Research On Visual Slam Based On Semantic Information In Dynamic Environments

Simultaneous Localization and Mapping(SLAM)is one of the key technologies in the field of mobile robotics.Most mature visual SLAM algorithms assume that the scene is static.When there are moving objects in the scene,the feature matching will be interfered,which will have a great impact on the positioning and mapping accuracy of the SLAM system.Therefore,in a dynamic environment,the SLAM algorithm needs to identify the moving objects in the scene.This dissertation designs a visual SLAM algorithm to improve the positioning accuracy of the algorithm in a dynamic environment,and generates a semantic map of the dynamic environment.The main research contents of this dissertation include:(1)This dissertation introduces the research background and significance of the subject,and briefly analyzes the development history of SLAM and the current research status of visual SLAM.Next,this dissertation introduces the classic visual SLAM framework,camera position estimation process,and map construction.(2)We analyze the influence of moving objects on the camera pose estimation,combine semantic information to design a visual odometry method based on object position relationship graph and a visual odometry method based on semantic feature points.In the visual odometry method based on object location relationship graph,the Mask R-CNN network is used to segment the original image at the instance level,and then the object location relationship graph is calculated based on the depth image and semantic information,and finally the motion states of objects are determined by combining with the prior knowledge database.Finally,the feature points extracted on the moving object are excluded.In the visual odometry method based on semantic feature points,Mask R-CNN is also used to semantically segment the original image.In the feature points extraction stage,semantic information and sparse optical flow are first used to detect moving objects,and finally the feature points extracted on the moving objects are excluded.In the feature point matching stage,a semantic feature point matching algorithm based on semantic information is proposed.The visual odometry method proposed in this dissertation is compared with ORB-SLAM3 and other dynamic scene SLAM systems.Experimental results show that our method greatly improves the positioning accuracy of the camera and the robustness of the system in a highly complex dynamic environment.(3)Finally,the key frames output of the visual odometry method above are used to construct the semantic map by combining depth images and semantic information.Aiming at the interference caused by moving objects in the dynamic environment,we designed a moving object removal algorithm.In addition,in order to improve the accuracy of point cloud construction,we designed a depth image repair algorithm based on fast guide filter.Experimental results show that the semantic map construction algorithm in this dissertation has good robustness and accuracy.