High-precision Mapping And Localization Of Point Cloud For Unmanned Driving

High-precision maps are a hot research topic in the field of unmanned vehicles in recent years.It can improve the stability of unmanned vehicles in various environments and ensure the capacity of unmanned vehicles in complex and variable traffic flows.High-precision maps have two main functions for unmanned vehicles: storing a large amount of a priori information to provide navigation for unmanned vehicles,such as lane geometry,lane line attributes,intersection topologies,etc.;using sparse features or dense point cloud extracting from sense data models the geometry of the environment and matches real-time sensory data to obtain the current position and attitude of the unmanned vehicle.There is some common equipment of unmanned vehicles.The integrated navigation system can directly obtain pose's change.But it is often affected by GPS signal occlusion,so it can't meet the positioning requirements of unmanned vehicles in tunnels,viaducts,and community environments.Camera's price is cheaper and it can obtain rich color information.Camera can restore 3D information through polar geometry,but its sensitivity to light and low measurement accuracy make it mostly used in indoor environments.In the other hands,Lidar has high measurement accuracy.The imaging characteristics that are not affected by the environment and can directly acquire geometric information have made Lidar become the main source of high-precision maps.In order to meet the positioning requirements of unmanned vehicles and improve the traffic capacity of unmanned vehicles,this paper proposes a high-precision map construction and positioning method based on point cloud.The main contents are as follows:1)Design a new map structure based on key frame of point cloud.The point cloud registration is used to construct the pose graph among key frames,which make full use of the repeated observation between the point cloud frames.Then use the map optimization method to eliminate the error generated during the registration process.This method is compared with the existing point cloud mapping method on two datasets in the small-scale but feature-rich parks and the open roads with large scale and sparse features.The experiment result proves that this method can significantly introduce the error caused by the registration algorithm.2)It is proposed to convert the point cloud into a height map.In the height map,the corresponding image features are extracted according to the gray gradient value.The bag-of-word method is used to add the global recognition of the image features in the height map.It solves the problem of global search in point cloud maps because high level feature nearly can't be extracted directly from point cloud.Comparison with existing algorithms verifies the superiority of the method.3)With global feature detection method,the mapping scheme is extended to the combination of global optimization and local optimization.The registration error can be continuously optimized by local redundant observation,and the global loopback detection is used to eliminate the longtime accumulation error.There is an experiment between only local optimization scheme and combination scheme.4)Based on the method of mapping,a high-precision map location method based on point cloud keyframe is proposed.The global pose detection and geometric relationship verification are used to initialize the global pose.With optimization of pose graph,the error caused by the dynamic change of the environment is corrected so vehicle can obtain accurate and smooth real-time pose information.