Mapping And Localization Based On Lightweight Semantic Maps In Autonomous Driving Scenarios

It is a challenging problem to achieve high-precision mapping and robust localization using low-cost sensors for autonomous vehicles.Li DAR is considered to be the most reliable sensor in autonomous driving.However,its high cost is a limitation on its extensive usage.Localization methods based on traditional vision will fail due to factors such as illumination,weather,viewing angle,and appearance changes in spite of it low cost.In view of the above problems,we propose a mapping and localization algorithm based on lightweight semantic maps in autonomous driving scenarios.Semantic features are widely appeared on urban roads and robust to illumination,weather,viewing and appearance changes.Moreover,the size of the map is greatly reduced.First,a lightweight semantic mapping algorithm based on image perception and point cloud is proposed.Then,a lightweight semantic mapping algorithm using camera is proposed since not all scenes are equipped with lidar sensors and incremental mapping is uaually required during the mapping process.Finally,an algorithm for localization using realtime images and prior maps is proposed to perform a robust localization.In view of the above problems,this thesis mainly completes the following work:1.A lightweight semantic mapping algorithm based on point cloud and image perception is proposed for traffic signs and pole objects.This thesis perform European clustering to filter out the point cloud in the background area.For traffic signs and pole objects,fitting plane and line models are carried out to select the target and compute the 3D position of the target.2.An algorithm based on vision to establish a lightweight semantic map for traffic signs and pole objects is proposed.For the given perception information,tracking between frames is first performed.If the association can be established,the established map point is used as the initial value of the coordinates.On the contrary,tracking between frames is performed and coordinates are initialized for different semantic information.Finally,this thesis added three types of factors based on semantic features,namely point observation factor,coplanar prior factor,and collinear prior factor.This thesis performed sliding window optimization and global optimization to obtain the final map.3.A localization algorithm based on lightweight semantic map is proposed.This thesis first completes the calibration of vision and IMU.A GPS-VIO algorithm is proposed to estimate the pose in the ENU coordinate system.The residual of the semantic feature is added to the system for optimization to obtain a robust localization result after finishing the data association.