摘要
为了实现系统在动态场景中准确定位和构建包含语义信息的地图,提出一种基于目标检测算法改进的视觉SLAM系统。该系统在ORB-SLAM2的跟踪线程中,采用目标检测算法YOLOv4去除动态特征提高系统位姿估计精度。建图线程中,对点云地图进行超体素聚类,与YOLOv4获取的物体标签融合构建语义地图。实验结果表明该视觉SLAM系统有效地去除动态特征,减少位姿估计误差,构建分层清晰的语义地图,且计算速度达到15帧/s,满足实时性要求。
In order to accurately locate in the dynamic scene and construct the map with semantic information,an improved visual SLAM system based on object detection algorithm is proposed.In the tracking thread of ORB-SLAM2,the system used the object detection algorithm YOLOv4 to remove dynamic features and improve the accuracy of pose estimation.In the mapping thread,the point cloud map was clustered by supervoxel,and the semantic map was constructed by merging the object labels obtained by YOLOv4.The experimental results show that the visual SLAM system can effectively eliminate the dynamic features,reduce the error of pose estimation and construct a clear and hierarchical semantic map,and the calculation speed reaches 15 fps,which meets the real-time requirements.
作者
王晓超
王春林
袁成祥
Wang Xiaochao;Wang Chunlin;Yuan Chengxiang(College of Automation(Artificial Intelligence),Hangzhou Dianzi University,Hangzhou 310018,Zhejiang,China;College of Computer and Information Engineering,Zhejiang Gongshang University,Hangzhou 310018,Zhejiang,China)
出处
《计算机应用与软件》
北大核心
2023年第5期214-220,共7页
Computer Applications and Software
基金
浙江省自然科学基金项目(ZL20f030013)。
关键词
同时定位与地图构建
动态场景
深度学习
目标检测
语义地图
Simultaneous localization and mapping(SLAM)
Dynamic scene
Deep learning
Object detection
Semantic map
作者简介
王晓超,硕士生,主研领域:视觉SLAM,深度学习;王春林,副教授;袁成祥,讲师。
