Research On Traffic Marker Reconstruction And Closed-loop Correction Of Camera Trajectory In Dynamic Scenes

Mobile robots are widely used in indoor storage,outdoor transportation and other services.Visual SLAM(simultaneous localization and mapping)technology is an important part of mobile robots' perception of unknown environments.This thesis mainly focus on the localization method of mobile robots in dynamic scenes,the semantic dense reconstruction of traffic markers and the closed-loop correction of the camera trajectory.The purpose is to use the semantic information to enable the robot to better deal with the localization problem in the dynamic scene,and complete the three-dimensional semantic reconstruction of the traffic markers.Finally,the closed-loop correction of the camera trajectory is used to improve the overall accuracy of the visual SLAM system.Accurate pose estimation is the premise for robots to build an accurate map.At present,most visual SLAM algorithms can perform correct pose estimation in static scenes.However,when there are moving objects in the scene,the localization accuracy of the system will have a huge error.In order to solve this problem,this paper first uses semantic images to identify potential moving objects,then design a dynamic object recognition model based on optical flow method,and integrates the model into the visual SLAM system,and then the public dataset is used to verify the effectiveness of the method.Experimental results show that this method can greatly improve the localization accuracy and robustness of visual SLAM system in dynamic scenes.It is important for visual SLAM to build reusable maps.In order to build a reusable semantic map,this paper proposes a dense reconstruction algorithm for traffic markers based on plane fitting.The paper first uses the semantic information of the image and the visual SLAM system to generate a semi-dense semantic map,and uses the Hungarian algorithm to perform multi-target tracking on the semantic images,and then the semidense semantic map and multi-target tracking results are used to densely reconstruct specific traffic markers.The algorithm can construct a semantic map containing seven traffic markers,which can provide a priori information for subsequent robot navigation,closed-loop correction and other functionsFinally,because visual SLAM system based on direct sparse method do not extract feature points during the tracking process,they cannot perform closed-loop correction.This paper makes use of the semantic map to make up for this defect.The closed-loop candidate frame is preliminarily judged by the distance threshold,and then the semantic map is converted into a graph model to accelerate the selection of the closed-loop candidate frame.Finally,the closed-loop correction is completed by the ICP algorithm and the pose graph.This algorithm can help the SLAM system to eliminate the cumulative error caused by long-term operation,thus improving the overall accuracy of the system.In summary,this paper improves the overall performance of the traditional SLAM system.The recognition of moving objects and trajectory closed-loop correction algorithm makes the system positioning more accurate.The traffic marker reconstruction algorithm helps the robot to construct a reusable map,which provides a new idea for the semantic map reconstruction of outdoor road scenes.