SLAM Based Multi-session Map Fusion And Dense Reconstruction

Map construction of large-scale environment is one of the important tasks of Map Surveying and Mapping,providing essential data foudations for autonomous driving,localization ang navigation and so on.However,Map Surveying and Mapping requires a variety of high-precision sensors and resources,which leads to low efficiency.In recent years,map construction based on SLAM(Simultaneous Localization and Mapping)system has drawn the widespread attention of scholars.However,due to the hardware foundation and high computational complexity,it's a great challenge for map construction of large-scale environment.Therefore,in this paper,we studies some of the disadvantages of the existing SLAM system.On the premise of ensuring the accuracy of the construction results,a multi-session map fusion algorithm is designed to complete the consisitent fusion of the map,and a dense reconstruction of the environment is added combined with depth information.Papers mainly include the following work:First of all,in terms of the overall system,in order to solve the data unity of the pure vision SLAM system,we add inertial measurement data.And a VIO(Visual-Inertial Odometry)that tightly couples visual information and inertial measurement information is implemented as the front end of the SLAM system by using Iterative Extended Kalman Filter algorithm and obtain the maximum posterior estimation of the camera pose and landmark positions.By applying triangulation method,the three-dimensional landmark points are mapped to a point cloud map of the target scene.Secondly,for the fusion of multi-session maps,we first analyze the principle of the classic ICP(Iterative Closest Point)algorithm for point cloud alignment.Then we introduce an algorithm for the overlapping searching among different sessions based on feature matching using Inverted Multi-Index.Combine descriptor projection and product quantization strategies to increase the speed for matching.Implement graph optimization to optimize the pose maps with closed-loop constraints to remove the loops among maps and fuse the repeated landmarks to obtain a globally consistent large-scale map.What's more,keyframe detection and landmark summarization strategies are added to compress the map and ensure the richness of the map information while reducing the map size.Finally,in order to meet the needs of dense scenes,the TSDF(Truncated signed distance function)model algorithm is implemented with the map information(including camera poses and positions of landmarks)and depth information measured by sensors and complet the dense reconstruction of the map to generate a dense model of the target scene.