High Precision 3D Mapping And Localization Based On Lidar

Recently one has witnessed a rapid progress for autonomous driving,and several companies have released their prototype autonomous driving cars.This rapid progress on one hand relies on the progress from traditional pattern recognition field,whilst on the other hand benefits a lot from the utilization of high precision maps.High precision maps differ from the traditional ones in that they contain much more accurate information which could aid the Autonomous Land Vehicle(ALV)in at least three aspects.Firstly,through a sophisticated scan-to-map matching,the ALV could be accurately localized on the high precision map without GPS signals.Secondly,by manually annotating the objects of interest on the map,an accurate prior for the static part of the surrounding scene could be perceived by the ALV.These objects of interest include lane markings,traffic lights,curbs,etc.Thirdly,the shape and region of the road obtained from the map could aid the detection of other dynamic traffic participants,and it could also be helpful for path planning.The issue researched in this paper is about the high precision three dimensional mapping and localization.The main contents and innovations are illustrated as follows:1.This paper proposes an adaptive Lidar scan matching algorithm.Lidar matching is the basis of mapping and localization.As the current popular matching algorithms only focus on urban environment,this paper proposes a combination matcher which could deal with both urban and rural data,improving the matching method's adaptability to the environment.This matcher naturally combines the merits of the local gradient method and the global searching method.Experiments on real-world dataset demonstrate that the accuracy of our approach exceeds many popular algorithms.2.This paper proposes a Lidar compensation algorithm which maintains the environmental consistency.Lidar point cloud is distorted during collecting data as the platform assembled with Lidar have a motion in this period.However,accurate scan matching relies on suitable Lidar compensation,which is ignored by a majority of matching algorithms.This paper proposes a compensation method based on matching and local optimization network by re-projecting multi-local point cloud frames to the same timestamp to obtain the best registration which means the best compensation.Compared with the Inertial Navigation System(INS),a costly way acquiring compensation,our method is much more accurate and would not be affected by the GPS bias.In contrast to the traditional matching compensation,our algorithm is not affected by the dropped frames and maintains the consistency of the environment.3.This paper proposes a localization algorithm tightly fusing the Inertial Measurement Unit(IMU)and Lidar based on the prior map.When the environmental structure is too simple or cluttered,existing methods which are only based on Lidar matching tend to be divergent.To make the localization more robust,our method predicts motion by pre-integrating IMU on manifold and optimizes the pose with observation obtained by matching the Lidar scan to the prior map.Moreover,the observation helps to modify the IMU bias which would come to a better prediction in turn.And,this paper adopts a multiresolution map representation based on the octo-tree.By setting different resolutions to the ground and obstacle in the map,the localization could be accelerated.Experiments on the real-world dataset show our method is accurate with translation error around 5 cm and rotation error around 0.01 degree.