Research On High-precision Mapping And Real-time Localization Methods In Unmanned Driving

High-precision mapping and localization technology is the key technology for unmanned ground vehicles(UGVs)to realize autonomous driving.Based on the autonomous navigation of UGVs,this thesis conducts research on mapping and localization in structured and unstructured environments respectively.In structured scenarios such as parks,the research results of this thesis show that the use of the low-cost two-dimensional LiDAR can meet the needs of mapping and localization.For more complex scenes,such as unstructured off-road scenes,three-dimensional mapping and positioning methods are required.In this thesis,the three-dimensional high-precision mapping and localization system are decoupled and researched separately.In the three-dimensional high-precision mapping system,for a large-scale complex environment,a multi-sensor fusion method based on factor degradation analysis is adopted to realize fully automatic high-precision map construction.In the localization system,for the problems of sparse features in offroad scenes,repeated features in highway scenes,and dynamic noise in crowded traffic scenes,a method of fusion of local motion constraints and point cloud matching observation constraints is used to solve the problem of positioning degradation in complex scenes.Finally,to achieve practical applications,this thesis proposes a robust positioning method based on multi-source information fusion,which realizes unmanned vehicle positioning that does not rely on satellite signals in mixed known and unknown scenarios.In summary,the innovations of this thesis are as follows:1.A two-dimensional simultaneous localization and mapping method for large-scale structured scenes has been proposed.This method is based on the correlation scan matching method,and its front-end matching algorithm and back-end optimization method are both improved.At the same time,to improve the accuracy of closed-loop detection,this thesis also designs and implements an efficient closed-loop detection method based on classifiers and a false closed-loop elimination mechanism.Real vehicle tests show that this method can quickly and accurately construct two-dimensional maps of structured scenes such as campuses and parks.2.Facing the off-road unstructured scene,this thesis designs and implements a robust 3D high-precision map construction method.This method can overcome the problem of point cloud matching degradation caused by feature sparseness in off-road scenes.By directly modeling the degradation situation,the mapping method is also suitable for crowded traffic scenes with excessive dynamic interference and highway scenes with single characteristics.3.Applying the constructed three-dimensional high-precision map,this thesis proposes a robust localization method based on point cloud-map matching.Aiming at the problem of point cloud matching degradation that may occur in the localization process,an optimization method combining point cloud matching constraints and local motion constraints is proposed.It achieves accurate and robust three-dimensional positioning in unstructured off-road scenes,crowded traffic scenes,and highway scenes without satellite navigation signals,and the lateral localization accuracy reaches centimeter level.4.For real-vehicle applications,a combined localization system based on multisource information fusion has been designed and implemented.The system can realize the effective integration and seamless switching of different localization methods in various states such as with/without high-precision maps,with/without satellite navigation signals.A large number of actual vehicle tests show that the system can meet the localization needs of unmanned vehicles under different conditions.