Research On Multi-sensor Fusion Based Point-cloud Map Creating And Updating

Autonomous driving has become a central issue in the automotive field.For automatic driving vehicles,autonomous driving map has become an indispensable requirement.Autonomous driving map can be mainly divided into two types,point cloud map and vector map.Point cloud map contains abundant environmental prior information,which supports intelligent vehicle multi functions such as positioning and obstacle detection.Consequently,how to construct and update point cloud map has became a main part in autonomous driving map development.In order to construct the automatic driving point cloud map,it is necessary to estimate ego-pose accurately when collecting point cloud data.To meet the different requirements of automatic driving,the point cloud map should contain color and semantic information,which requires the alignment calibration between Li DAR and camera.Because of dynamic property of ICVs' driving environment,the dynamic updating algorithm of point cloud map is necessary.This work will mainly focus on the following issues:The current point cloud map construction relies on differential GNSS positioning,which makes pose estimation unstable in satellite signal occlusion area.This paper will propose a pose estimation algorithm based on Light Detection and Ranging(Li DAR),monocular vision and Inertial Measurement Unit(IMU).Firstly,a correction algorithm will be studied for eliminating Li DAR pose estimation drift to improve the accuracy of estimation result.Then,the multi-sensor pose estimation tightly-coupled optimization framework will be proposed to ensure the consistency of accurate pose estimation when Li DAR fails for a short time.In order to add visual information into point cloud map,the alignment calibration method between 2D pixels and 3D points will be studied.After calibration,the corresponding pixel information can be correlated with the point in the map.Consequently,the color information and semantics segmentation results from the image can be obtained in autonomous driving map.In consideration of the non-perfect image semantics segmentation results,we will further optimize the semantics point cloud map based on the geometric and semantic information of objects according to a voting mechanism.As for the problem of updating point cloud map,the point cloud based re-localization algorithm is necessary to achieve alignment between the real-time updating data and the original map.Then,based on octree structure and binary Bayes filter,the efficient organization of point cloud map data and modeling the uncertainty of point cloud map can be achieved.Finally,traditional ray casting method needs to be improved to update the point cloud map dynamically for ICVs driving scene.The proposed algorithm will be firstly validated on the open source data set.After that,the hardware platform of the map acquisition vehicle will be designed.The point cloud map in Tsinghua University campus will be constructed and the map accuracy needs to be verified.In the experiment,the scene of ICV re-localization and point cloud map updating without differential GNSS is designed,which verifies the validity of the proposed updating algorithm of point cloud map for autonomous driving.