High-precision map data management and development method and system

Abstract

The application provides a high-precision map data management and development method and system, comprising the following steps: S1, collecting data by an automatic driving vehicle and uploading the data to distributed cloud storage, loading the data to perform fusion model training; S2, analyzing and archiving the stored data to perform map labeling, and storing the labeled data to short-term distributed storage; S3, performing data link intercommunication; and S4, performing cloud data storage management and business unit data storage management on the data. The application is suitable for different business scenarios and provides quick data retrieval, standard data retrieval and batch data retrieval services for automatic driving business, and provides quick use and data storage services for algorithm models. The application comprises a high-precision map data source, algorithm model training, data storage and a management system and method for development and use, realizes data resource connection, and ensures data effectiveness, availability and reliability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automatic driving, in particular, to a high-precision map data management and development method and system. BACKGROUND

[0002] The technical iteration and commercial landing of automatic driving vehicles for environmental sanitation cannot be separated from the construction of road, communication and other infrastructures. However, at present, the infrastructure construction is only based on the needs of enterprises or different departments, resulting in a series of problems such as resource duplication, difficulty in interconnection, low utilization efficiency, high construction and system maintenance cost, difficulty in management and service, low data capacity level, and so on.

[0003] The high-precision map includes quasi-dynamic map and dynamic map. The quasi-dynamic map includes information such as traffic lights, road congestion status, road construction, road surface information, and weather. The dynamic map includes high real-time information such as surrounding traffic participants, traffic accident information, and sudden roadblocks. Figure One Generally, the information is obtained by terminal sensors such as vision, laser radar, and millimeter wave radar at the vehicle end and the roadside end. The terminal platform processes the perceived information through cleaning, classification, coding, positioning, and extraction, and uploads it to the map cloud platform or differentiates it with the terminal local basic data and static data to generate or update dynamic map data. The cloud storage does not establish an effective storage strategy. The automatic driving vehicle for environmental sanitation exists independently in the entire data network of the data monitoring and management platform as an intelligent terminal. The vehicle end serves as both the collection end of high-precision map data and the application end of high-precision map. The dynamic high-precision map data, processed data, and result data are stored uniformly without effective differentiation, resulting in difficult resource management and space waste. Therefore, there is currently a lack of a system that can effectively store and manage high-precision map static and dynamic data.

[0004] Patent document CN114238538A discloses a data storage system and method based on high-precision map. The system includes a message buffer area, a real-time storage area, a short-term storage area, and a long-term storage area. The short-term storage area includes an HBase storage area for storing high-precision dynamic map data, a PostgreSQL spatial library for storing high-precision static map vector data, and a FastDFS storage area for storing file data and spatial data. The long-term storage area is used to store expired data in the short-term storage area in a file manner, establish an index directory, and store historical data and archived data. However, the data resource connectivity of this invention is poor, and it is limited to data classification storage. It cannot realize an integrated management system and method of high-precision map data source, algorithm model training, data storage, and development and use. The data resource connectivity is poor, and the effectiveness, availability, and reliability of the data need to be further improved. SUMMARY

[0005] In view of the defects in the prior art, the purpose of the present application is to provide a high-precision map data management and development method and system.

[0006] According to the high-precision map data management and development method provided by the present application, the following steps are included:

[0007] Step S1: vehicle collection data is uploaded to distributed cloud storage, and data is loaded for fusion model training;

[0008] Step S2: analyze the archived storage data to perform map labeling, and store the labeled data to short-term distributed storage;

[0009] Step S3: data link interworking is performed;

[0010] Step S4: data is stored and managed in the cloud and business unit data storage management.

[0011] Preferably, in the step S1:

[0012] Vehicle collection data is uploaded to distributed cloud storage, and positioning model and perception model load data for fusion model training, and the fusion model result set data is used as the data source for high-precision map development; dynamic data and static data of the high-precision map are classified and stored, and are respectively stored in real-time distributed message queue and archive storage;

[0013] Multi-source fusion model training is adopted, the first level is multi-sensor fusion, and the second level is multi-feature element fusion; by extracting feature points, line segments and gray information features, multiple feature elements are obtained; line features, surface features and normal distribution features of three-dimensional point clouds are obtained using laser, camera and laser are fused to obtain image and point cloud information, and the image and point cloud information is directly input into neural network and deep learning model to help extract semantic information; the geometric and semantic information is fused by adding the semantic information of the point cloud and the image.

[0014] Preferably, in the step S2:

[0015] Archive storage data preprocessing: analyze the archived storage data to perform map labeling, and store the labeled data to short-term distributed storage for static data use after passing the audit.

[0016] Preferably, in the step S3:

[0017] Vehicle collection end data is stored to the cloud, and positioning and perception models pull corresponding storage data for fusion model training, and the fusion result set is displayed and used according to the dynamic and static requirements of the high-precision map.

[0018] Preferably, in the step S4:

[0019] Cloud data storage management: vehicle collected data is uploaded to the cloud, and the cloud data provides different service data for data classification storage, and each model pulls and stores data according to the requirement to the storage medium;

[0020] Business unit data storage management: the business scenarios of the high-precision map are divided into dynamic data display and static data display, the dynamic data output by the fusion model is stored in the distributed message middleware queue, and the static data is stored in the distributed short-term storage component after data analysis, data labeling and data development;

[0021] According to different business scenarios, storage is divided into file storage, archive storage and object storage.

[0022] According to the high-precision map data management and development system provided by the application, comprising:

[0023] Module M1: vehicle collected data is uploaded to the distributed cloud storage, and the data is loaded for fusion model training;

[0024] Module M2: archive storage data is analyzed for map labeling, and the labeled data is stored in the short-term distributed storage;

[0025] Module M3: data link interworking is performed;

[0026] Module M4: cloud data storage management and business unit data storage management are performed on the data.

[0027] Preferably, in the module M1:

[0028] Vehicle collected data is uploaded to the distributed cloud storage, and the positioning model and the perception model load data for fusion model training, and the fusion model result set data is used as a data source for high-precision map development; according to the dynamic data and the static data of the high-precision map, data classification storage is performed, and the data is stored in the real-time distributed message queue and the archive storage, respectively;

[0029] Multi-source fusion model training is adopted, the first level is multi-sensor fusion, and the second level is multi-feature element fusion, a plurality of feature elements are obtained by extracting feature points, line segments and gray information features; the line features, surface features and normal distribution features of the three-dimensional point cloud obtained by the laser are used to fuse the camera and the laser to obtain image and point cloud information, and the image and point cloud information is directly input into the neural network and the deep learning model to help extract semantic information; the geometric and semantic information is fused by adding the semantic information of the point cloud and the image.

[0030] Preferably, in the module M2:

[0031] Archive storage data preprocessing: parse the archive storage data for map annotation, and store the annotated data to the short-term distributed storage for static data use after passing the audit.

[0032] Preferably, in the module M3:

[0033] The vehicle collects end data to cloud storage, and a positioning and perception model pulls corresponding storage data for fusion model training. The fusion result set is displayed and used according to the dynamic and static requirements of the high-precision map.

[0034] Preferably, in the module M4:

[0035] Cloud data storage management: vehicle data is uploaded to the cloud, and the cloud data provides different business data for data classification storage. Each model pulls and stores data according to the requirements of the storage medium.

[0036] Business unit data storage management: the business scenarios of the high-precision map are divided into dynamic data display and static data display. The dynamic data output by the fusion model is stored in the distributed message middleware queue, and the static data is stored in the distributed short-term storage component after data analysis, data annotation and data development.

[0037] According to different business scenarios, storage is divided into file storage, archive storage and object storage.

[0038] Compared with the prior art, the present application has the following beneficial effects:

[0039] 1. The present application is suitable for different business scenario storage, and provides quick data retrieval, standard data retrieval and batch data retrieval services for autonomous driving business, and provides quick use and data storage services for algorithm models.

[0040] 2. The present application includes a high-precision map data source, algorithm model training, data storage and development management system and method, realizes data resource connection, and ensures data effectiveness, availability and reliability. BRIEF DESCRIPTION OF DRAWINGS

[0041] Other features, objects and advantages of the present application will become more apparent through reading the following detailed description of the non-limiting embodiments with reference to the accompanying drawings:

[0042] Figure 1 The present application is a flowchart;

[0043] Figure 2 The present application is a model training schematic diagram;

[0044] Figure 3 The present application is a high-precision map index model schematic diagram;

[0045] Figure 4The ELK distributed data retrieval schematic diagram. DETAILED DESCRIPTION

[0046] The application will be described in detail below with specific examples. The following examples will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the application. These are within the scope of the present application.

[0047] Example 1

[0048] The present disclosure relates to the field of autonomous driving, especially to the field of high-precision maps, and in particular to a high-precision map data management and development method and system

[0049] The present application provides a management system and method for high-precision map data sources, algorithm model training, data storage and use, which realizes the connection of data resources and ensures the effectiveness, availability and reliability of data, including:

[0050] (1) Data development and storage business flow management of high-precision map of automatic driving vehicle for environmental sanitation;

[0051] (2) Algorithm model combined with big data technology to provide intelligent decision optimization scheme for data retrieval, analysis, mining and other high-precision map dynamic and static scene data;

[0052] (3) High-precision map index model optimization, which can meet the general data set retrieval of multiple business scenarios.

[0053] According to the high-precision map data management and development method provided by the present application, as shown in Figures 1-4 , including:

[0054] Step S1: vehicle collected data is uploaded to distributed cloud storage, and data is loaded for fusion model training;

[0055] Specifically, in the step S1:

[0056] The vehicle collected data is uploaded to distributed cloud storage, and the positioning model and the perception model are loaded with data for fusion model training. The fusion model result set data is used as the data source for high-precision map development. According to the dynamic data and static data of the high-precision map, the data is classified and stored, and is respectively stored in the real-time distributed message queue and the archive storage;

[0057] The first level is the fusion of multiple sensors, and the second level is the fusion of multiple feature primitives. Feature points, line segments, and gray information features are extracted to obtain multiple feature primitives. Line features, surface features, and normal distribution features of three-dimensional point clouds are obtained using a laser. The camera and the laser are fused to obtain image and point cloud information. The image and point cloud information are directly input into a neural network and a deep learning model to help extract semantic information. The geometric and semantic information is fused.

[0058] Step S2: Analyze the archived storage data for map annotation, and store the annotated data to the short-term distributed storage.

[0059] Specifically, in the step S2:

[0060] Archive storage data preprocessing: Analyze the archived storage data for map annotation, and store the annotated data to the short-term distributed storage for static data use after passing the audit.

[0061] Step S3: Perform data link interworking.

[0062] Specifically, in the step S3:

[0063] The vehicle collection end data is stored to the cloud, the positioning and perception model pulls the corresponding storage data for fusion model training, and the fusion result set is displayed and used according to the dynamic and static requirements of the high-precision map.

[0064] Step S4: Perform cloud data storage management and business unit data storage management on the data.

[0065] Specifically, in the step S4:

[0066] Cloud data storage management: The vehicle collection data is uploaded to the cloud, the cloud data provides different business data for data classification storage, and each model pulls and stores data to the storage medium according to the requirements.

[0067] Business unit data storage management: The business scenarios of the high-precision map are divided into dynamic data display and static data display. The dynamic data output by the fusion model is stored in the distributed message middleware queue, and the static data is stored in the distributed short-term storage component after data analysis, data annotation, and data development.

[0068] According to different business scenarios, the storage is divided into file storage, archive storage, and object storage.

[0069] Embodiment 2:

[0070] The application further provides a high-precision map data management and development system, which can be realized by performing the process steps of the high-precision map data management and development method, that is, the high-precision map data management and development method can be understood as a preferred embodiment of the high-precision map data management and development system by those skilled in the art.

[0071] According to the high-precision map data management and development system provided by the application, as shown in the figure, comprising: Figures 1-4

[0072] Module M1: vehicle collected data uploading distributed cloud storage, loading data for fusion model training;

[0073] Specifically, in the module M1:

[0074] The vehicle collected data is uploaded to the distributed cloud storage, the positioning model and the perception model load data for fusion model training, and the fusion model result set data is used as a data source for high-precision map development; the dynamic data and the static data of the high-precision map are classified and stored, and are respectively stored in a real-time distributed message queue and an archival storage;

[0075] A multi-source fusion model training is adopted, the first level is multi-sensor fusion, and the second level is multi-feature primitive fusion; a plurality of feature primitives are obtained by extracting feature points, line segments and gray information features; line features, surface features and normal distribution features of three-dimensional point clouds are obtained by using a laser, image and point cloud information is obtained by fusing a camera and the laser, the image and the point cloud information are directly input into a neural network and a deep learning model to help extract semantic information; the geometric and semantic information is fused by adding the semantic information of the point cloud and the image.

[0076] Module M2: analyzing archival storage data for map annotation, and storing the annotated data to short-term distributed storage;

[0077] Specifically, in the module M2:

[0078] Archival storage data preprocessing: analyzing the archival storage data for map annotation, and storing the annotated data to short-term distributed storage for static data use after the annotated data are audited.

[0079] Module M3: data link interworking;

[0080] Specifically, in the module M3:

[0081] The vehicle collected data is uploaded to the cloud storage, the positioning model and the perception model pull corresponding storage data for fusion model training, and the fusion result set is displayed and used according to the dynamic and static requirements of the high-precision map.

[0082] ​Module M4: cloud data storage management and business unit data storage management of data.

[0083] Specifically, in the module M4:

[0084] Cloud data storage management: vehicle collected data is uploaded to the cloud, and the cloud data provides different business data for data classification storage, and each model pulls and stores data according to the requirements of the storage medium.

[0085] Business unit data storage management: the business scenarios of high-precision maps are divided into dynamic data display and static data display, the dynamic data output by the fusion model is stored in the distributed message middleware queue, and the static data is stored in the distributed short-term storage component after data analysis, data labeling and data development.

[0086] According to different business scenarios, the storage is divided into file storage, archive storage and object storage.

[0087] Embodiment 3:

[0088] The management system provided by the application divides the storage according to different business scenarios:

[0089] File storage: provides scalable shared file storage services for vehicle-side original data and model output data

[0090] Archive storage: provides high-availability, low-storage and long-term object storage services. Provides fast data retrieval, standard data retrieval and batch data retrieval services for autonomous driving business.

[0091] Object storage: provides console, API, SDK and tools for uploading, downloading, sharing and managing multi-format files, and provides fast use and data storage services for algorithm models.

[0092] The technical implementation of the application, as shown in Figure 1 includes:

[0093] (1) The autonomous vehicle collects data and uploads it to the distributed cloud storage, the positioning model and the perception model load data for fusion model training, and the fusion model result set data is used as the data source for high-precision map development. According to the dynamic data and static data of the high-precision map, the data is classified and stored in the real-time distributed message queue and the archive storage.

[0094] (2) Archive storage data preprocessing: analyze the archive storage data for map labeling, and store the labeled data in the short-term distributed storage for static data use after passing the data audit.

[0095] (3) Data management:

[0096] a. Data link interworking: collection end data of sanitation vehicle to cloud storage. Positioning, perception model pulls corresponding storage data for fusion model training, and fusion result set is displayed and used according to dynamic and static requirements of high-precision map;

[0097] b. Data management:

[0098] 1) Cloud data storage management: automatic driving sanitation vehicle collects data and uploads to cloud, cloud data provides different business data for data classification storage, and each model pulls and stores data according to the needs of the storage medium;

[0099] 2) Business unit data storage management: the business scenarios of high-precision map are divided into dynamic data display and static data display, and the dynamic data such as vehicle, road and cloud output by the fusion model are stored in the distributed message middleware queue. The data of vector map (static) is stored in the distributed short-term storage component after data analysis, data labeling and data development.

[0100] Among them:

[0101] (1) Model training:

[0102] As shown in Figure 2 , the patent adopts multi-source fusion model training. The first level is the fusion of multiple sensors, such as laser or GPS. The second level is the fusion of multiple feature primitives. We can extract features such as feature points, line segments and gray information to obtain multiple feature primitives. After using laser, we can also get line features, surface features and normal distribution features of three-dimensional point cloud, which are all features used by SLAM. Fusing camera and laser can get information of two channels, i.e. image and point cloud. Directly inputting image and point cloud information into neural network and deep learning model can help to extract semantic information. Then adding the semantic information of point cloud and image can fuse geometry and semantics.

[0103] (2) High-precision map development index model:

[0104] As shown in Figure 3As shown, the shp vector file and the picture file are loaded on the cloud server side, the data of the specified label layer in the shp file is read, the longitude and latitude and the corresponding time sequence timestamp are obtained, the picture file name under the storage path is read and written into the postgres time sequence database, the longitude and latitude and the corresponding timestamp are obtained to the memory, then the picture name written into the postgres database is queried, the timestamp in the picture name is compared with the corresponding timestamp, and the absolute value of the difference satisfies the preset value, then the mapping is written back to the postgres time sequence database, the preset value is set according to the business, finally, the picture name is requested to the interface of the object storage server according to the picture name in the above mapping, the picture resource link corresponding to the picture name is obtained, and the postgres time sequence database is written back, thus, the longitude, latitude, picture name and picture access resource path are mapped and maintained in the postgres database.

[0105] Among them, ELK (ElasticSearch) distributed data retrieval can meet the general data set retrieval of multiple business scenarios; Spark is a big data parallel computing framework based on memory computing, which can perform batch processing; Flink is a distributed computing framework for stream processing and batch processing (such as Figure 4 As shown).

[0106] Those skilled in the art know that, in addition to implementing the system, device and each module thereof provided by the present application in a pure computer readable program code manner, the same program can also be realized in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers by logically programming the method steps. Therefore, the system, device and each module thereof provided by the present application can be considered as a hardware component, and the modules included therein for realizing various programs can also be considered as structures within the hardware component; the modules for realizing various functions can also be considered as both software programs for realizing methods and structures within the hardware component.

[0107] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essential content of the present application. In the case of no conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A high-precision map data management and development method, characterized by, Comprise: Step S1: vehicle acquisition data upload distributed cloud storage, load data for fusion model training; Step S2: parse the archived storage data for map annotation, and store the annotated data to short-term distributed storage; Step S3: data link interworking; Step S4: data storage management and business unit data storage management on cloud; In the step S1: Vehicle acquisition data upload distributed cloud storage, positioning model and perception model load data for fusion model training, fusion model result set data as the data source of high-precision map development; according to the dynamic data and static data of high-precision map, data classification storage is carried out, and real-time distributed message queue and archive storage are respectively stored; Multi-source fusion model training, the first level is multi-sensor fusion, the second level is multi-feature element fusion, through extracting feature points, line segments and gray information features, multiple feature elements are obtained; using laser to obtain three-dimensional point cloud line features, surface features and normal distribution features, camera and laser are fused to obtain image and point cloud information, which is directly input into neural network and deep learning model to help extract semantic information; plus point cloud and image semantic information, geometry and semantic information are fused; In the step S2: Archive storage data preprocessing: parse the archived storage data for map annotation, and store the annotated data to short-term distributed storage for static data use after passing the audit; In the step S3: Vehicle acquisition end data to cloud storage, positioning and perception model pull corresponding storage data for fusion model training, and the fusion result set is displayed and used according to the dynamic and static requirements of high-precision map; In the step S4: Cloud data storage management: vehicle acquisition data upload cloud, cloud data provides different business data for data classification storage, and each model pulls and stores data according to the requirement to the storage medium; Business unit data storage management: the business scenarios of high-precision map are divided into dynamic data display and static data display, the dynamic data output by the fusion model is stored in the distributed message middleware queue, and the static data is stored in the distributed short-term storage component after data analysis, data annotation and data development; According to different business scenarios, the storage is divided, including: file storage, archive storage and object storage.

2. A high-precision map data management and development system characterized by comprising: Comprise: Module M1: vehicle acquisition data upload distributed cloud storage, load data for fusion model training; Module M2: parse the archived storage data for map annotation, and store the annotated data to short-term distributed storage; Module M3: data link interworking; Module M4: data storage management and business unit data storage management on cloud; In the module M1: Vehicle acquisition data upload distributed cloud storage, positioning model and perception model load data for fusion model training, fusion model result set data as the data source of high-precision map development; according to the dynamic data and static data of high-precision map, data classification storage is carried out, and real-time distributed message queue and archive storage are respectively stored; Adopting multi-source fusion model training, the first level is the fusion of multiple sensors, and the second level is the fusion of multiple feature primitives. Multiple feature primitives are obtained by extracting feature points, line segments and gray information features. Line features, surface features and normal distribution features of three-dimensional point clouds are obtained using a laser. Camera and laser are fused to obtain image and point cloud information. Image and point cloud information are directly input into a neural network and a deep learning model to help extract semantic information. Geometric and semantic information are fused by adding point cloud and image semantic information. In the module M2: Archive storage data preprocessing: parse the archive storage data for map annotation, and store the annotated data to the short-term distributed storage for static data use after passing the audit; In the module M3: Vehicle acquisition end data to cloud storage, positioning and perception model pull corresponding storage data for fusion model training, and the fusion result set is displayed and used according to the dynamic and static requirements of the high-precision map; In the module M4: Cloud data storage management: vehicle acquisition data is uploaded to the cloud, and the cloud data provides different business data for data classification storage. Each model pulls and stores data according to the requirements of the storage medium. Business unit data storage management: the business scenarios of high-precision maps are divided into dynamic data display and static data display. Dynamic data output by the fusion model is stored in the distributed message middleware queue, and static data is stored in the distributed short-term storage component after data analysis, data annotation and data development. According to different business scenarios, storage is divided into file storage, archive storage and object storage.

Images