为了提高室内动态场景下定位与建图的准确性与实时性,提出了一种基于目标检测的室内动态场景同步定位与建图(simultaneous localization and mapping,SLAM)系统。利用目标检测的实时性,在传统ORB_SLAM2算法上结合YOLOv5目标检测网络识...为了提高室内动态场景下定位与建图的准确性与实时性,提出了一种基于目标检测的室内动态场景同步定位与建图(simultaneous localization and mapping,SLAM)系统。利用目标检测的实时性,在传统ORB_SLAM2算法上结合YOLOv5目标检测网络识别相机图像中的动态物体,生成动态识别框,根据动态特征点判别方法只将识别框内动态物体上的ORB特征点去除,利用剩余特征点进行相机位姿的估计,最后建立只含静态物体的稠密点云地图与八叉树地图。同时在机器人操作系统(robot operating system,ROS)下进行仿真,采用套接字(Socket)通信方式代替ROS中话题通信方式,将ORB_SLAM2算法与YOLOv5目标检测网络相结合,以提高定位与建图的实时性。在TUM数据集上进行多次实验结果表明,与ORB_SLAM2系统相比,本文系统相机位姿精确度大幅度提高,并且提高了每帧跟踪的处理速度。展开更多
视觉同步定位与建图(VSLAM)技术常常用于室内机器人的导航与感知,然而VSLAM的位姿估算方法是针对静态环境的,当场景中存在运动对象时,可能会导致定位和建图失败。针对此问题,提出了一个结合实例分割与聚类的VSLAM系统。所提系统使用实...视觉同步定位与建图(VSLAM)技术常常用于室内机器人的导航与感知,然而VSLAM的位姿估算方法是针对静态环境的,当场景中存在运动对象时,可能会导致定位和建图失败。针对此问题,提出了一个结合实例分割与聚类的VSLAM系统。所提系统使用实例分割网络生成场景中动态对象的概率掩膜,同时利用多视图几何的方法检测场景中的动态点,并将检测到的动态点与获得的概率掩膜匹配之后确定动态物体的精确动态掩膜;利用动态掩膜删除动态物体的特征点,然后利用剩余的静态特征点准确估计摄像机的位置。为了解决实例分割网络欠分割的问题,采用深度填充算法和聚类算法保证动态特征点完全删除。最后,重建图片被动态物体遮挡的背景,在正确的相机位姿下建立静态稠密点云地图。在公开的TUM(Technical University of Munich)数据集上的实验结果表明,在动态环境中,所提系统在保证实时性的同时能实现鲁棒的定位与建图。展开更多
室内动态场景下的同步定位与地图构建(simultaneous localization and mapping,SLAM)系统容易受到运动障碍物的影响,从而导致其位姿估计精度和视觉里程计的稳定性降低。本文提出一种基于YOLOv4目标检测网络的视觉SLAM算法,获取语义信息...室内动态场景下的同步定位与地图构建(simultaneous localization and mapping,SLAM)系统容易受到运动障碍物的影响,从而导致其位姿估计精度和视觉里程计的稳定性降低。本文提出一种基于YOLOv4目标检测网络的视觉SLAM算法,获取语义信息,并利用LK光流法判断动态特征,在传统的ORB-SLAM2系统上将动态特征点剔除,只使用静态特征点来估计相机的位姿;建立稠密点云地图,并转化成节约内存空间的八叉树地图。在TUM公开数据集上对该方法进行测试和评估,实验结果表明:在动态环境下,该系统与ORB-SLAM2相比,相机位姿估计精度提高83%,且减少了生成的环境地图的存储空间,为后续实现机器人导航具有重要意义。展开更多
Automatic Digital Orthophoto Map(DOM)generation plays an important role in many downstream works such as land use and cover detection,urban planning,and disaster assessment.Existing DOM generation methods can generate...Automatic Digital Orthophoto Map(DOM)generation plays an important role in many downstream works such as land use and cover detection,urban planning,and disaster assessment.Existing DOM generation methods can generate promising results but always need ground object filtered DEM generation before otho-rectification;this can consume much time and produce building facade contained results.To address this problem,a pixel-by-pixel digital differential rectification-based automatic DOM generation method is proposed in this paper.Firstly,3D point clouds with texture are generated by dense image matching based on an optical flow field for a stereo pair of images,respectively.Then,the grayscale of the digital differential rectification image is extracted directly from the point clouds element by element according to the nearest neighbor method for matched points.Subsequently,the elevation is repaired grid-by-grid using the multi-layer Locally Refined B-spline(LR-B)interpolation method with triangular mesh constraint for the point clouds void area,and the grayscale is obtained by the indirect scheme of digital differential rectification to generate the pixel-by-pixel digital differentially rectified image of a single image slice.Finally,a seamline network is automatically searched using a disparity map optimization algorithm,and DOM is smartly mosaicked.The qualitative and quantitative experimental results on three datasets were produced and evaluated,which confirmed the feasibility of the proposed method,and the DOM accuracy can reach 1 Ground Sample Distance(GSD)level.The comparison experiment with the state-of-the-art commercial softwares showed that the proposed method generated DOM has a better visual effect on building boundaries and roof completeness with comparable accuracy and computational efficiency.展开更多
家居机器人技术一般应用视觉同步定位与建图(SLAM,Simultaneous Localization and Mapping)来实现定位与构建导航地图,如何实现视觉SLAM系统快速准确定位和构建丰富环境信息的地图已经成为视觉SLAM研究的热点问题。本文将光流法与关键...家居机器人技术一般应用视觉同步定位与建图(SLAM,Simultaneous Localization and Mapping)来实现定位与构建导航地图,如何实现视觉SLAM系统快速准确定位和构建丰富环境信息的地图已经成为视觉SLAM研究的热点问题。本文将光流法与关键点结合,加快视觉SLAM的数据处理速度,并添加稠密点云地图和八叉树地图构建线程来获取环境信息,实现一个较为优秀的视觉SLAM系统。在公开数据集上进行的定位实验表明,该视觉SLAM系统在绝对轨迹误差和相对位姿误差上与ORB-SLAM2系统保持基本一致,并且在其中几项数据中具有更小的误差结果,整体系统对图像的处理速度约为40 FPS(Frames Per Second),是ORB-SLAM2系统的1.4倍左右,说明该系统在提高系统速度的基础上保持了较高的准确度。展开更多
文摘视觉同步定位与建图(VSLAM)技术常常用于室内机器人的导航与感知,然而VSLAM的位姿估算方法是针对静态环境的,当场景中存在运动对象时,可能会导致定位和建图失败。针对此问题,提出了一个结合实例分割与聚类的VSLAM系统。所提系统使用实例分割网络生成场景中动态对象的概率掩膜,同时利用多视图几何的方法检测场景中的动态点,并将检测到的动态点与获得的概率掩膜匹配之后确定动态物体的精确动态掩膜;利用动态掩膜删除动态物体的特征点,然后利用剩余的静态特征点准确估计摄像机的位置。为了解决实例分割网络欠分割的问题,采用深度填充算法和聚类算法保证动态特征点完全删除。最后,重建图片被动态物体遮挡的背景,在正确的相机位姿下建立静态稠密点云地图。在公开的TUM(Technical University of Munich)数据集上的实验结果表明,在动态环境中,所提系统在保证实时性的同时能实现鲁棒的定位与建图。
文摘室内动态场景下的同步定位与地图构建(simultaneous localization and mapping,SLAM)系统容易受到运动障碍物的影响,从而导致其位姿估计精度和视觉里程计的稳定性降低。本文提出一种基于YOLOv4目标检测网络的视觉SLAM算法,获取语义信息,并利用LK光流法判断动态特征,在传统的ORB-SLAM2系统上将动态特征点剔除,只使用静态特征点来估计相机的位姿;建立稠密点云地图,并转化成节约内存空间的八叉树地图。在TUM公开数据集上对该方法进行测试和评估,实验结果表明:在动态环境下,该系统与ORB-SLAM2相比,相机位姿估计精度提高83%,且减少了生成的环境地图的存储空间,为后续实现机器人导航具有重要意义。
基金supported by the National Natural Science Foundation of China[Grant No.41771479]the National High-Resolution Earth Observation System(the Civil Part)[Grant No.50-H31D01-0508-13/15]the Japan Society for the Promotion of Science[Grant No.22H03573].
文摘Automatic Digital Orthophoto Map(DOM)generation plays an important role in many downstream works such as land use and cover detection,urban planning,and disaster assessment.Existing DOM generation methods can generate promising results but always need ground object filtered DEM generation before otho-rectification;this can consume much time and produce building facade contained results.To address this problem,a pixel-by-pixel digital differential rectification-based automatic DOM generation method is proposed in this paper.Firstly,3D point clouds with texture are generated by dense image matching based on an optical flow field for a stereo pair of images,respectively.Then,the grayscale of the digital differential rectification image is extracted directly from the point clouds element by element according to the nearest neighbor method for matched points.Subsequently,the elevation is repaired grid-by-grid using the multi-layer Locally Refined B-spline(LR-B)interpolation method with triangular mesh constraint for the point clouds void area,and the grayscale is obtained by the indirect scheme of digital differential rectification to generate the pixel-by-pixel digital differentially rectified image of a single image slice.Finally,a seamline network is automatically searched using a disparity map optimization algorithm,and DOM is smartly mosaicked.The qualitative and quantitative experimental results on three datasets were produced and evaluated,which confirmed the feasibility of the proposed method,and the DOM accuracy can reach 1 Ground Sample Distance(GSD)level.The comparison experiment with the state-of-the-art commercial softwares showed that the proposed method generated DOM has a better visual effect on building boundaries and roof completeness with comparable accuracy and computational efficiency.
文摘家居机器人技术一般应用视觉同步定位与建图(SLAM,Simultaneous Localization and Mapping)来实现定位与构建导航地图,如何实现视觉SLAM系统快速准确定位和构建丰富环境信息的地图已经成为视觉SLAM研究的热点问题。本文将光流法与关键点结合,加快视觉SLAM的数据处理速度,并添加稠密点云地图和八叉树地图构建线程来获取环境信息,实现一个较为优秀的视觉SLAM系统。在公开数据集上进行的定位实验表明,该视觉SLAM系统在绝对轨迹误差和相对位姿误差上与ORB-SLAM2系统保持基本一致,并且在其中几项数据中具有更小的误差结果,整体系统对图像的处理速度约为40 FPS(Frames Per Second),是ORB-SLAM2系统的1.4倍左右,说明该系统在提高系统速度的基础上保持了较高的准确度。
