Dynamic Visual SLAM System Integrating Optical Flow and Multi-View Geometry
Visual Simultaneous Localization and Mapping(SLAM)reduces the positioning accuracy and cannot accurately construct static maps under dynamic interference.A dynamic visual SLAM system combining optical flow and multi-view geometry is proposed,which is improved based on ORB-SLAM2.It introduces processed optical flow information into the tracking thread,which,when combines with multi-view geometry,yields dynamic-region masks for segmenting image frames in the field of view,thus achieving dynamic-region detection and the filtering of feature points in dynamic regions.This improves the tracking accuracy while ensuring the real-time performance of the visual SLAM system by replacing the original map's construction thread.In the new map's construction thread,optical flow information and the MobileNetV2 instance segmentation network are introduced.By combining the segmentation results of the instance segmentation network with the optical flow dynamic-region mask,an ordered point cloud is obtained and segmented by layer to solve the"dragging"issue caused by dynamic objects during map construction.Simultaneously,semantic information is fused into the segmented point-cloud cluster to construct a static semantic OctoMap.Experimental results on the TUM Dynamic Objects dataset show that compared with ORB-SLAM2,the positioning accuracy of the proposed algorithm improves by an average of 70.4％,with a maximum improvement of 90％in high dynamic scene sequence testing.
国家科技期刊平台
NSTL
万方数据
