Research On Key Technologies For Road Infrastructure Map Construction Based On Visual SLAM

Constructing road infrastructure maps is of great significance for promoting the construction of a powerful transportation nation,optimizing transportation networks,and improving road management capabilities.Traditional visual SLAM mainly focuses on real-time positioning and map construction,generating sparse point clouds,while the construction of road infrastructure maps requires dense point cloud data.Therefore,by improving the ORB-SLAM2 algorithm,a road infrastructure map construction system is designed to help improve road infrastructure management efficiency and promote the development of intelligent transportation systems.Academic dissertation implements a road infrastructure map construction system based on an improved ORB-SLAM2 framework,models the road environment,and outputs dense point clouds.The dense point clouds are then used to recognize facilities,ultimately constructing road infrastructure maps.Firstly,in the data acquisition part,a data acquisition system consisting of a stereo camera and a visual controller is designed and built.Next,in the algorithm input part,stereo disparity maps are constructed using the SGBM algorithm with a stereo camera,further generating depth maps.Then,feature extraction is improved by adopting the ORB feature extraction algorithm combined with the Q-tree uniform algorithm to reduce redundant feature points.Furthermore,keyframes are selected through the tracking thread,and pose estimation and mapping are carried out through local mapping and loop closure detection.Based on this,a dense point cloud map construction thread is added,with the left camera image and depth map of the keyframes used for mapping as input data.Pixel coordinates are converted to threedimensional spatial coordinates,resulting in the dense point cloud of keyframes.The dense point clouds of keyframes are stitched together and subjected to three-layer filtering,outputting a dense point cloud map.Finally,facilities in the dense point cloud map are classified using the Point Net++ network.Data is collected in various road scenarios such as campus roads,expressways,and ordinary urban roads,and the proposed road infrastructure map construction method is experimentally validated.In the campus road environment,the facility recognition accuracy reaches 79.17%,and the mapping average error is 9.35%.In the expressway scenario,the facility recognition accuracy reaches 73.02%,and the mapping average error is 11.61%.In ordinary urban roads,the facility recognition accuracy reaches 72.03%,and the mapping average error is 13.24%.The accuracy of various indicators in the experiments meets the requirements.The research results of this paper provide a reference for the construction of road infrastructure maps to a certain extent.