Research On Clustermap Building And Unmanned Vehicle Re-localization In Urban Environment

Map and map-based re-localization techniques are very important for unmanned vehicles driving in urban environments.So in order to make the process of map creation more simple and the achievement of re-localization more reliable,in this paper,we proposed a method for creating the ClusterMap and an approach for achieving re-localization based on it.The ClusterMap is generated by segmenting point clouds into different point clusters and filtering out clusters that belong to dynamic objects.Besides,each cluster in this map will be accompanied by a location descriptor to distinguish its location feature.By matching and verifying location descriptors,correspondences between the temporary ClusterMap and the original one can be established,then the transformation matrix between them can be recovered.This method does not require high-density point clouds and high-precision segmentation algorithms,which can save a lot of time for point cloud processing.Besides,it can overcome many changes in environmental conditions,such as light intensity,object appearance and observing direction,etc.These changes may adversely affect the accuracy and reliability of the re-localization and are difficult to overcome in methods that rely on visual features or appearance characteristics to achieve re-localization.Furthermore,in order to enhance the consistency of the ClusterMap with real scenes,we improved existing SLAM frameworks and proposed a 3D visual-lidar SLAM system.Its front-end can extract depth information for 2D visual features directly from 3D point clouds to complement the advantages of different sensors.Its back-end can detect loop-closure and use global nonlinear optimization algorithms to synchronously adjust trajectories and all map points so that the SLAM system can provide a good pose reference for the creation of ClusterMap.Algorithms proposed in this paper are tested on both KITTI dataset and our customized dataset.Experiments showed that the 3D visual-lidar SLAM can obtain stable localization and mapping results with small-scale errors,and the re-localization based on ClusterMap can also be reliably achieved in the urban environment,where the original ClusterMap was created several months ago and the temporary one is generated based on the current environmental status.