Crowdsourcing data based automatic driving high-precision map incremental updating method and system

Abstract

The application provides an automatic driving high-precision map incremental updating method and system based on crowdsourcing data, relates to the technical field of automatic driving, and obtains image and point cloud data of a temporary construction area of a city road uploaded by a plurality of crowdsourcing vehicles, wherein the image and point cloud data jointly constitute crowdsourcing data; the crowdsourcing data is subjected to key frame screening, spatiotemporal correlation analysis is performed on the screened key frame data, and a region feature data package with a time stamp is generated; based on a deep learning model, semantic segmentation is performed on the region feature data package to identify a map change boundary, and map change boundaries respectively corresponding to the plurality of crowdsourcing vehicles are processed; a dynamic weight fusion algorithm is used to calculate the processed results, an incremental updating patch is generated, and the automatic driving high-precision map is subjected to incremental updating based on the incremental updating patch, so that the updating precision of the automatic driving high-precision map for a temporary construction scene of a city road can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, and particularly relates to an automatic driving high-precision map incremental updating method and system based on crowdsourcing data. BACKGROUND

[0002] With the evolution of automatic driving technology to a higher level, high-precision maps, as the core foundation supporting perception decision-making and path planning, need to adapt to the dynamic changes of urban roads in real time. Frequent temporary construction, emergency repair, temporary adjustment of road markings and other scenarios in urban roads require high-precision maps to adopt an incremental updating mode to modify only the changed areas, rather than redraw the entire map, so as to reduce resource consumption and time cost in the updating process, while ensuring updating efficiency. The current technical requirements focus on: how to rely on crowdsourcing data to achieve efficient incremental updating of high-precision maps, both to ensure real-time capture of dynamic scenario changes such as temporary construction, and to ensure updating accuracy through precise fusion of multi-source crowdsourcing data, while filling the road network coverage gaps that cannot be reached by traditional collection methods.

[0003] At present, for the above technical requirements, the more typical existing scheme in the industry is a high-precision map updating scheme based on fixed sensor nodes and uniform weight fusion. The core logic of this scheme is: fixed laser radars and high-definition cameras are deployed at key nodes of urban roads (such as crossroads and main road sections), and road environment data is collected through a preset period; the collected data is transmitted to the cloud processing center through the network, and a uniform weight fusion algorithm is used to integrate the collection data of multiple nodes to eliminate data redundancy and repeated information; then the fused data set is compared with the original high-precision map to identify the road change area and generate the corresponding update patch, and finally the update patch is pushed to the automatic driving vehicle to complete the updating of the high-precision map.

[0004] However, the existing solutions mentioned above have significant drawbacks and are difficult to adapt to the incremental update requirements of high-precision maps in crowdsourced data scenarios: First, the coverage is inherently limited. Fixed sensors can only be deployed at key nodes with high traffic flow, lacking coverage for end-point road networks such as suburban side roads and community alleys. This results in temporary construction, road adjustments, and other changes in these areas not being captured, creating blind spots in high-precision map updates. Second, the data fusion accuracy is insufficient. The unified weight fusion algorithm does not consider the differences in data quality between different fixed sensors. For example, some sensors may experience decreased acquisition accuracy due to equipment aging, or data distortion due to severe weather such as rain or fog. Under unified weights, these low-quality data are included in the fusion process equally with high-quality data, which can easily lead to deviations in the identification of road change areas, thus affecting the update accuracy of high-precision maps. Third, the update response is lagging. The acquisition cycle of fixed sensors is fixed, and data transmission and fusion processing require additional time. For sudden and potentially rapidly changing scenarios such as temporary construction, it is impossible to capture changes in real time, easily leading to situations where "construction has started but the high-precision map has not been updated in time," posing a threat to the traffic safety of autonomous vehicles. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for incremental updating of high-precision maps for autonomous driving based on crowdsourced data, so as to solve the problems of limited coverage, low fusion accuracy, and slow response in the existing technology that affect the accuracy of updating high-precision maps for temporary construction scenarios on urban roads.

[0006] To address the aforementioned technical problems, in a first aspect, this application provides a method for incremental updating of high-precision maps for autonomous driving based on crowdsourced data, comprising:

[0007] The images and point cloud data of temporary construction areas on urban roads uploaded by multiple crowdsourced vehicles are obtained, and the images and point cloud data together constitute crowdsourced data.

[0008] The crowdsourced data is filtered for keyframes, and the filtered keyframe data is subjected to spatiotemporal correlation analysis to generate a time-stamped regional feature data package.

[0009] Based on a deep learning model, semantic segmentation is performed on the regional feature data packets to identify map change boundaries, and the map change boundaries corresponding to the multiple crowdsourced vehicles are processed respectively.

[0010] The processed results are calculated using a dynamic weight fusion algorithm to generate an incremental update patch, which is then used to incrementally update the high-precision map for autonomous driving.

[0011] Optionally, the key frame screening of the crowdsourcing data comprises:

[0012] extracting image frame data and point cloud frame data in the crowdsourcing data, and confirming that the image frame data and the point cloud frame data correspond to the same collection time;

[0013] marking a region containing a construction identifier in the image frame data, dividing the point cloud frame data into point cloud clusters matching the shape of the construction identifier, and regarding the image frame data and the point cloud frame data as candidate key frame data if the region and the point cloud cluster have a corresponding relationship;

[0014] arranging the candidate key frame data in a collection time sequence, and selecting candidate key frame data with a time interval not exceeding a preset time length as final key frame data;

[0015] based on collection location information of the final key frame data, distributing the final key frame data to corresponding target road segments according to city road segment division rules to form a continuous key frame data sequence;

[0016] annotating and integrating the collection time of each frame of data in the continuous key frame data sequence to generate a time-stamped regional feature data package.

[0017] Optionally, based on the collection location information of the final key frame data, distributing the final key frame data to corresponding target road segments according to city road segment division rules to form a continuous key frame data sequence comprises:

[0018] obtaining the location information of fixed identifiers arranged on both sides of city roads, and taking the road interval between two adjacent fixed identifiers as the division range of the target road segment;

[0019] comparing the collection location information of the final key frame data with the division range of each target road segment, and when the comparison result indicates that the location information of the final key frame data is within the division range of any target road segment, distributing the final key frame data to the corresponding target road segment;

[0020] arranging the final key frame data distributed to the same target road segment in a collection time sequence, calculating the distance between the locations of two adjacent final key frame data after sorting, retaining adjacent final key frame data with a distance less than a preset value, and eliminating the next final key frame data with a distance greater than or equal to the preset value to form a continuous key frame data sequence.

[0021] In a second aspect, the present application provides an automatic driving high-precision map incremental updating system based on crowdsourcing data, comprising:

[0022] an acquisition module configured to acquire images and point cloud data of temporary construction areas on urban roads uploaded by a plurality of crowd-sourcing vehicles, the images and the point cloud data collectively constituting crowd-sourced data;

[0023] a screening module configured to perform key frame screening on the crowd-sourced data, and perform spatio-temporal correlation analysis on screened key frame data to generate a time-stamped area feature data package;

[0024] a processing module configured to perform semantic segmentation on the area feature data package based on a deep learning model to identify a map change boundary, and process the map change boundary corresponding to each of the plurality of crowd-sourcing vehicles;

[0025] a generation module configured to calculate a processed result according to a dynamic weight fusion algorithm to generate an incremental update patch, and perform incremental update on an automatic driving high-definition map based on the incremental update patch.

[0026] In a third aspect, the present application provides an electronic device, comprising:

[0027] a memory configured to store a computer program;

[0028] a processor configured to implement steps of the method for incremental update of an automatic driving high-definition map based on crowd-sourced data according to the first aspect when executing the computer program.

[0029] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executable by a processor to implement steps of the method for incremental update of an automatic driving high-definition map based on crowd-sourced data according to the first aspect.

[0030] In the present application, a method for incremental update of an automatic driving high-definition map based on crowd-sourced data is provided, comprising: acquiring images and point cloud data of temporary construction areas on urban roads uploaded by a plurality of crowd-sourcing vehicles, the images and the point cloud data collectively constituting crowd-sourced data; performing key frame screening on the crowd-sourced data, and performing spatio-temporal correlation analysis on screened key frame data to generate a time-stamped area feature data package; performing semantic segmentation on the area feature data package based on a deep learning model to identify a map change boundary, and processing the map change boundary corresponding to each of the plurality of crowd-sourcing vehicles; calculating a processed result according to a dynamic weight fusion algorithm to generate an incremental update patch, and performing incremental update on an automatic driving high-definition map based on the incremental update patch.

[0031] The application has the following advantages: by acquiring image and point cloud data of the temporary construction area of the urban road uploaded by multiple crowd-sourcing vehicles and forming crowd-sourcing data, the construction area data with wide coverage and high collection frequency can be acquired, the limitation that the fixed sensor only covers key nodes is made up, and rich and comprehensive data source is provided for high-precision map updating; by performing key frame screening on the crowd-sourcing data, and performing spatio-temporal correlation analysis on the screened key frame data to generate a region feature data packet with a time stamp, redundant invalid frames in the crowd-sourcing data can be eliminated, and effective key frames related to the construction area are extracted, and meanwhile, structured data is formed in combination with the spatio-temporal information, thereby providing high-quality input for subsequent semantic segmentation; by performing semantic segmentation on the region feature data packet based on a deep learning model to identify a map change boundary, and processing the map change boundaries corresponding to the multiple crowd-sourcing vehicles respectively, the map change area boundary caused by temporary construction can be accurately identified, boundary deviation caused by single data or extensive identification mode is avoided, and the accuracy of the change boundary is ensured; by calculating the processed result according to a dynamic weight fusion algorithm to generate an incremental update patch, and performing incremental updating on the autonomous driving high-precision map based on the patch, the multi-source map change boundary data can be differentiated and fused, the update patch is generated only for the map change part, the whole map does not need to be redrawn, resource consumption for updating is reduced, and the updating efficiency and precision of the high-precision map are improved.

[0032] Further, image frame data and point cloud frame data at the same collection time in the crowd-sourcing data are extracted, a region containing a construction identifier in the image frame is marked and a point cloud cluster matching the region in the point cloud frame is divided, and frame data with a corresponding relationship is taken as a candidate key frame; the candidate key frame is arranged according to collection time and selected as a final key frame if the time interval meets the requirements; meanwhile, an interval between adjacent fixed identifiers on both sides of the urban road is taken as a target road section range, the final key frame is distributed to the corresponding target road section according to the collection position, the final key frames of the same target road section are sorted according to time, the distance between adjacent frames is calculated, and frame data with a distance meeting the requirements is reserved to form a continuous key frame data sequence, and finally the collection time of the sequence is labeled and integrated to generate a region feature data packet with a time stamp. The same collection time of the image frame and the point cloud frame and the corresponding relationship between the two can be confirmed, the data consistency and effectiveness of the candidate key frame are ensured, redundant and discrete invalid data is eliminated by selecting the final key frame with a time interval meeting the requirements, distributing to the target road section according to the position and reserving frame data with an adjacent distance meeting the requirements, a spatio-temporal continuous key frame sequence is formed, a region feature data packet is generated by labeling the time and integrating, and structured and high-quality data source is provided for subsequent semantic segmentation and map change boundary identification, and the accuracy and efficiency of subsequent processing are further improved.

[0033] These aspects or other aspects of the application will be more apparent in the following description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the accompanying drawings required by the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0035] Figure 1 A flowchart of an automatic driving high-precision map incremental updating method based on crowd-sourced data provided by an embodiment of the present application;

[0036] Figure 2 A specific implementation diagram of an automatic driving high-precision map incremental updating method based on crowd-sourced data provided by an embodiment of the present application;

[0037] Figure 3 A structural diagram of an automatic driving high-precision map incremental updating system based on crowd-sourced data provided by an embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to solve the problems of coverage limitation, low fusion accuracy and response lag in the prior art, an automatic driving high-precision map incremental updating method based on crowd-sourced data is provided by an embodiment of the present application, which adopts the following design concept: first, collect pictures and point cloud data uploaded by multiple social vehicles about temporary construction areas of urban roads, and the two types of data together constitute crowd-sourced data for updating the map; then, filter out key data frames useful for updating the map from the crowd-sourced data, and analyze the data collection time and collection location to sort out area feature data with time markers; then, process the area feature data by deep learning to find out the boundaries changed in the map due to construction, and sort out the changed boundaries collected by different crowd-sourced vehicles; finally, according to the actual situation of the data, assign different importance weights to different changed boundaries, use appropriate algorithms to fuse and calculate the data, generate update patches containing only the changed part of the map, and then use the patches to update the automatic driving high-precision map. This method can effectively solve the previous problems: relying on a large number of crowd-sourced vehicles to collect data can cover various areas from main roads to terminal roads, eliminating the update blind area; by filtering key data, analyzing spatio-temporal information and assigning weights according to data quality, the accuracy of map updating can be improved; only the changed part is used to generate patches for updating, and the frequency of data collection by crowd-sourced vehicles is high, which can quickly respond to temporary construction changes, avoid update lag, and ensure the safety of automatic driving.

[0039] In order for those skilled in the art to better understand the present application, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0040] The core of the present application is to provide an automatic driving high-precision map incremental updating method based on crowd-sourced data, and a specific embodiment of the method is shown in the flowchart Figure 1 As shown in the figure, the method comprises:

[0041] S11, acquiring image and point cloud data of the temporary construction area of the urban road uploaded by a plurality of crowd-sourced vehicles, and the image and point cloud data jointly constitute crowd-sourced data.

[0042] Among them, the crowd-sourced vehicle is a social ordinary vehicle installed with an image acquisition device (such as a camera) and a collection device capable of acquiring the actual spatial position information of an object, which can collect road-related information during driving; the image of the temporary construction area of the urban road is a picture containing construction-related elements (such as construction fences and cone barrels) taken by the image acquisition device of the crowd-sourced vehicle when passing through the temporary construction road section; the point cloud data is a set of information collected by the special spatial position collection device of the crowd-sourced vehicle, which can reflect the actual spatial position of the object (such as construction fences and cone barrels) on the road; the crowd-sourced data is a data set composed of the above image and point cloud data uploaded by a plurality of crowd-sourced vehicles.

[0043] In the embodiments of the present application, first, let the crowd-sourced vehicles collect the image and point cloud data of the area when driving to the temporary construction area of the urban road through the image acquisition device and spatial position collection device carried by themselves, then these crowd-sourced vehicles will upload the collected image and point cloud data to the cloud data processing platform through the network, finally the cloud platform receives these data uploaded by a plurality of crowd-sourced vehicles and integrates them to form the crowd-sourced data needed for subsequent processing, for example, there is a temporary construction road section in the suburbs of a city, when a certain brand of crowd-sourced vehicle passes through here, its camera takes pictures containing construction fences and cone barrels, and at the same time the spatial position collection device collects the spatial position information of these fences and cone barrels on the road, then the vehicle uploads these pictures and spatial position information to the cloud, then other multiple brands of crowd-sourced vehicles also pass through the road section and upload similar data, the cloud platform integrates these data from different brands of vehicles together to form the crowd-sourced data of the construction road section.

[0044] S12, key frame screening is performed on the crowd-sourced data, and spatio-temporal correlation analysis is performed on the screened key frame data to generate a region feature data packet with a time stamp.

[0045] Among them, the key frame screening is an operation of selecting image frames and point cloud frames containing effective construction information and having high data quality from a large amount of crowdsourcing data; the key frame data is a combination of image frames and point cloud frames obtained after key frame screening, which is helpful for subsequent map updating; the space-time correlation analysis is an operation of analyzing the correlation between the data in time and space in combination with the collection time and collection location of the key frame data; and the region feature data packet with time stamp is a data set reflecting the key features of the construction region, which is formed by labeling the collection time for each frame of data on the basis of the key frame data and then integrating the labeled key frame data according to the construction region.

[0046] In the embodiments of the present application, first, image frame data and point cloud frame data are extracted from crowdsourcing data, it is confirmed whether the image frames and point cloud frames correspond to the same collection time, then the region containing the construction identifier is marked in the image frame, the point cloud frame is divided into point cloud clusters matching the shape of the construction identifier, if the marked region and the divided point cloud cluster have a corresponding relationship, such image frame and point cloud frame are taken as candidate key frame data; then the candidate key frame data is arranged in the order of collection time, and the candidate key frame data with a time interval meeting the preset requirement is selected as the final key frame data; then the position information of the fixed identifier on both sides of the urban road is obtained, the road interval between the adjacent two fixed identifiers is taken as the division range of the target road section, the collection position information of the final key frame data is compared with the division range of each target road section, and the final key frame data with the collection position in the range of a certain target road section is distributed to the corresponding target road section; after the final key frame data distributed to the same target road section is sorted according to the collection time, the distance between the collection positions of adjacent two frames of data is calculated, the adjacent frame data with a distance meeting the preset requirement is retained, and the latter frame of data with a distance not meeting the requirement is removed, to form a continuous key frame data sequence; finally, the collection time is labeled for each frame of data in the continuous key frame data sequence and integrated, to generate the region feature data packet with time stamp. For example, from the crowdsourcing data of a construction region in a suburban area of a city, first, the image frame and point cloud frame collected at the same time are found out, the region with construction fence is marked in the image frame, the point cloud frame is divided into point cloud clusters matching the shape of the fence, and after confirming that the two correspond to each other, the candidate key frame is obtained, these candidate frames are arranged according to the collection time, the final key frame meeting the time interval requirement is selected, the interval between the adjacent fixed identifiers (such as street lamps) on both sides of the road is taken as the target road section range, the final key frame is distributed to the target road section according to the collection position, the final key frame of the same road section is sorted according to the time, the distance between the collection positions of adjacent frames is calculated, the frame with a short distance is retained, and the frame with a long distance is removed, to form a continuous key frame sequence, and finally the collection time is labeled for each frame and integrated, to generate the region feature data packet with time stamp of the construction region.

[0047] S13, performing semantic segmentation on the regional feature data packet based on the deep learning model to identify the map change boundary, and processing the map change boundaries respectively corresponding to the plurality of crowd-sourcing vehicles.

[0048] The deep learning model is a computer model that can autonomously master data characteristics by learning a large amount of data, and then process the data. The semantic segmentation is an operation of dividing the image frames and the point cloud frames in the regional feature data packet into different parts according to content categories by using the deep learning model. The map change boundary is an edge line between an original normal road region and a newly added construction region in the high-definition map for autonomous driving, that is, a boundary position on the map that changes due to construction. The processing of the map change boundaries respectively corresponding to the plurality of crowd-sourcing vehicles is an operation of organizing and standardizing the map change boundaries identified by the data collected by different crowd-sourcing vehicles, so that the format and range description of the boundaries are consistent, and finally obtaining the map change boundaries with uniform format.

[0049] In the embodiments of the present application, the regional feature data packet with a timestamp is first input into the deep learning model, the image frames in the data packet are segmented semantically by using the construction marker features learned by the model in advance, so as to distinguish the part belonging to the construction region from the part belonging to the normal road region in the image frames. Then, based on the semantic segmentation result of the image frames, the point cloud frames are also segmented semantically by using the deep learning model in combination with the corresponding point cloud frame data, so as to determine the position range of the construction region in the actual space, and then identify the map change boundary between the construction region and the normal road region. Finally, the map change boundaries respectively corresponding to the plurality of crowd-sourcing vehicles are collected, and the boundaries are processed in a uniform format, so that the boundary information formats corresponding to different crowd-sourcing vehicles are consistent. For example, the regional feature data packet with a timestamp of a construction region in a suburban area of a city is input into the deep learning model, the model divides the region in the image frames where the construction fences and the cone barrels are located into the construction region according to the learned features of the construction fences and the cone barrels, and divides other regions into the normal road region, and then determines the actual range of the construction region in combination with the spatial position information of the construction fences and the cone barrels in the point cloud frames, finds the edge outside the construction fence as the map change boundary, and then collects the map change boundaries of the construction road section identified by the plurality of brand crowd-sourcing vehicles, and unifies and standardizes the boundary information in different description modes into the same format, so that the formats of all the boundaries are consistent.

[0050] S14, calculating the processed result according to a dynamic weight fusion algorithm to generate an incremental update patch, and performing incremental update on the high-definition map for autonomous driving based on the incremental update patch.

[0051] The dynamic weight fusion algorithm is an algorithm for assigning different importance to data from different sources according to the actual situation of the data, and then integrating and calculating the data according to the assigned weights; the incremental update patch is a file containing only the part of the high-precision map for autonomous driving that needs to be updated, and does not contain the normal part of the map that has not changed; the incremental update of the high-precision map for autonomous driving based on the incremental update patch is to load the generated incremental update patch into the update system of the high-precision map, and only replace or supplement the part of the map that needs to be changed, without the need to redraw the entire map, so as to finally realize the incremental update of the high-precision map for autonomous driving.

[0052] In the embodiment of the present application, first, the map change boundary data corresponding to the plurality of processed crowd-sourced vehicles is collected, the accuracy of the collection time and the collection position of each boundary data is determined, and the corresponding weight is assigned to each boundary data through a dynamic weight fusion algorithm according to these conditions. Generally, the more recent the collection time and the smaller the position deviation of the boundary data, the higher the weight assigned, and the older the collection time and the larger the position deviation, the lower the weight assigned. Then, all the processed map change boundary data is fused and calculated according to the assigned weights, the calculation method is to multiply the boundary information corresponding to each boundary data by its weight, and then add all the products to obtain a comprehensive map change boundary. Then, the original high-precision map for autonomous driving is compared to find the part of the map corresponding to the comprehensive map change boundary that needs to be updated, and an incremental update patch containing only this part of the update content is generated. Finally, the incremental update patch is loaded into the update system of the high-precision map for autonomous driving, and the system only modifies the part of the map that needs to be updated to complete the incremental update of the high-precision map. For example, for a construction road section in a suburban area of a city, a plurality of processed boundary data of crowd-sourced vehicles of different brands is collected, the boundary data of a certain brand of vehicle has a relatively new collection time and a small position deviation, and is assigned a high weight, the collection time of the boundary data of another brand of vehicle is medium, the position deviation is medium, and the weight is medium, and the collection time of the boundary data of another brand of vehicle is old, the position deviation is large, and the weight is low. In the fusion calculation, the boundary information of each vehicle is multiplied by the corresponding weight, and the results are added to obtain a comprehensive boundary. Comparing with the original high-precision map, it is found that the original map does not have construction annotations on the road section, and an incremental update patch containing the comprehensive boundary and the construction annotations is generated. After the patch is loaded into the update system, the system adds these contents to the corresponding road section of the original map to complete the update.

[0053] The present application provides the following specific examples: a section of road in a suburban area of a city has a temporary construction area set up for maintenance, and multiple crowd-sourced vehicles of different brands pass through the construction area one after another in daily driving. Each crowd-sourced vehicle captures pictures of the construction area through the camera carried by itself, and the pictures can see the construction barriers and cones, and the equipment on the vehicle that can collect spatial position also synchronously collects the actual spatial position information of the barriers and cones on the road. Then each crowd-sourced vehicle uploads the pictures and the collected spatial position information to a designated cloud processing platform through the network. The cloud platform receives and aggregates these data to form crowd-sourced data of the temporary construction area. Image frames and point cloud frames are extracted from the crowd-sourced data in the cloud, it is first confirmed that each group of image frames and point cloud frames are collected at the same time, then the area containing the barriers and cones is marked in the image frames, the point cloud frames are divided into point cloud clusters that match the shape of these construction markers, and after confirming that the marked area corresponds to the point cloud cluster, the candidate key frame data is obtained. The candidate key frames are arranged according to the collection time, and the ones with a time interval meeting the requirements are selected as the final key frame data. Then, the interval between the adjacent street lamps (fixed markers) on both sides of the construction section is taken as the target road section range, and the final key frame data is distributed to the target road section according to the collection position. After sorting the final key frame data distributed to the road section according to the collection time, the distance between adjacent frames is calculated, and the adjacent frames with a distance meeting the requirements are retained, and the frames not meeting the requirements are removed, to form a continuous key frame data sequence. The collection time is labeled for each frame data in the sequence and integrated to generate a region feature data package with a timestamp for the construction area. The generated region feature data package is input into a deep learning model, and the model performs semantic segmentation on the image frames according to the learned features of the construction markers, divides the area where the barriers and cones are located in the image frames into a construction area, and divides the remaining area into a normal road area. Then, combined with the corresponding point cloud frame data, the model also performs semantic segmentation on the point cloud frame to determine the actual spatial range of the construction area, and further identifies the map change boundary between the construction area and the normal road area. Then, the map change boundaries of the construction section identified by multiple brand crowd-sourced vehicles are collected, and the boundary information in different descriptions is uniformly standardized to the same format, ensuring that all boundary formats are consistent. The map change boundary data of the processed multiple brand crowd-sourced vehicles is collected, the collection time and position accuracy of each boundary data are analyzed, the boundary data with newer collection time and smaller position deviation are assigned higher weights, the boundary data with medium collection time and medium position deviation are assigned medium weights, and the boundary data with older collection time and larger position deviation are assigned lower weights. According to the assigned weights, the boundary information of each vehicle is multiplied by the corresponding weight, and all results are added to obtain the comprehensive map change boundary of the construction section. Compared with the original autonomous driving high-precision map, it is found that the original map shows the section as a normal road without construction-related annotations, and therefore an incremental update patch containing only the comprehensive map change boundary and construction annotations is generated.The incremental update patch is loaded into the update system of the automatic driving high-precision map, and the system only adds the comprehensive map change boundary and the construction mark at the corresponding position of the suburban construction road section in the original map, without modifying other parts of the map, so as to finally complete the incremental update of the automatic driving high-precision map of the road section.

[0054] By performing S11-S14, through the entire incremental update process, relying on the data collected by multiple crowd-sourcing vehicles, the application embodiment can collect construction area information covering a wide range, including not only the construction area of the trunk road but also the construction area of the end road such as the suburban road, thereby making up for the limited coverage when relying only on fixed sensors to collect data and solving the problem that the construction change of the end road is difficult to be captured. In the data processing process, the key frame is screened to remove redundant invalid data, high-quality regional feature data packets are formed by combining the spatio-temporal correlation analysis, the deep learning model is used for accurate segmentation and identification of the map change boundary, and the dynamic weight fusion algorithm is used to preferentially adopt the data with higher quality to generate a comprehensive boundary, thereby effectively improving the accuracy of the map change boundary identification and avoiding the boundary deviation problem caused by the difference in data quality or extensive identification method. The generated incremental update patch only contains the map change part, and there is no need to redraw the entire map, thereby greatly reducing the data amount and computing resources required for updating and reducing the update cost. At the same time, the crowd-sourcing vehicles collect data frequently, and the data processing and updating process is efficient, so that the change of the sudden scene such as temporary construction can be quickly responded, the update lag problem when the fixed sensor collects data is avoided, the updating efficiency of the high-precision map is improved, and it is ensured that the updated map can accurately reflect the construction area situation, thereby providing protection for the safe passing of the automatic driving vehicle.

[0055] In a possible embodiment, as shown in Figure 2 S12, the crowd-sourcing data is subjected to key frame screening, and the key frame data after screening is subjected to spatio-temporal correlation analysis to generate a time-stamped regional feature data packet, including:

[0056] Step 121, the image frame data and the point cloud frame data in the crowd-sourcing data are extracted, and it is confirmed that the image frame data and the point cloud frame data correspond to the same collection time.

[0057] Wherein, the image frame data is a single picture in the continuous road pictures collected by the crowdsourcing vehicle through the camera, and each picture contains the road scene information at a specific time; the point cloud frame data is a set of data points collected by the crowdsourcing vehicle through the device for collecting spatial position (such as laser radar), which can reflect the actual spatial position of the object on the road; the same collection time refers to the time when the image frame data and the point cloud frame data are collected, which ensures that the two types of data reflect the road scene at the same time point; this step needs to extract the two types of frame data from the previously summarized crowdsourcing data, and then confirm that they are collected at the same time through time comparison, and finally form a combination of time-matched image frames and point cloud frames, which provides a basis for subsequent association processing.

[0058] In the embodiments of the present application, first, all image frame data and point cloud frame data are extracted from the already summarized crowdsourcing data through a data extraction tool, wherein each image frame data and each set of point cloud frame data is marked with a time label of the collection time, then a time comparison tool is used to check the time label of each image frame data and the time label of each set of point cloud frame data, if the time labels of the two are exactly the same, it is confirmed that they correspond to the same collection time, if the time labels are different, this set of image frame and point cloud frame is temporarily not included in the subsequent processing, for example, from the crowdsourcing data of the temporary construction area in the suburb of A place, a plurality of road pictures and a plurality of sets of spatial position data points collected by a certain brand of crowdsourcing vehicle are extracted, each picture and each set of data points are marked with the collection time, it is found that the time label of a certain picture is the same as the time label of the corresponding set of data points, it is confirmed that the two correspond to the same collection time, and the pictures and data points with different time labels are temporarily not included in the subsequent processing.

[0059] Step 122, marking the region containing the construction identifier in the image frame data, dividing the point cloud frame data into point cloud clusters matching the shape of the construction identifier, if there is a corresponding relationship between the region and the point cloud cluster, the image frame data and the point cloud frame data are taken as candidate key frame data.

[0060] Wherein, the region of the construction identifier is a specific picture part in the image frame data containing the construction-related object (such as construction fence, cone barrel); the point cloud cluster is a set of data points formed by dividing the point cloud frame data according to the shape of the object, which is similar to the shape of the construction identifier; the corresponding relationship refers to the fact that the construction identifier region marked in the image matches the point cloud cluster divided in the point cloud in terms of spatial position, that is, both reflect the same construction identifier; the candidate key frame data is the combination of the image frame data and the point cloud frame data that meet the condition of "containing the construction identifier region and corresponding matching point cloud cluster", which is the basic material for subsequent screening of the final key frame data.

[0061] In the embodiments of the present application, first, the image recognition tool is used to identify construction markers such as construction fences, cone barrels, etc. in the image frame data confirmed at the same collection time, and mark the screen area where these markers are located on the image. Then, the point cloud division tool is used to divide the point cloud frame data corresponding to the image frame according to the object shape, aggregate the data points into multiple point cloud clusters, and filter out the point cloud clusters matching the shape of the construction markers. Then, the spatial position comparison tool is used to check whether the construction marker area marked in the image and the point cloud cluster matching the shape in the point cloud exist a corresponding relationship in the spatial position. If the corresponding relationship exists, the image frame data and the point cloud frame data are taken as the candidate key frame data. If the corresponding relationship does not exist, the candidate key frame data is not included in the candidate range. For example, in a certain image frame of the construction area in the suburb of A, the area where the construction markers are located is marked by the image recognition tool, and the corresponding point cloud frame data is divided into multiple point cloud clusters. The point cloud cluster matching the shape of the construction markers is filtered out. After the matching is confirmed by the spatial position comparison, the image frame and the point cloud frame are taken as the candidate key frame data.

[0062] Step 123, arrange the candidate key frame data in the collection time sequence, and select the candidate key frame data with a time interval not exceeding a preset time length as the final key frame data.

[0063] Among them, the collection time sequence refers to the sequence arranged from early to late according to the collection time of the image frame (or the point cloud frame) in the candidate key frame data; the preset time length is a time standard set in advance for judging whether the time interval of adjacent candidate key frame data is reasonable; the time interval refers to the difference value of the collection time of adjacent two candidate key frame data after arrangement; the final key frame data is a frame data combination selected from the candidate key frame data, with a time interval not exceeding the preset time length and having continuity in time.

[0064] In the embodiments of the present application, first, the collection time of all candidate key frame data is extracted by the time sorting tool, and these candidate key frame data are arranged in the order from early to late according to the collection time. Then, a preset time length is set, the collection time interval of adjacent two candidate key frame data after arrangement is calculated, and if the calculated time interval does not exceed the preset time length, the two candidate key frame data are retained, and if the time interval exceeds the preset time length, the latter candidate key frame data is removed. Finally, the retained candidate key frame data is taken as the final key frame data. For example, from the candidate key frame data of the construction area in the suburb of A, multiple groups of data are extracted and sorted according to the collection time. After setting the preset time length, the time interval of adjacent frames is calculated, the candidate frames with a time interval meeting the requirements are retained, the candidate frames with a time interval not meeting the requirements are removed, and finally the final key frame data is obtained.

[0065] Step 124, based on the collection position information of the final key frame data, the final key frame data is allocated to the corresponding target road section according to the city road section division rule to form a continuous key frame data sequence.

[0066] Wherein, the collection position information is the specific road position information of the crowd-sourcing vehicle when collecting the final key frame data, which is usually obtained by positioning tools; the city road section division rule is a standard for dividing the city road into different sections, which generally takes the fixed signs (such as street lamps, traffic signs) on both sides of the road as a reference to determine the range of the road section; the target road section is a specific road section formed according to the division rule; the continuous key frame data sequence refers to the combination of the final key frame data allocated to the same target road section and keeping coherent in time and space.

[0067] In the embodiments of the present application, first, the collection position information of each group of final key frame data is obtained by a positioning information extraction tool, and the city road section division rule is determined, that is, the road interval between adjacent fixed signs (such as street lamps, traffic signs) on both sides of the road is taken as the range of a target road section, then the collection position information of each group of final key frame data is compared with the range of each target road section by a position comparison tool, if the collection position of a group of final key frame is within the range of a target road section, the group of final key frame data is allocated to the target road section, then the final key frame data allocated to the same target road section is sorted by collection time from early to late, and it is confirmed that their collection positions are continuously distributed in the road section, the sorted and spatially coherent frame data is arranged into an ordered combination to form a continuous key frame data sequence, for example, the A suburban road is divided into target road sections by adjacent fixed signs, the final key frame with collection position within a target road section is allocated to the road section, after sorting by time and confirming spatial coherence, the continuous key frame data sequence of the road section is arranged.

[0068] Step 125, labeling the collection time of each frame data in the continuous key frame data sequence and integrating to generate a time-stamped regional feature data package.

[0069] Wherein, the integration refers to classifying the continuous key frame data sequence with labeled collection time by target road section, adding road section identification and construction scene description to arrange into a complete data package; the time-stamped regional feature data package is a data package containing specific target road section information, each frame data with collection time label (i.e. time stamp) and reflecting the construction scene characteristics of the road section, which is the core input data for subsequent semantic segmentation processing.

[0070] In the embodiments of the present application, first, the recorded collection time is extracted from each group of frame data of the continuous key frame data sequence by a time labeling tool, and the time information is clearly labeled in the corresponding frame data file. Then, the continuous key frame data sequence of the same target road section with labeled collection time is sorted by collection time from early to late through a data integration tool, and the road section identifier and construction scene description (such as the types of construction markers contained) are added in the data packet, ensuring that the time labeling of all frame data in the data packet is complete, and the road section information is accurate. Finally, the arranged contents are combined into a unified data packet, that is, a regional feature data packet with timestamp is generated. For example, the continuous key frame data sequence of a target road section in the suburb of A is processed, each group of frame data is labeled with collection time, and the road section identifier and scene description are added after sorting by time, and the regional feature data packet with timestamp of the road section is generated after integration.

[0071] The present application provides the following specific examples: in the temporary construction area in the suburb, first, the image frame data and point cloud frame data uploaded by each brand of crowd-sourced vehicle are extracted from the aggregated crowd-sourced data, and it is confirmed that each image frame and the corresponding point cloud frame are collected at the same time through time comparison. Frame data with time mismatch is temporarily excluded from processing. Then, in the image frames with confirmed time matching, the construction fence and cone barrel area is recognized and marked by the tool, and the corresponding point cloud frame data is divided into multiple point cloud groups. The point cloud groups matching the shape of the construction markers are screened out, and after confirming the matching of the construction area in the image and the point cloud group through spatial position comparison, the image frame and the point cloud frame are taken as candidate key frame data. Then, the collection time of all candidate key frames is extracted, arranged in order from early to late, and the time interval of adjacent candidate frames is calculated after setting a preset time length. The candidate frames with interval meeting the requirements are retained to obtain the final key frame data. Then, the target road section is divided according to the fixed markers on both sides of the road, the collection position information of the final key frame is obtained, the final key frame with collection position within the range of a target road section is assigned to the road section, the frame data is sorted by collection time and confirmed to be continuously distributed within the road section, and a continuous key frame data sequence is formed. Finally, the collection time of each group of frame data in the continuous key frame data sequence is labeled, and after arranging in time sequence, the identifier of the target road section and the scene description containing the types of construction markers are added, and the regional feature data packet with timestamp of the road section is integrated.

[0072] By performing steps 121-125, the embodiment of the application extracts and confirms time-matched image frames and point cloud frame data through step 121, avoiding scene mismatch problems caused by time differences, and preliminarily screening invalid data combinations; step 122 focuses on effective data containing construction markers, further ensures that the data reflects the same construction scene through spatial position matching, and provides high-quality candidate materials for subsequent screening; step 123 selects the final key frame according to the time continuity, eliminates data with too large time interval, and guarantees the continuity of data in the time dimension, which is convenient for reflecting the dynamic changes of the construction scene; step 124 distributes the final key frame according to the road section and forms a spatially coherent sequence, so that the data accurately corresponds to the specific road section, avoids confusion of data in different road sections, and improves the processing pertinence; step 125 generates a region feature data packet with timestamp by labeling time and integrating, so that the data is structured and standardized, and the association between road section, time and construction feature is clearly presented, providing ordered and accurate input data for subsequent semantic segmentation based on deep learning model, continuously improving the data effectiveness and pertinence of the overall process, and laying a solid foundation for the subsequent link of high-precision map incremental update.

[0073] In a possible embodiment, step 124, based on the collection position information of the final key frame data, distributes the final key frame data to the corresponding target road section according to the city road section division rule, to form a continuous key frame data sequence, including:

[0074] a1, obtaining the position information of the fixed markers arranged on both sides of the city road, and taking the road section between the two adjacent fixed markers as the division range of the target road section.

[0075] Wherein, the fixed marker is a marker installed on both sides of the road for a long time and the position of which will not change at will, including: street lamps, traffic signs, road milestones, etc.; the position information of the fixed marker is the specific position description of these markers on the road obtained by the positioning device, such as the corresponding coordinates or road section position; the road section between the two adjacent fixed markers refers to the section of road between the two fixed markers next to each other; the division range of the target road section is to define the road section between the two adjacent fixed markers as the boundary range of an independent road section.

[0076] In the embodiments of the present application, first, the position information of all fixed markers on both sides of the urban road is obtained by a device with positioning function or a map query tool, to ensure that the position of each fixed marker can be accurately recorded; then the fixed markers are arranged in order according to the extension direction of the road, and each two adjacent fixed markers are found out, and then the road part between the two adjacent fixed markers is determined as an independent road section; finally, the independent road section is defined as the division range of the target road section, and other adjacent fixed markers are processed in the same way to obtain the division range of all target road sections, for example, on a section of road in the A suburb, first, the position information of 10 street lamps on both sides of the road is obtained by the positioning device, the street lamps are sorted as street lamp 1 to street lamp 10 according to the direction of the road from east to west, then the road section between street lamp 1 and street lamp 2 is defined as the division range of target road section 1, the road section between street lamp 2 and street lamp 3 is defined as the division range of target road section 2, and so on to determine the division range of 9 target road sections.

[0077] a2, compare the collection position information of the final key frame data with the division range of each target road section, when the comparison result indicates that the position information of the final key frame data is within the division range of any target road section, the final key frame data is assigned to the corresponding target road section.

[0078] Wherein, the collection position information of the final key frame data is the specific road position recorded by the positioning device of the crowd-sourcing vehicle when obtaining the final key frame data; the comparison result is the conclusion obtained by comparing the collection position information of the final key frame data with the division range of each target road section, whether the collection position is within the range of a target road section; the final key frame data is assigned to the corresponding target road section, which means that when the collection position is within the range of a target road section, the group of final key frame data is attributed to the target road section, and this step allows each group of final key frame data to correspond to a specific target road section, to realize accurate matching of data and road sections.

[0079] In the embodiments of the present application, firstly, the corresponding collection position information is extracted from the attribute information of each set of final key frame data, and the division range of all target road segments determined through the a1 step is called out to ensure that the description formats of the collection position information and the division range of the road segment are consistent, facilitating subsequent comparison; then the position comparison tool is used to compare the collection position information of each set of final key frame data with the division range of each target road segment one by one to determine whether the collection position falls within the boundary range of a target road segment; finally, according to the comparison result, if the collection position of a set of final key frame data is within the division range of a target road segment, the set of final key frame data is assigned to the target road segment, and the data assignment is completed, and all final key frame data are processed in the same way. For example, in the processing of road data in the suburb of A, the collection position information of 5 sets of final key frame data is extracted, the division range of target road segments 1 to 9 determined in the a1 step is called out, and then position 1 is compared with the range of each target road segment one by one. It is found that position 1 is within the range of target road segment 2, so the final key frame data corresponding to position 1 is assigned to target road segment 2, position 2 is within the range of target road segment 3, and is assigned to target road segment 3, until the assignment of all data is completed.

[0080] a3, arrange the final key frame data assigned to the same target road segment in the order of collection time, calculate the distance between the positions of the two adjacent final key frame data after sorting, retain the adjacent final key frame data with a distance less than a preset value, and eliminate the latter final key frame data with a distance greater than or equal to the preset value, to form a continuous key frame data sequence.

[0081] In the embodiments of the present application, firstly, the corresponding collection position information is extracted from the attribute information of each set of final key frame data, and the division range of all target road segments determined through the a1 step is called out to ensure that the description formats of the collection position information and the division range of the road segment are consistent, facilitating subsequent comparison; then the position comparison tool is used to compare the collection position information of each set of final key frame data with the division range of each target road segment one by one to determine whether the collection position falls within the boundary range of a target road segment; finally, according to the comparison result, if the collection position of a set of final key frame data is within the division range of a target road segment, the set of final key frame data is assigned to the target road segment, and the data assignment is completed, and all final key frame data are processed in the same way. For example, in the processing of road data in the suburb of A, the collection position information of 5 sets of final key frame data is extracted, the division range of target road segments 1 to 9 determined in the a1 step is called out, and then position 1 is compared with the range of each target road segment one by one. It is found that position 1 is within the range of target road segment 2, so the final key frame data corresponding to position 1 is assigned to target road segment 2, position 2 is within the range of target road segment 3, and is assigned to target road segment 3, until the assignment of all data is completed.

[0082] In the embodiments of the present application, firstly, the collection time is extracted from the attribute of each group of data for all the final key frame data allocated to the same target road section, and these data are arranged into an ordered list in the order of collection time from early to late; then the collection position information of the adjacent two groups of data after sorting is obtained by a distance calculation tool, and the road distance between the two collection positions is calculated according to the position information; then the preset value set in advance is called out, and the calculated distance between the adjacent positions is compared with the preset value, if the distance is less than the preset value, the two adjacent data are retained, if the distance is greater than or equal to the preset value, the latter group of data is removed; finally, all the retained data are arranged into an ordered combination to form the continuous key frame data sequence of the target road section, for example, in the target road section 2 of the A suburb, 4 groups of final key frame data are allocated, the collection times are 8 o'clock, 8:10, 8:25 and 8:30 respectively, and the data are arranged in the order of time as data 1 (8 o'clock), data 2 (8:10), data 3 (8:25) and data 4 (8:30); the collection position of data 1 is position X, and the collection position of data 2 is position Y, the road distance from X to Y is 200 meters, the preset value is 300 meters, 200 meters is less than 300 meters, and data 1 and 2 are retained; the distance between data 2 (position Y) and data 3 (position Z) is 350 meters, which is greater than 300 meters, and data 3 is removed; data 3 is removed, and there is no need to calculate the distance between data 3 and 4, finally, data 1 and 2 are retained, and the continuous key frame data sequence of the target road section 2 is arranged.

[0083] The present application provides the following specific examples: on a section of urban road in the suburbs of A, first, the staff uses a device with positioning function to obtain the position information of the fixed markers on both sides of the road, a total of 9 position data of fixed markers; according to the extension direction of the road from north to south, the 9 fixed markers are sequentially numbered as marker 1 to marker 9, and the road interval between adjacent markers is defined as target road section A to target road section H, completing the target road section division. Then extract the collection position information from the 6 groups of final key frame data, call out the 8 target road section division ranges that have been determined, and compare them one by one through the position comparison tool: position A is within the range of target road section B, and is allocated to target road section B; position B is within the range of target road section C, and is allocated to target road section C; positions C, D, E and F are allocated to target road sections B, D, E and C respectively. Finally, for target road section B, the collection times are 9:05 and 9:12 on a certain date in the morning, after sorting by time, the road distance between the two collection positions is calculated as 180 meters, the preset value is 300 meters, 180 meters is less than 300 meters, and the two groups of data are retained to form the continuous key frame data sequence of target road section B; target road section C is processed in the same way, the distance between adjacent positions is calculated as 220 meters (less than 300 meters), and the data is retained to form a continuous sequence; other target road sections are processed in this way, and finally a continuous key frame data sequence is generated for each target road section to which data is allocated.

[0084] By performing a1-a3, the embodiment of the present application obtains the fixed marker position and divides the target road section through the a1 step, providing independent road section division with clear boundaries and conforming to the actual spatial law for urban roads, avoiding the problem of ambiguous road section range in subsequent data processing; the a2 step realizes the accurate matching of the final key frame data and the target road section through the collection position comparison, preventing the confusion of data of different road sections and ensuring the clear data attribution; the a3 step generates the continuous key frame data sequence with time order and position continuity through time sorting and distance screening, eliminating invalid data with discontinuous positions and improving the data effectiveness. The three are connected to form a complete "road section division-data allocation-sequence generation" process, providing a structured and high-quality data source for subsequent annotation of collection time and integration of the generation of region feature data packets with timestamps, laying the foundation for the subsequent link of high-precision map incremental update.

[0085] In a possible embodiment, S13, based on a deep learning model, performs semantic segmentation on the region feature data packet to identify the map change boundary, comprising:

[0086] Step 131, input the image frame data and the point cloud frame data in the timestamped region feature data packet into a deep learning model, perform semantic segmentation on the region containing the construction identifier in the image frame data and the point cloud cluster in the point cloud frame data corresponding to the region by the deep learning model, and obtain an image segmentation result and a point cloud segmentation result.

[0087] The deep learning model is a computer model that can autonomously learn data features and process data, which is used to distinguish different content categories in the data in this embodiment; the semantic segmentation is an operation of separating the image frame or the point cloud frame according to the content by using the deep learning model; the region containing the construction identifier is a picture part in the image frame with a construction-related object (such as a fence or a cone barrel); the point cloud cluster corresponding to the region is a combination of data points in the point cloud frame that matches the shape of the object in the region; the image segmentation result is a result of marking the region of the construction identifier in the image after semantic segmentation; and the point cloud segmentation result is a result of marking the point cloud cluster of the corresponding construction identifier in the point cloud after semantic segmentation.

[0088] In the embodiment of the present application, first, all image frame data and corresponding point cloud frame data are extracted from the timestamped region feature data packet, and the data are input into the deep learning model together; second, the input image frame data are processed by the deep learning model, the model finds the region containing the construction identifier in the image frame according to the learned construction identifier features, and distinguishes the region from the normal road region to form an image segmentation result; finally, the point cloud frame data corresponding to the image segmentation result are processed by the deep learning model, the model finds the point cloud cluster in the point cloud frame that matches the shape of the construction identifier region in the image, and distinguishes the point cloud cluster from other point clouds to form a point cloud segmentation result.

[0089] Step 132, associate the image segmentation result and the point cloud segmentation result of the same data frame to determine the overall region range containing the construction identifier.

[0090] The same data frame refers to a group of image frame data and a group of point cloud frame data corresponding to the same collection time in the timestamped region feature data packet; the association is a spatial position comparison of the image segmentation result and the point cloud segmentation result of the same data frame to confirm that both reflect the same construction identifier; and the overall region range containing the construction identifier is a complete construction region range description formed by combining the image segmentation result (the construction region in the visual picture) and the point cloud segmentation result (the construction region in the spatial position) of the same data frame.

[0091] In the embodiment of the present application, first, the image segmentation result and the point cloud segmentation result corresponding to the same data frame are screened out from the result obtained in step 131; second, the position of the "construction marker area" in the image segmentation result of the same data frame is compared with the spatial position of the "construction marker point cloud cluster" in the point cloud segmentation result by a spatial position comparison tool to confirm whether the two reflect the same construction marker; finally, the image segmentation result and the point cloud segmentation point cloud cluster that are consistent after comparison are combined to form an overall area range containing the construction marker, which contains both the picture range of the construction marker in the image and the spatial position range of the construction marker in the point cloud.

[0092] Step 133, compare the overall area range with the original area information of the corresponding target road section in the automatic driving high-precision map to mark the non-overlapping area edge.

[0093] The automatic driving high-precision map is a high-precision map used for navigation of an automatic driving vehicle, wherein the original area information of the corresponding target road section is the area description of the target road section in the map before updating; the comparison is an operation of placing the overall area range and the original area information in the same spatial reference system to check whether there is a difference between the two; the non-overlapping area edge is the boundary line of the non-overlapping part of the overall area range (construction area) and the original area information (normal road), that is, the boundary line between the construction area and the original normal road.

[0094] In the embodiment of the present application, first, the original area information of the target road section corresponding to the current processing object is called from the automatic driving high-precision map system to ensure that the original information can clearly reflect the state of the road section before updating; second, the multiple sets of overall area ranges obtained in step 132 are compared with the called original area information under the same spatial reference to check whether the overall area range has a corresponding part in the original area information; finally, a marking tool is used to mark the boundary line of the non-overlapping part of the overall area range and the original area information in the comparison result.

[0095] Step 134, arrange the non-overlapping area edges in chronological order to form the map change boundary of the target road section.

[0096] The arrangement is an operation of removing duplicates and splicing the non-overlapping area edges sorted by time to form a continuous line; the map change boundary of the target road section is the boundary line obtained after the arrangement, which can continuously reflect the difference between the construction area and the original road of the target road section, and is also the boundary that needs to be updated in the high-precision map.

[0097] In the embodiments of the present application, first, all non-overlapping region edges corresponding to the target road section marked by the collecting step 133 are collected, and the collection time of the overall region range corresponding to each edge is recorded; second, these non-overlapping region edges are sorted in the order from early to late according to the collection time; third, the sorted edges are arranged: repeated edges are removed, and edges with slight differences are spliced to ensure the continuity of the spliced lines; finally, the arranged continuous lines are determined as the map change boundary of the target road section.

[0098] The present application provides the following specific examples: in the processing of the suburban target road section A, first, three groups of image frame data and corresponding point cloud frame data with collection times of 9:05 am, 9:12 am and 9:18 am on a certain date are extracted from the timestamped region feature data packet of the road section, and the six groups of data are input into the deep learning model; the model performs semantic segmentation on each group of image frames, identifies and labels the "construction sign region" where the red enclosures and yellow cone barrels are located in the image, and performs segmentation on the corresponding point cloud frames to find and label the point cloud clusters matching the shape of the construction sign, obtaining three groups of image segmentation results and three groups of point cloud segmentation results. Then, the image segmentation results and the point cloud segmentation results of the same collection time of each group are screened out, and it is confirmed through spatial position comparison that the construction region in the image and the corresponding point cloud cluster position in the point cloud are completely matched, and three groups of overall construction region ranges are formed in combination. Then, the original region information of the target road section A is called from the autonomous driving high-precision map system, and it is displayed that the road section is "a two-lane normal road without construction region", the three groups of overall construction region ranges are compared with the original information, it is found that the construction region is a newly added part in the original road, and there are non-overlapping regions outside the construction region and at the intersection with the original road. The tool is used to mark the edges of these non-overlapping regions. Finally, the three groups of non-overlapping edges are collected, sorted according to the collection time, the 9:12 am edge completely coinciding with the 9:05 am edge is removed, the main body of the 9:05 am edge is spliced with the small adjustment part of the 9:18 am edge, a continuous line is formed, and the line is determined as the map change boundary of the suburban target road section A.

[0099] By performing steps 131-134, the embodiment of the application accurately identifies the construction marker related part in the image and the point cloud through semantic segmentation of step 131, provides two types of basic data from vision and space for subsequent determination of the construction range, and avoids the limitation of a single data type; step 132 forms a complete and accurate overall construction area range by associating the segmentation results of the same data frame, combining the visual picture and the spatial position information, and ensures that the reference basis for subsequent comparison is comprehensive and reliable; step 133 accurately finds out the difference between the newly added area of construction and the original road by comparing the overall range with the original information of the high-precision map, marks the non-overlapping edge that needs to be updated, and avoids missing or misjudging the part to be updated; step 134 forms a coherent map change boundary by arranging the edges by time and de-duplicating splicing, and provides a clear and unified "change reference" for subsequent multi-vehicle boundary fusion and generation of incremental update patches. The application continuously improves the accuracy and effectiveness of the data, and lays a key foundation for incremental update of the high-precision map of autonomous driving.

[0100] In a possible embodiment, step 131 performs semantic segmentation on the region containing the construction marker in the image frame data and the point cloud cluster in the point cloud frame data corresponding to the region through a deep learning model, to obtain image segmentation results and point cloud segmentation results, including:

[0101] b1, based on the acquisition position information and the acquisition angle information corresponding to the image frame data and the point cloud frame data of the same data frame, respectively, a mapping relationship between each pixel position in the image frame data and each point cloud coordinate in the point cloud frame data is established.

[0102] Wherein, the same data frame is a group of image frame data and a group of point cloud frame data collected by a crowdsourcing vehicle at the same time, and the two groups of data reflect the road scene at the same time; the acquisition position information is the specific road position of the crowdsourcing vehicle when collecting the data, and the acquisition angle information is the shooting or scanning angle of the camera for collecting the image and the device for collecting the point cloud relative to the road; the pixel position in the image frame data is the specific position of each small dot (pixel) in the picture, such as the pixel position of the upper left corner or the center of the picture; the point cloud coordinate in the point cloud frame data is a group of data reflecting the actual position of an object in three-dimensional space, which can be understood as a "position marker" of the object in the real space; the mapping relationship is a "one-to-one correspondence" relationship between each pixel position in the image and the corresponding point cloud coordinate in the point cloud, through which the real space position corresponding to a certain pixel in the image can be known.

[0103] In the embodiment of the present application, firstly, the collection position information and the collection angle information of the image frame data and the point cloud frame data in the same data frame are extracted respectively, and the information is carefully checked to ensure that the information is complete and error-free; secondly, using a spatial position corresponding calculation method, based on the collection position and the collection angle, combining the picture distribution rule of the image pixels and the spatial distribution rule of the point cloud coordinates, the corresponding point cloud coordinates of each pixel position in the image are calculated one by one; finally, all the calculated “pixel position-point cloud coordinate” corresponding relationships are sorted and recorded to form a complete mapping relationship.

[0104] b2, identifying a pixel set presenting the appearance features of the construction markers in the image frame data by the deep learning model to segment a region containing the pixel set as an image segmentation result.

[0105] The deep learning model is a computer processing tool that autonomously masters the appearance features of the construction markers by learning a large number of construction scene pictures and has been trained in advance; the appearance features of the construction markers are the external characteristics of commonly used objects in construction (such as red enclosures and yellow cone barrels), including color (red, yellow), shape (enclosure is rectangular, cone barrel is sharp at the top and round at the bottom), etc.; the pixel set is a group of pixels formed by all the pixels presenting the appearance features of the construction markers in the image, which are concentrated in the picture; the region containing the pixel set is the whole region of the concentrated pixel set in the image enclosed by lines to form a clear picture range; the image segmentation result is the picture region containing the construction markers enclosed to clearly show the position of the construction markers in the image.

[0106] In the embodiment of the present application, firstly, the image frame data in the same data frame processed in the b1 step is input into the deep learning model trained in advance, which has mastered the appearance features of the construction markers such as red enclosures and yellow cone barrels; secondly, the model analyzes the image frame data pixel by pixel to judge whether each pixel meets the appearance features of the construction markers, filters out all the pixels meeting the features, and aggregates to form a pixel set; finally, the model encloses the pixel set in the image as a whole with lines to form a continuous picture region, which is the image segmentation result containing the construction markers, ensuring that the segmentation region completely covers all the pixels corresponding to the construction markers.

[0107] b3, based on the mapping relationship, positioning the range of the point cloud coordinates corresponding to the pixel positions in the image segmentation result in the point cloud frame data, and determining the point cloud cluster in the range of the point cloud coordinates as the point cloud segmentation result.

[0108] Wherein, the mapping relationship is the one-to-one correspondence relationship of "pixel position-point cloud coordinate" established in the b1 step; the pixel position in the image segmentation result is the specific picture position of all pixels in the segmentation region obtained in the b2 step; the point cloud coordinate range is the real space range formed by collecting all point cloud coordinates corresponding to these pixel positions through the mapping relationship; the point cloud cluster is the point cloud set in the point cloud frame data located in the coordinate range and gathered according to the shape of the construction marker; and the point cloud segmentation result is the whole of these point cloud clusters, which can clearly indicate the position and shape of the construction marker in the real space.

[0109] In the embodiment of the present application, firstly, the mapping relationship established in the b1 step is called, and the positions of all pixels in the image segmentation result in the b2 step are extracted; secondly, according to the mapping relationship, the point cloud coordinates corresponding to each pixel position are found one by one, the distribution range is analyzed after the coordinates are collected, and the point cloud coordinate range is determined; finally, the point clouds located in the coordinate range are filtered out in the point cloud frame data, the point cloud gathering situation is observed, the point clouds gathered in the form of a long strip are classified into one point cloud cluster (corresponding to the enclosure), and the point clouds gathered in the form of three conical shapes are classified into three point cloud clusters (corresponding to the cone barrels), which together form the point cloud segmentation result, ensuring that each point cloud cluster corresponds to an actual construction marker.

[0110] The present application provides the following specific examples: In a suburban temporary construction road section, a group of the same data frames are collected by B brand crowd-sourcing vehicles. Firstly, the b1 step is performed: the collection position of the data frame is extracted as "A suburban construction road section near marker D", the image and point cloud collection angles are both "horizontally forward, with an angle of 25 degrees with the ground", the image has 1920 horizontal pixels and 1080 vertical pixels, the corresponding point cloud coordinates are calculated for each pixel one by one through spatial position correspondence calculation, and finally the mapping relationship of pixel position and point cloud coordinate is formed. Then, the b2 step is performed: the image frame data is input into the deep learning model, the model identifies 800 red pixels and 300 yellow pixels pixel by pixel, and the pixels are collected as a pixel set to form an image segmentation result by circling the pixels with lines. Finally, the b3 step is performed: the mapping relationship of b1 is called, the positions of 1100 pixels in the segmentation region are extracted, the corresponding point cloud coordinates are found and collected one by one, and the coordinate range is determined as "5-9 meters away from the device, 0.6-2.2 meters to the left"; the point clouds in the range are filtered out in the point cloud frame, the point clouds in the form of a long strip are classified into one point cloud cluster (corresponding to the enclosure), and the point clouds in the form of a conical shape are classified into three point cloud clusters (corresponding to the cone barrels), forming the point cloud segmentation result.

[0111] By performing b1~b3, the mapping relationship established in step b1 of the embodiment b1 of the present application realizes the effective association of the image "flat picture" and the point cloud "real space position", solves the problem that the two cannot be directly corresponded, and provides a key foundation for subsequent cross-data type processing; the step b2 accurately identifies the pixel set of the construction marker through the deep learning model and segments the area, avoiding the missed judgment and misjudgment of the construction area caused by manual identification or simple screening, and ensuring the clear and accurate position of the construction marker in the image; the step b3 positions the point cloud coordinate range based on the mapping relationship and forms a point cloud cluster, converts the flat segmentation result of the image into a point cloud segmentation result in the real space, makes the actual position and shape of the construction marker more explicit, and makes up for the deficiency that the space position cannot be perceived only by the image. The present application forms a complete construction marker identification process from data association to image recognition to space positioning, and provides accurate and comprehensive data support for subsequent determination of the overall range of the construction area and comparison of the high-precision map to find out the change boundary.

[0112] In a possible embodiment, S14, the processed results are calculated according to the dynamic weight fusion algorithm to generate an incremental update patch, and the automatic driving high-precision map is incrementally updated based on the incremental update patch, including:

[0113] Step 141, obtaining a plurality of map change boundaries respectively corresponding to a plurality of crowd-sourced vehicles, and collection time and position accuracy information of the map change boundaries.

[0114] Among them, the plurality of crowd-sourced vehicles are different social vehicles participating in the collection of urban road data, and these vehicles can record road construction related information; the map change boundary is the edge of the area changed in the automatic driving high-precision map due to construction identified by each crowd-sourced vehicle in the previous step, that is, the demarcation line between the construction area and the normal road; the collection time is the specific time when each crowd-sourced vehicle obtains the map change boundary; the position accuracy information is the degree of coincidence between the map change boundary and the actual construction area edge, which can reflect the accuracy of the boundary.

[0115] In the embodiment of the present application, first, the target processing section is determined, and the map change boundaries respectively corresponding to a plurality of crowd-sourced vehicles in the section are screened out from the pre-period crowd-sourced vehicle data, ensuring that each boundary can clearly reflect the edge of the construction area; second, for each screened map change boundary, the collection time (such as 9:10 am on a certain date) is extracted from the collection record of the corresponding vehicle, and the position accuracy information of the boundary is also extracted; finally, the "map change boundary", "collection time" and "position accuracy information" of each crowd-sourced vehicle are arranged as a group of data, forming a plurality of groups of basic data to be processed.

[0116] Step 142, according to the new and old degree of the collection time and the accurate degree of the position accuracy information, a corresponding weight value is assigned to the map change boundary corresponding to each crowd-sourced vehicle.

[0117] Wherein, the new and old degree of the collection time is the difference between the collection time of the map change boundary and the current time, the smaller the difference is, the newer the time is; the precision degree of the position accuracy information is the fitting degree of the map change boundary and the actual construction area edge, the higher the fitting degree is, the more accurate the precision is; the weight value is a value reflecting the "importance" of each map change boundary in subsequent integration, the higher the importance (the newer the time is, the more accurate the precision is), the larger the weight value is.

[0118] In the embodiment of the present application, first, the judgment standard of the new and old degree of the collection time and the judgment standard of the precision degree of the position accuracy are set; second, for each map change boundary, the new and old degree of the collection time and the precision degree of the position accuracy are judged respectively; finally, according to the preset "grade-weight" rule (such as "new + accurate" corresponding to high weight, "new + relatively accurate" corresponding to medium-high weight, "medium + inaccurate" corresponding to low weight), the weight value is assigned to each boundary.

[0119] Step 143, calculating and integrating each map change boundary by a dynamic weight fusion algorithm to obtain a comprehensive map change boundary.

[0120] Wherein, the dynamic weight fusion algorithm is a method of integrating after weighted calculation of multiple sets of boundary data according to the weight value of each map change boundary, the core is to make the boundary with high weight have greater influence on the final result; the weighted calculation is to multiply the key information (such as the position of the boundary feature point) of each map change boundary by its weight value; the integration is to summarize all the weighted boundary information and merge it into unified boundary data; the comprehensive map change boundary is the boundary after integration, which is closer to the actual construction situation.

[0121] In the embodiment of the present application, first, the key information (such as the position of the feature point on the boundary, represented by simplified coordinates) of each map change boundary is extracted to ensure the consistency of the information format; second, the position of the feature point of each boundary is weighted calculated by the dynamic weight fusion algorithm, that is, the coordinates of each feature point are multiplied by the corresponding weight; finally, the weighted coordinates of the same feature point are added to obtain the comprehensive feature point coordinates, and then the comprehensive feature points are connected in order to form the comprehensive map change boundary.

[0122] Step 144, generating an incremental update patch containing the part to be updated in the automatic driving high-precision map according to the comprehensive map change boundary.

[0123] Firstly, the original information of the target road section in the automatic driving high-precision map is called to view the original map content (such as whether there is construction annotation); secondly, the comprehensive map change boundary is compared with the original map to find out the "blank area" in the original map which has no construction annotation but actually has construction, and the "blank area" is determined as the part to be updated; finally, the content to be updated is arranged into a specific format file to ensure that only the part to be updated is contained, and an incremental update patch is formed.

[0124] Step 145, loading the incremental update patch to the update system of the automatic driving high-precision map to complete the incremental update of the high-precision map.

[0125] In the embodiment of the present application, the incremental update patch is a file generated in step 144 and containing the part of the map to be updated; the update system of the automatic driving high-precision map is a software system receiving the patch and modifying the existing map; loading is transmitting the patch to the update system and letting the system recognize the update content; and the incremental update is that the system only modifies the part to be updated without modifying the normal area.

[0126] In the embodiment of the present application, firstly, the incremental update patch is transmitted to the update system (such as a cloud platform) through a network to ensure that the patch is complete and undamaged; secondly, after the update system receives the patch, the patch content is automatically analyzed (such as identifying the coordinates and annotation information of the area to be updated) and located to the target road section in the high-precision map; finally, the system modifies the map according to the patch content, draws the comprehensive boundary in the specified area, adds the construction annotation, saves the map data after modification, and completes the update.

[0127] The present application provides the following specific examples: in the temporary construction section of the suburb of A, first perform step 141: collect the basic data of 3 crowd-sourced vehicles of B brand, C brand and D brand-B brand, 9:10 collection, small deviation, C brand, D brand, and organize into 3 groups of data. Then step 142: set the rule at the current time 10 o'clock, “new after 9 o'clock, small deviation for precision”, assign the weight of B brand 0.6, C brand 0.3, and D brand 0.1. Then step 143: calculate the comprehensive feature point through the dynamic weight fusion algorithm: A1 point X coordinate = 10*0.6+11*0.3+9*0.1 = 10.2, Y coordinate = 20*0.6+20*0.3+20*0.1 = 20; A2 point X coordinate = 30*0.6+31*0.3+29*0.1 = 30.2, Y coordinate = 20, and the comprehensive boundary (10.2, 20) to (30.2, 20) is obtained after connection. After that, step 144: compare the original map (north lane (10, 20) to (30, 20) without construction annotation), determine (10.2, 20) to (30.2, 20) as the area to be updated, and generate the incremental update patch containing only “the area adds construction annotation + red comprehensive boundary”. Finally, step 145: transmit the patch to the cloud update system, and the system is positioned to the road section in the suburb of A after analysis, draws a red boundary in the specified area, adds “temporary construction area” annotation, retains other normal information, and completes the incremental update of the high-precision map.

[0128] By performing steps 141-145, the present application collects multi-source crowd-sourced data and supporting information through step 141, avoids single data deviation, and provides a comprehensive basis for subsequent processing; step 142 assigns weights according to data quality, so that high-quality data dominates in integration and reduces the interference of inferior data; step 143 generates a high-precision comprehensive boundary through dynamic weight fusion, accurately reflects the actual construction area edge, and improves the boundary accuracy; step 144 generates an incremental patch containing only the updated part, avoids resource waste caused by full map redrawing, and reduces the data volume; step 145 quickly loads the patch to complete the update without damaging normal map data, and improves the update efficiency. The present application not only ensures the accuracy of the high-precision map update for autonomous driving, but also realizes efficient update, so that the map can timely reflect temporary construction changes and provide reliable navigation basis for safe driving of autonomous vehicles.

[0129] In a possible embodiment, step 142 assigns a corresponding weight value to the map change boundary corresponding to each crowd-sourced vehicle according to the new and old degree of the collection time and the precision degree of the position accuracy information, including:

[0130] c1, determine the current time, calculate the time interval of the collection time of each map change boundary and the current time, the length of the time interval is used to determine the new and old degree of the collection time, and the position error value in the position accuracy information corresponding to each map change boundary is extracted, the size of the position error value is used to determine the accurate degree of the position accuracy.

[0131] Wherein, the current time is the specific time when the data processing is performed, which is used to measure the early or late of the collection time; the map change boundary is the boundary line between the construction area and the normal road identified by each crowd-sourced vehicle in the previous step; the collection time is the specific time when each crowd-sourced vehicle obtains the corresponding map change boundary; the time interval is the difference value obtained by subtracting the collection time from the current time, and the length of the difference value can reflect the new and old of the collection time; the position accuracy information is the information reflecting the fitting degree of the map change boundary and the actual construction edge, and the position error value in the information is the specific number, and the size of the value can reflect the accuracy of the position accuracy, and the smaller the value is, the more accurate the accuracy is.

[0132] In the embodiment of the application, first, the specific time when the data processing is performed is determined by the clock tool to ensure the accuracy of the time; second, the collection time of each map change boundary is extracted from the previously collected multiple sets of "map change boundary to collection time to position accuracy information" data, the time interval is calculated by subtracting the collection time from the current time, and the new and old of the collection time is judged according to the length of the interval; finally, the specific number reflecting the deviation of the boundary and the actual construction edge (i.e. the position error value) is extracted from the position accuracy information of each map change boundary, and the accuracy of the position accuracy is judged according to the size of the value (the smaller the value is, the more accurate the accuracy is, and the larger the value is, the less accurate the accuracy is).

[0133] c2, determine the time weight value of each map change boundary based on the corresponding relationship between the preset time interval range and the time weight value, and determine the accuracy weight value of each map change boundary based on the corresponding relationship between the preset position error value range and the accuracy weight value.

[0134] Wherein, the corresponding relationship between the preset time interval range and the time weight value is a rule set in advance, which clearly shows that different time interval intervals correspond to different time weight values; the time weight value is a value determined according to the time interval range, which is used to reflect the contribution of the new and old of the collection time to the importance of the data; the corresponding relationship between the preset position error value range and the accuracy weight value is another set of rules set in advance, which clearly shows that different position error intervals correspond to different accuracy weight values; the accuracy weight value is a value determined according to the position error range, which is used to reflect the contribution of the accuracy of the position accuracy to the importance of the data.

[0135] In the embodiment of the present application, firstly, the "time interval range-time weight value" corresponding rule set in advance is called; secondly, the time interval of each map change boundary calculated in the c1 step is checked to determine which interval range it belongs to, and then the time weight value is determined according to the rule; then the "position error value range-precision weight value" corresponding rule set in advance is called; finally, the position error value of each boundary extracted in the c1 step is checked to determine which error range it belongs to, and then the precision weight value is determined according to the rule.

[0136] c3, the comprehensive value of the time weight value and the precision weight value of each map change boundary is calculated, and the comprehensive value is taken as the weight value of the corresponding map change boundary.

[0137] Among them, the time weight value is used to reflect the contribution of the new and old of the collection time to the importance of the data; the precision weight value is used to reflect the contribution of the accuracy of the position to the importance of the data; the comprehensive value is a single value obtained by merging the "time weight value+precision weight value", which can comprehensively reflect the weight contribution of two dimensions; the weight value of the corresponding map change boundary is the comprehensive value, which can reflect the importance of the data quality (combination of time new and old and accuracy accurate) of the boundary to the subsequent integration process.

[0138] In the embodiment of the present application, the calculation method of the comprehensive value is "comprehensive value = time weight value + precision weight value", which is simple and intuitive and can combine the contribution of two dimensions.

[0139] The present application provides the following specific examples: in the data processing of the temporary construction section in the A suburb, firstly, the c1 step is executed: the current time is determined to be 10:00 on a certain date in the morning through the system clock, the collection time is extracted from the boundary data of the B brand, C brand and D brand crowdsourcing vehicles, and the time interval is calculated: 10:00-9:10=50 minutes (B brand), 10:00-9:30=30 minutes (C brand), 10:00-8:20=100 minutes (D brand); then the error value (B brand 0.5 meters, C brand 1.2 meters, D brand 2.5 meters) is extracted from the position accuracy information. Then, the c2 step is executed: the preset "time interval-time weight" rule is called to determine the time weight value of B and C brands 0.4 and D brand 0.2; the "position error-precision weight" rule is called to determine the precision weight value of B brand 0.2, C brand 0.1 and D brand 0.05. Finally, the c3 step is executed: according to "comprehensive value = time weight value + precision weight value", B brand 0.4+0.2=0.6 (weight value 0.6), C brand 0.4+0.1=0.5 (weight value 0.5), D brand 0.2+0.05=0.25 (weight value 0.25), and finally the final weight value of the three boundaries is obtained.

[0140] By performing c1-c3, the application embodiment c1 step converts the two fuzzy concepts of "acquisition time new and old" and "position precision accurate" into quantifiable values by determining the current time, calculating the time interval and extracting the position error value, avoids subjective judgment deviation, and provides an objective basis for subsequent weight calculation; the c2 step converts the quantified interval and error into specific time weight value and precision weight value relying on the preset corresponding rule, realizes the ordered conversion from data features to weight contribution, ensures that the weight distribution has rules to follow, and can reflect the data quality difference; the c3 step combines the weights of the two dimensions by a simple and intuitive calculation method to obtain the final weight value that can comprehensively reflect the data quality, simplifies the subsequent fusion calculation, ensures that high-quality data (time new and precision accurate) can play a greater role in integration, forms a complete process from quantified features, split weights to combined weights, lays a solid foundation for subsequent dynamic weight fusion to generate high-precision comprehensive map change boundaries, and effectively improves the accuracy and rationality of overall data processing.

[0141] Figure 3 A structure schematic diagram of an automatic driving high-precision map incremental update system based on crowd-sourced data provided by the application embodiment is shown in Figure 3 The system comprises:

[0142] The acquisition module 31 is configured to acquire image and point cloud data of a city road temporary construction area uploaded by a plurality of crowd-sourced vehicles, and the image and point cloud data jointly constitute crowd-sourced data.

[0143] The screening module 32 is configured to perform key frame screening on the crowd-sourced data, perform spatio-temporal correlation analysis on the screened key frame data, and generate a region feature data packet with a time stamp.

[0144] The processing module 33 is configured to perform semantic segmentation on the region feature data packet based on a deep learning model, to identify a map change boundary, and to process a plurality of map change boundaries corresponding to the plurality of crowd-sourced vehicles, respectively.

[0145] The generation module 34 is configured to calculate the processed result according to a dynamic weight fusion algorithm, to generate an incremental update patch, and to perform incremental update on the automatic driving high-precision map based on the incremental update patch.

[0146] The automatic driving high-precision map incremental update system based on crowd-sourced data of the application embodiment is used to implement the foregoing automatic driving high-precision map incremental update method based on crowd-sourced data, and therefore the specific embodiments in the automatic driving high-precision map incremental update system based on crowd-sourced data can be seen from the embodiment part of the automatic driving high-precision map incremental update method based on crowd-sourced data in the foregoing, and the specific embodiments can be referred to the description of the corresponding embodiment part, which will not be described here.

[0147] The application further provides an electronic device, comprising a memory for storing a computer program, and a processor for executing the computer program to implement the steps of the automatic driving high-precision map incremental updating method based on crowdsourcing data.

[0148] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the automatic driving high-precision map incremental updating method based on crowdsourcing data.

[0149] In an example embodiment, the computer readable storage medium can include, but is not limited to, a U disk, a read-only memory, a random access memory, a mobile hard disk, a magnetic disk or an optical disk, and various media capable of storing computer programs.

[0150] The embodiments of the application further provide a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the steps in the automatic driving high-precision map incremental updating method based on crowdsourcing data.

[0151] The skilled person can further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been described in general terms in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the application.

[0152] The above describes in detail the automatic driving high-precision map incremental updating method and system based on crowdsourcing data provided by the application. The principles and implementation modes of the application are described by applying specific examples in this paper, and the above example descriptions are only used to help understand the method of the application and its core idea. It should be pointed out that for ordinary skilled persons in the technical field, some improvements and modifications can be made to the application without departing from the principles of the application, and these improvements and modifications also fall within the protection scope of the application.

Claims

1. An automatic driving high-precision map incremental updating method based on crowdsourcing data, characterized in that, The method comprises the following steps: acquiring image and point cloud data of temporary construction area of urban road uploaded by multiple crowd-sourcing vehicles, wherein the image and the point cloud data jointly constitute crowd-sourcing data; performing key frame screening on the crowd-sourcing data, and performing spatio-temporal correlation analysis on screened key frame data to generate time-stamped regional feature data packet; performing semantic segmentation on the regional feature data packet based on a deep learning model to identify map change boundary, and processing the map change boundary corresponding to the multiple crowd-sourcing vehicles respectively; calculating the processed result according to a dynamic weight fusion algorithm to generate an incremental update patch, and performing incremental update on an automatic driving high-precision map based on the incremental update patch; the key frame screening on the crowd-sourcing data and the spatio-temporal correlation analysis on the screened key frame data to generate the time-stamped regional feature data packet comprise the following steps: extracting image frame data and point cloud frame data in the crowd-sourcing data, and confirming that the image frame data and the point cloud frame data correspond to the same collection time; labeling a region containing construction markers in the image frame data, and dividing the point cloud frame data into point cloud clusters matching the shape of the construction markers, if the region and the point cloud cluster have a corresponding relationship, the image frame data and the point cloud frame data are taken as candidate key frame data; arranging the candidate key frame data in the order of collection time, and selecting candidate key frame data with a time interval not exceeding a preset time length as final key frame data; based on the collection location information of the final key frame data, the final key frame data is distributed to the corresponding target road section according to the urban road section division rule to form a continuous key frame data sequence; labeling the collection time of each frame of data in the continuous key frame data sequence and integrating to generate a time-stamped regional feature data packet.

2. The method of claim 1, wherein, the distribution of the final key frame data to the corresponding target road section according to the collection location information of the final key frame data and the urban road section division rule to form a continuous key frame data sequence comprises the following steps: acquiring the position information of the fixed markers arranged on both sides of the urban road, and taking the road interval between two adjacent fixed markers as the division range of the target road section; comparing the collection location information of the final key frame data with the division range of each target road section, and when the comparison result indicates that the position information of the final key frame data is within the division range of any target road section, the final key frame data is distributed to the corresponding target road section; arranging the final key frame data distributed to the same target road section in the order of collection time, calculating the distance between the positions of two adjacent final key frame data after sorting, retaining adjacent final key frame data with a distance less than a preset value, and eliminating the latter final key frame data with a distance greater than or equal to the preset value to form a continuous key frame data sequence.

3. The method of claim 1, wherein, the semantic segmentation on the regional feature data packet based on the deep learning model to identify the map change boundary comprises the following steps: Input the image frame data and the point cloud frame data in the timestamped area feature data packet into a deep learning model, perform semantic segmentation on an area containing a construction identifier in the image frame data and on a point cloud cluster in the point cloud frame data corresponding to the area by the deep learning model, and obtain an image segmentation result and a point cloud segmentation result; Associate the image segmentation result and the point cloud segmentation result of the same data frame, and determine an overall area range containing the construction identifier; Compare the overall area range with original area information of a corresponding target road segment in an automatic driving high-definition map, and mark out a non-overlapping area edge; Arrange the non-overlapping area edge in a time sequence, and form a map change boundary of the target road segment.

4. The method of claim 3, wherein, The method for performing semantic segmentation on an area containing a construction identifier in image frame data and on a point cloud cluster in point cloud frame data corresponding to the area by the deep learning model, and obtaining an image segmentation result and a point cloud segmentation result, comprises: Establish a mapping relationship between each pixel position in the image frame data and each point cloud coordinate in the point cloud frame data based on the acquisition position information and the acquisition angle information corresponding to the image frame data and the point cloud frame data of the same data frame; Identify a pixel set presenting an appearance feature of a construction identifier in the image frame data by the deep learning model, divide an area containing the pixel set as an image segmentation result, and determine a point cloud cluster in the point cloud frame data corresponding to the pixel position in the image segmentation result as a point cloud segmentation result. The method for calculating the processed result according to a dynamic weight fusion algorithm, generating an incremental update patch, and performing incremental update on the automatic driving high-definition map based on the incremental update patch, comprises:

5. The method of claim 1, wherein, Obtain a plurality of map change boundaries corresponding to a plurality of crowd-sourcing vehicles respectively, and acquisition time and position accuracy information of the map change boundaries; Determine a corresponding weight value for each map change boundary corresponding to a crowd-sourcing vehicle according to a new and old degree of the acquisition time and a precise degree of the position accuracy information; Calculate and integrate each map change boundary by a dynamic weight fusion algorithm to obtain a comprehensive map change boundary; Generate an incremental update patch containing a part to be updated in the automatic driving high-definition map according to the comprehensive map change boundary; Load the incremental update patch to an update system of the automatic driving high-definition map to complete the incremental update of the high-definition map. The method for determining a corresponding weight value for each map change boundary corresponding to a crowd-sourcing vehicle according to a new and old degree of the acquisition time and a precise degree of the position accuracy information, comprises:

6. The method of claim 5, wherein, Determine a current time, calculate a time interval between the acquisition time of each map change boundary and the current time, and determine a position error value in the position accuracy information corresponding to each map change boundary, wherein the length of the time interval is used to determine the new and old degree of the acquisition time, and the size of the position error value is used to determine the precise degree of the position accuracy. ​ The time weight value of each map change boundary is determined based on a preset correspondence between a time interval range and the time weight value, and the precision weight value of each map change boundary is determined based on a preset correspondence between a position error numerical range and the precision weight value; A comprehensive value of the time weight value and the precision weight value of each map change boundary is calculated, and the comprehensive value is taken as the weight value of the corresponding map change boundary.

7. An automatic driving high-precision map incremental updating system based on crowdsourcing data, characterized in that, It comprises: An acquisition module is configured to acquire image and point cloud data of a temporary construction area of an urban road uploaded by a plurality of crowd-sourcing vehicles, wherein the image and the point cloud data jointly constitute crowd-sourcing data; A screening module is configured to perform key frame screening on the crowd-sourcing data, and perform spatio-temporal correlation analysis on the screened key frame data to generate a region feature data package with a time stamp; A processing module is configured to perform semantic segmentation on the region feature data package based on a deep learning model to identify a map change boundary, and process the map change boundary corresponding to each of the plurality of crowd-sourcing vehicles; A generation module is configured to calculate a processed result according to a dynamic weight fusion algorithm, generate an incremental update patch, and perform incremental update on an automatic driving high-precision map based on the incremental update patch; The key frame screening on the crowd-sourcing data and the spatio-temporal correlation analysis on the screened key frame data to generate a region feature data package with a time stamp comprises: Image frame data and point cloud frame data in the crowd-sourcing data are extracted, and it is confirmed that the image frame data and the point cloud frame data correspond to the same collection time; Regions containing construction markers in the image frame data are marked, and the point cloud frame data is divided into point cloud clusters matching the shape of the construction markers, and if the regions and the point cloud clusters have a corresponding relationship, the image frame data and the point cloud frame data are taken as candidate key frame data; The candidate key frame data is arranged in a collection time sequence, and candidate key frame data with a time interval not exceeding a preset time length is selected as final key frame data; Based on collection location information of the final key frame data, the final key frame data is distributed to a corresponding target road segment according to a city road segment division rule to form a continuous key frame data sequence; Each frame of data in the continuous key frame data sequence is labeled with collection time and integrated to generate a region feature data package with a time stamp.

8. An electronic device, comprising: It comprises: A memory is configured to store a computer program; A processor is configured to execute the computer program to implement the steps of the automatic driving high-precision map incremental update method based on crowd-sourcing data according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer program is stored in the computer readable storage medium, and the computer program can be executed by the processor to implement the automatic driving high-precision map incremental update method based on crowd-sourcing data according to any one of claims 1 to 6.

Images