Research On Data Quality For HD Map Crowdsourcing In Autonomous Driving

High-Definition Map(HD Map)plays a key role in developing autonomous driving technology by providing highly accurate road information to support vehicle driving.However,its frequently updating becomes a critical barrier.Running a professional instrumented fleet for map data acquisition can be too time-consuming and expensive to guarantee the spatiotemporal coverage of map service.Conversely,the crowdsourcing approach collects and screens massive vehicular data in the edge,contributing to realizing a higher-quality map service with a wider coverage area,higher real-time performance,and lower expense.In this pattern,autonomous/intelligent vehicles are both service requesters and data providers,since they continuously capture and analyze images to understand the driving environment before making subsequent decisions.Their key semantic outputs(traffic signs,etc.)can be directly facilitated to judge the map changing and quantify the data quality.In order to realize data quality measurement and long-term data quality assurance in HD Map crowdsourcing,this paper concentrates on a data quality framework for semanticbased HD Map crowdsourcing and a reverse auction mechanism based on participant contribution.On the one hand,this paper fully utilizes these meaningful semantics and propose a semantic-based data quality framework of a vehicle-edge-cloud structure in HD map crowdsourcing.This paper divides the target area into several sub-areas and specify the edges to perform intra-subarea data filtering by using the semantic classification tree(SCT)and sensing data selection based on the proposed semantic utility metrics for the data quality measurement.Meanwhile,this paper assigns the cloud to macroscopically assess and control the overall crowdsourcing quality to achieve a subarea-balanced map crowdsourcing.A set of evaluation rules are designed to validate the performance of HD Map crowdsourcing.On the other hand,in order to motivate more vehicles to participate in long-term HD Map crowdsourcing and provide high-quality image and semantic data continuously,this paper proposes a reverse auction mechanism based on participant contribution for HD Map crowdsourcing.This mechanism includes a participant contribution measurement method,which dynamically quantifies participants' real contribution to crowdsourcing by utilizing vehicle semantic utility and sub-area overall crowdsourcing quality in the data quality framework;a real contribution-based participant selection algorithm(RCB),which conducts participant selection periodically,including two selection modes of priority-coverage contribution first(PCF)and semantic contribution first(PCF);and a potential contributionbased incentive mechanism(PC-VPC),which provides virtual participation credits(VPC)as compensation for failed participants who are prospective contributors,preventing them from leaving the crowdsourcing.Experiments show that the semantic-based data quality framework can filter out at least70% of invalid semantic objects at the edge,and realize the effective extraction of semantics required by HD Map.The selection method based on semantic data quality can effectively select high-quality semantic data which are meaningful for map updating.The proposed reverse auction mechanism based on participant contribution improves the quality of longterm crowdsourcing HD Map with a limited budget.The RCB-PCF algorithm can conduct more rapid,large-scale,and uniform map covering,and increase the number of roads that achieve sensing target by 50% compared with the algorithm that only maximizes the coverage without considering the priority of road.RCB-SCF algorithm is proven to be beneficial to significantly improve the overall semantic data quality of roads.PC-VPC incentive mechanism is useful for improving the fairness of participant selection and maintaining the number of participants in crowdsourcing.This work is of great significance to existing knowledge of data quality assessment and overall crowdsourcing quality controlling as well as the long-term data quality assurance in semantic-based HD Map crowdsourcing.