基于平面匹配与目标检测的视觉SLAM算法
Visual SLAM algorithm based on plane matching and object detection
娄路 1张忍 1李一天 1隗寒冰 2王桂平1
作者信息
- 1. 重庆交通大学信息科学与工程学院,重庆 400074
- 2. 重庆交通大学机电与车辆工程学院,重庆 400074
- 折叠
摘要
针对移动机器人同时定位与建图(SLAM)算法在动态复杂、低纹理场景中存在算法精度下降甚至无法正常工作等问题,提出一种基于快速运动目标检测与平面匹配的算法.将视觉特征提取与运动目标检测方法相结合,降低动态目标对定位的干扰;采用平面匹配技术,克服低纹理环境特征缺少问题;同时生成一个稠密的三维点云地图,用于机器人环境解析等应用.该算法在数据集KITTI和TUM上的绝对轨迹误差RMSE指数相对于ORB-SLAM2算法分别降低了66.67％、98.77％,算法运行速率为22.20 fps.结果表明该算法具有良好定位精度、运行效率和鲁棒性.
Abstract
Aiming at the problem that the simultaneous localization and mapping algorithm of mobile robots has poor accuracy and even fails to work normally in dynamic,complex and low texture scenes,an algorithm based on fast moving object detection and plane matching was proposed.Visual feature extraction and moving target detection were combined to reduce the interference of dynamic target to SLAM localization.Plane matching technology was used to overcome the lack of features in low-texture envi-ronment,so as to improve the positioning accuracy in dynamic and complex environment.At the same time,a dense 3D recon-structed point cloud map was generated,which was used for robot environment analysis and other applications.Compared with the ORB-SLAM2 algorithm,the absolute trajectory error RMSE index on KITTI and TUM decreases by 66.67％and 98.77％respectively,and the algorithm runs at 22.20 fps.Results show that the algorithm has good positioning accuracy,running efficiency and robustness.
关键词
移动机器人/同时定位与建图/特征提取/目标检测/平面匹配/定位精度/稠密三维重建Key words
mobile robot/SLAM/feature extraction/object detection/plane matching/positioning accuracy/dense 3D recon-struction引用本文复制引用
基金项目
国家自然科学基金面上基金项目(52172381)
重庆市自然科学基金面上基金项目(cstc2021jcyjmsxmX1121)
出版年
2024
NETL
NSTL
万方数据