HD Map Update Based On Crowdsourcing Vehicles

With the development of AI technology and 5G communication technology,vehicle automatic driving technology has also been developed rapidly.As an important foundation for autonomous driving,HighDefinition map(HD map)provides vehicles with high-precision positioning,rich dynamic traffic information,driving decision-making and path planning capabilities,ensuring the safety and comfort of autonomous driving and improving traffic efficiency.The real-time feature determines that HD map need to be updated frequently.It is difficult for professional vehicles to complete real-time update of HD map due to high cost,and crowdsourcing vehicles have the characteristics of low cost and large quantity.Therefore,the research of HD map update based on crowdsourcing vehicles is of great significance for the real-time maintenance of HD map.HD map update procedure based on crowdsourcing vehicles includes vehicle selection,data collection,map change detection and update,map upload and distribution,etc..Vehicle selection and map change detection and update will directly affect the high accuracy and real-time.For vehicle selection method,the existing research mainly focuses on high region coverage,low road overlap rate and low crowdsourcing cost as the main optimization goals,but ignores the impact of data quality on subsequent HD map detection and update.For the map change detection and update method,the existing research has different solutions based on different data type,but they all have similar solutions,include object detection,highprecision positioning,road feature status update,etc..However,the impact of road feature occlusion on object detection and map change update is barely considered in current research.Therefore,this paper conducts research from two aspects:crowdsourcing vehicle selection and map change detection and update,and the main research contents include:(1)A crowdsourcing vehicle selection method for data quality differences is proposed to achieve the goal of selecting as few suitable vehicles as possible from vehicles with different data quality,completing full road coverage and providing high-quality data.This method maps the vehicle selection problem as a multi-armed slot machine problem without predicting the vehicle trajectory,and proposes an upper confidence bound(UCB)algorithm to solve it.At the same time,this paper proposes an method based active learning to solve the cold start problem encountered in vehicle selection.Comparative experiments show that the crowdsourcing vehicle selection method proposed in this paper can select a suitable subset of vehicles from many crowdsourcing vehicles under the premise limited budget of crowdsourcing platform,and has performance advantages compared with existing vehicle selection methods in terms of platform utility,road coverage,data quality,etc.,and also has faster cold start speed.(2)A map change detection and update method for road feature occlusion problem is proposed to achieve the goal of high robustness of target detection and high accuracy of map change update.This method considers the influence of road occlusion problem on the accuracy of object detection,and proposes a object detection algorithm based on contrastive learning,which randomly masks the occluded road features and inputs them to the contrastive learning network,so that the network can focus more on the features of the unoccluded part of the road and improve the robustness of recognition.Combining the traditional HD map update process and the research on map change update based on deep learning,this paper proposes a new map change update method.Comparative experiments show that the map change detection update method proposed in this paper has higher robustness than the traditional object detection algorithm,and at the same time,it maintains a high accuracy rate of change update and simplifies the map change update process.