基于语义与几何约束的动态SLAM系统
Dynamic SLAM system based on semantic and geometric constraints
丁国强 1赵朋朋 1李涛 1娄泰山1
作者信息
- 1. 郑州轻工业大学电气信息工程学院,河南郑州 450002
- 折叠
摘要
针对传统视觉SLAM在动态环境中出现的漂移、定位丢失和精度差等问题,提出一种基于语义先验信息和几何约束的面向动态环境的视觉SLAM系统.在传统ORB-SLAM2系统框架下加入语义分割模块,以提取物体的先验信息;利用先验的静态特征点进行几何约束,剔除图像中的动态特征点;针对剔除动态特征点会导致特征缺失等问题,加入图像补全线程,合成一个不包含动态物体的图像帧;利用合成帧求解相机位姿.实验结果表明,与传统算法相比,该算法在高动态环境中具有较高精度.
Abstract
Aiming at the problems of drift,localization loss and poor precision in traditional visual SLAM in dynamic environ-ments,a visual SLAM system for dynamic environments based on semantic prior information and geometric constraints was pro-posed.The semantic segmentation module was added under the traditional ORB-SLAM2 system framework to extract the prior information of the object.The static feature points were used for geometric constraints to eliminate the dynamic feature points in the image.The image completion thread was added to synthesize an image frame without dynamic object,aiming at the problem of feature missing caused by removing dynamic feature points.The composite frame was used to solve the camera pose.Experi-mental results show that compared with traditional algorithms,the proposed algorithm has higher accuracy in high dynamic envi-ronment.
关键词
同时定位与建图/动态环境/语义分割/几何约束/运动一致性检测/图像补全/位姿估计Key words
simultaneous localization and mapping/dynamic environment/semantic segmentation/geometric constrained/mo-tion consistency detection/image completion/pose estimation引用本文复制引用
基金项目
2022年河南省教育厅重点研发计划基金项目(22A120010)
2022年河南省教育厅重点研发计划基金项目(22A590003)
出版年
2024
NETL
NSTL
万方数据