基于YOLACT网络的移动机器人视觉SLAM算法

YOLACT network⁃based algorithm of visual SLAM of mobile robots

提出一种室内动态场景的视觉SLAM算法,引入实例分割网络YOLACT,剔除大部分动态点,利用多视图几何进一步过滤分割掩膜外未被剔除的动态特征点,使用剩余的静态特征点作为相机位姿估计;同时构建点云地图,转换并建立八叉树地图;使用背景修复以恢复被剔除动态物体后的背景。为验证算法的有效性,使用TUM数据集测试,并与ORB-SLAM2算法和其他处理动态场景的SLAM算法对比,结果表明,提出的算法在高动态数据集上表现良好。相较于ORB-SLAM2算法,提出的算法在室内动态场景中的定位精度提升93.06%,可应用于后期机器人定位导航使用。

A visual SLAM algorithm for indoor dynamic scenes was proposed.The instance segmentation net-work YOLACT was introduced to eliminate most of the dynamic points.The multi-view geometry was used to further filter the dynamic feature points that were not eliminated outside the segmentation mask.The remaining static feature points were used as camera pose estimation.At the same time,the point cloud map was construc-ted,the octree map was transformed and established;background repair was used to restore the background after dynamic objects were removed.Finally,in order to verify the effectiveness of the proposed algorithm,the TUM dataset was used for testing,and compared with the ORB-SLAM2 algorithm and other SLAM algorithms processing dynamic scenarios,and results show that the proposed algorithm performs well on the highly dyna-mic dataset.Compared with the ORB-SLAM2 algorithm,the positioning accuracy of the proposed algorithm in indoor dynamic scenes is improved by 93.06%,and it can be applied to the later use of robot positioning and navigation.