Research On SLAM Algorithm Of Dynamic Vision Based On Case Segmentation And Optical Flow

In recent years,with the rapid development of visual navigation technology,SLAM(Simultaneous Localization and Mapping)technology has gradually become a hot research direction.SLAM technology has the function of map building for unknown scenes,and can also realize autonomous positioning of the platform.Visual SLAM system needs to obtain image information from the scene in the process of positioning itself.Mobile robots based on traditional SLAM design are generally oriented to static scenes,but the scenes faced by robots are very complex in application,which easily interfere with dynamic targets,affect the accuracy of feature matching,and reduce the positioning accuracy and robustness.In order to improve the positioning accuracy and robustness of visual SLAM system for dynamic scenes,this paper relies on the special project "Research on Key Technologies of Police Patrol Robot Based on Vision Navigation",which is a provincial key field research and development plan,and completes the following work:Firstly,a Mask-R-CNN network model based on Hdfi-FPN is proposed to improve the accuracy of object edge segmentation by instance segmentation.At the same time,a soft NMS(Non Maximum Suppression)algorithm based on the second-order difference equation is proposed to improve the accuracy and speed of selecting the best candidate box.Finally,a comparative experiment is designed to test the accuracy of Mask R-CNN network model fused with Hdfi-FPN.Experiments show that the improved Mask R-CNN network has higher segmentation accuracy for target edges.Then,aiming at the problem that visual SLAM is easily disturbed by dynamic objects in the scene when camera motion estimation,a visual odometer is designed,which combines traditional feature point detection method with instance segmentation and dense optical flow.Dynamic objects are detected based on optical flow velocity comparison,and dynamic object operators are eliminated before pose calculation,so as to improve the accuracy of pose calculation of SLAM system.The algorithm firstly uses Mask R-CNN for instance segmentation,then uses dense optical flow method to accurately track and measure the target velocity in the candidate frame,and uses sparse optical flow method to calculate the static background velocity,and then compares the target velocity with the background velocity to judge whether the object is in motion.The ORB(Oriented FAST and Rotated BRIEF)feature points in the candidate frame judged as dynamic objects are eliminated,and the static feature points are retained,and then feature matching and pose estimation are carried out.Then,aiming at the problem that the dynamic object interferes with visual SLAM location in the dynamic scene,a fusion dynamic object detection algorithm based on ORBSLAM2 system is proposed,which eliminates the dynamic feature points,thus reducing the interference to visual SLAM location and improving the robustness of visual SLAM in the dynamic scene.Finally,the performance of visual ORB-SLAM2 system with dynamic target detection is verified on the public data set.The experimental results show that it can accurately estimate the pose of the camera in high dynamic scenes,and the test trajectory is basically consistent with the real trajectory,which improves the RMSE value by more than 80%and other indexes significantly.Therefore,the vision ORB-SLAM 2 system with dynamic target detection can greatly improve the positioning drift of ORB-SLAM 2 in the face of dynamic scenes,and improve the accuracy of the original algorithm.It is proved that the dynamic vision SLAM algorithm based on instance segmentation and optical flow can improve the positioning accuracy and robustness of the system.