Research On High Precision Map Optimization And Localization For Autonomous Driving

Localization is one of the core issues in the field of autonomous vehicle research,and it is also one of the most fundamental parts in the three-layer of robotic system.Stable and high-frequency localization with high precision is the guarantee for safe driving of an autonomous car.The localization error of an autonomous car in public roads should be controlled within 20 cm.Localization methods are divided into two categories: method based on Global Navigation Satellite System(GNSS)and method based on high precision(HD)maps.Localization method based on GNSS is widely used in the field of autonomous car.It has features like all-weather,no communication and movable positioning.Their accuracy can vary from a few of meters to a few centimeters,depending on the quality of the equipment used.For the autonomous vehicle application,the localization accuracy can be further improved by combining GNSS with real-time differential positioning technology and inertial navigation system to achieve the centimeter-level accuracy.However,GNSS-based methods are susceptible to environmental topography.The environments such as natural canyons,tunnels and urban roads with high-rise buildings,will disturb the propagation of satellite signals,and resulting in reduced accuracy.This makes the GNSS based method unable to promote the application in the field of autonomous driving.Method based on HD map is the mainstream in the field of autonomous driving.The HD map can not only provide real-time,high-frequency and centimeter-level localization,but also contains a lot of road annotation information to help the perception.In this paper,a non-map element filtering method based on full convolutional network(FCN)is proposed.Based on this method,a HD map optimization method is proposed,which is called FCN-NDT.The stability of localization based on the optimized map and ordinary map is studied.We also compared the proposed method with Realtime Kinematic and Inertial Navigation System(RTK-INS)based method in urban road driving.The main research contents of this paper include:Firstly,a point cloud non-map element semantic segmentation method based on FCN is proposed.This method can segment and classify non-map elements end-to-end.Experimental results show that the proposed method is effective.The detection speed is much faster than VoxelNet.Secondly,the FCN-NDT HD map optimization method is proposed.This method uses FCN to extract map element points and non-map element points,and combines NDT matching method and RTK-INS to add map element points to the map.The proposed method can filter out non-map element(such as pedestrians,vehicles,trucks,etc.)in the point cloud during the construction of the HD map,which makes the road information more clear for the road marker recognition.Thirdly,the influence of optimized map on the stability of NDT matching algorithm is studied.The comparison experiments show that the optimized map is better than the original map in the stability of localization.The accuracy and speed of proposed method has reached the demand of urban autonomous driving.Fourthly,a comparative experiment of NDT method with optimized map and RTK-INS method is studied in urban autonomous driving.Experimental results show that the proposed method can optimize HD maps and filter out the non-map elements effectively.Based on the optimized map,NDT matching localization method can achieve high-frequency,high-precision in urban road.Compared with RTK-INS method,the proposed method has higher stability in urban road scenes.