摘要
针对室内动态场景下的视觉同步定位与地图构建(VSLAM)问题,提出了一种基于YOLACT实例分割融合光流约束的视觉同步定位与地图构建方法,以降低运动物体对VSLAM系统性能影响。该系统通过自适应阈值的方法提取到均匀分布的ORB特征点,然后利用YOLACT实例分割网络获取动态对象的掩膜,同时使用改进的光流约束对动态点进行检测。将动态点与动态对象掩膜进行匹配之后可以删除动态物体的特征点,之后使用剩余的静态特征点完成相机的位姿估计。最后使用静态区域的图像信息生成点云图,并通过滤波器对点云图进一步优化,同时引用八叉树存储点云,建立八叉树地图。在TUM数据集室内动态场景和真实室内动态场景下进行测试,相较于ORB-SLAM3算法,所提VSLAM算法在低动态场景中的定位精度有10%以上的提升,在高动态场景中对比DS-SLAM算法,也有5%左右的定位精度提升,验证了所提方法在室内动态场景下的可行性和有效性。
Aiming at the problem of visual simultaneous localization and mapping(VSLAM)in indoor dynamic scenes,a VSLAM method based on YOLACT instance segmentation and optical flow constraint is proposed to reduce the impact of moving objects on the performance of VSLAM system.The system extracts evenly distributed ORB feature points by adaptive threshold method,and then obtains the mask of dynamic objects by using YOLACT instance segmentation network.At the same time,the dynamic points are detected by using the improved optical flow constraint method.After matching the dynamic points with the dynamic object mask,the feature points of the dynamic object can be deleted,and then the remaining static feature points are used to complete the pose estimation of the camera.Finally,the image information of the static area is used to generate the point cloud map,and the point cloud map is further optimized by the filter.The octree is used to store the point cloud and establish the Octo-Map.The test in TUM data set indoor dynamic scene and real indoor dynamic scene shows that,the proposed VSLAM algorithm improves the positioning accuracy by more than 10%in low dynamic scenes compared with the ORB-SLAM3 algorithm and about 5%in the high dynamic scene compared with the DS-SLAM algorithm.The feasibility and effectiveness of the proposed method in indoor dynamic scenarios are verified.
作者
陈志环
王祖傲
李想成
CHEN Zhihuan;WANG Zuao;LI Xiangcheng(Engineering Research Center for Metallurgical Automation and Measurement Technology of Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,China;Institute of Robotics and Intelligent Systems,Wuhan University of Science and Technology,Wuhan 430081,China)
出处
《中国惯性技术学报》
EI
CSCD
北大核心
2023年第4期390-400,共11页
Journal of Chinese Inertial Technology
基金
国家自然科学基金资助项目(62203339,62173262,62073250,62003249)
湖北省重点研发计划项目(2020BAB021)。
关键词
视觉同步定位与地图构建
动态场景
实例分割
动态特征点过滤
稠密地图
visual simultaneous localization and mapping
dynamic scenes
instance segmentation
dynamic feature point filtering
dense map
