视觉同步定位与建图(visual simultaneous localiza-tion and mapping,VSLAM)技术是近年来机器人和计算机视觉领域的重点研究方向之一,但当前的主流算法主要面向静态环境,当场景中存在运动的物体时,算法的定位精度和稳定性会受到很大影响.为了解决上述问题,提出了一种惯性测量单元(inertial measurement unit,IMU)积分与 YO-LOv4语义分割结合的VSLAM前端动态特征点剔除算法,通过YOLOv4网络对图像进行语义分割,识别图像中有运动可能的物体;再将IMU积分与语义分割结合,对目标检测框内有运动可能的特征点进行重投影误差的解算,识别并剔除环境中运动的特征点.在TUM Visual-Inertial Dataset上验证该算法,结果表明,在包含运动物体的室内场景下,该算法可以有效剔除环境中的运动物体,显著提升SLAM系统的定位精度和稳定性.
Deep Learning-Based Visual SLAM Technology for Indoor Dynamic Scenes
Visual simultaneous localization and mapping(VSLAM)technology is one of the key research directions in the field of robotics and computer vision in recent years,but the current algorithms are mainly oriented to static environ-ment.When there are moving objects in the scene,the posi-tioning accuracy and stability of the algorithm will be greatly af-fected.To solve the above problems,we propose a VSLAM algorithm that combines inertial measurement unit(IMU)in-tegration and YOLOv4 semantic segmentation to eliminate front-end dynamic feature points.YOLOv4 network is used to perform semantic segmentation of images and identify ob-jects with possible movement in images.Then IMU integra-tion and semantic segmentation are combined to solve the re-projection error of the feature points with possible movement in the target detection frame,and identify and eliminate the moving feature points in the environment.The TUM Visual-Inertial Dataset is used to verify the proposed algorithm,and the experimental results show that the proposed algorithm can effectively remove the moving objects in the indoor scene con-taining moving objects,significantly improving the positioning accuracy and stability of SLAM system.
visual simultaneous localization and mapping(VSLAM)feature pointdynamic objectdeep learning
郑晓华、耿鑫雷、邓浩坤
展开 >
中铁一局集团建筑安装工程有限公司,陕西 西安,710054
阿里巴巴(北京)软件服务有限公司,北京,100015
测绘遥感信息工程国家重点实验室深圳研发中心,广东 深圳,518057
视觉同步定位与建图(visual simultaneous localization and mapping,VSLAM) 特征点 动态目标 深度学习
NETL
NSTL
万方数据
