinary frames,greatly reduce the running time,and the running error changes little.The experiment is tested on the tum data set.The average time of each frame is shortened by more than 0.01s after increasing LK optical flow.Compared with orb-slam2 and ds-slam,the average absolute trajectory error is improved by more than 95％and 30％,which proves the good oper-ation accuracy and robustness of the algorithm in dynamic scenes.
关键词
动态环境/语义分割/一致性检测算法/LK光流/动态阈值Key words
dynamic environment/semantic segmentation/consistency detection algorithm/LK optical flow/dynam-ic threshold引用本文复制引用
基金项目
山西省重点研发计划(201803D121123)
出版年
2024
NETL
NSTL
维普
万方数据