文献类型：期刊文献

中文题名：动态场景中基于神经网络特征提取的SLAM

英文题名：SLAM Based on Feature Extraction with Neural Network in Dynamic Scene

作者：孙润[1];刘百川[1];闫伊琳[1];徐卫星[1];和望利[1,2,3]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237;[3]华东理工大学工业控制技术全国重点实验室,上海200237

年份：2025

卷号：32

期号：7

起止页码：1233

中文期刊名：控制工程

外文期刊名：Control Engineering of China

收录：;北大核心:【北大核心2023】；

基金：上海市碳中和基础研究特区项目(22TQ1400100-3);工业控制技术全国重点实验室自主课题(ICT2024A14);自动化集装箱码头“大脑”Port-AI(Ⅰ期)人工智能算法模块研究与应用项目

语种：中文

中文关键词：SLAM;视觉里程计;弱纹理场景;目标检测;动态特征点剔除;多视图几何

外文关键词：SLAM;visual odometry;weak-texture scene;object detection;elimination of dynamic feature point;multiple view geometry

摘要：传统同时定位与地图构建(simultaneous localization and mapping,SLAM)在弱纹理场景中的鲁棒性差,在动态场景中受动态物体干扰。针对这些问题,提出了动态视觉SLAM。首先,在视觉前端使用几何对应网络2(geometric correspondence network version 2,GCNv2)提取特征点并生成二值描述子,提高SLAM在弱纹理场景中的鲁棒性;然后,引入目标检测网络对动态物体进行检测,获取当前帧的语义信息,结合多视图几何剔除动态物体,去除动态物体对SLAM的干扰。实验结果表明:在弱纹理场景中,所提方法可以持续提取足够数量的高质量特征点;在存在动态物体干扰的场景中,所提方法的绝对位姿误差和相对位姿误差较小;在静态场景中,所提方法的性能仍然较优。
Conventional simultaneous localization and mapping(SLAM)has poor robustness in weak-texture scenes and is disturbed by dynamic objects in dynamic scenes.To solve these problems,dynamic visual SLAM is proposed.Firstly,in the visual odometry,the geometric correspondence network version 2(GCNv2)is used to extract feature points and generate binary descriptors,which improves the robustness of SLAM in weak-texture scenes.Then,the object detection network is introduced to detect dynamic objects,obtain the semantic information of the current frame,and combine multiple view geometry to eliminate dynamic objects,which removes the interference of dynamic objects on SLAM.The experimental results show that the proposed method can continuously extract a sufficient number of high-quality feature points in weak-texture scenes.In scenes with the interference of dynamic objects,the absolute pose error and relative pose error of the proposed method are small.In static scenes,the performance of the proposed method is still superior.