Method for Generating and Updating High-Definition Maps in an Autonomous Driving Scenario Library

Abstract

The present invention discloses a method for generating and updating high-precision maps for an autonomous driving scenario library, belonging to the field of autonomous driving technology. The method comprises the following steps: S1: performing element feature recognition on the image obtained by frame extraction, extracting the road network element features within the target detection frame, and retaining the valid target detection frame; S2: constructing road network element connection relationships; S3: compensating for missing road network elements and connection relationships; S4: semanticizing the road network elements and configuring element attributes, and outputting them as high-precision map data. The present invention overcomes the problems of low element recognition accuracy, imperfect connection relationship construction, and poor robustness of missing compensation in existing maps. By utilizing high-altitude image video for frame extraction processing and element feature recognition, the method can obtain more comprehensive and accurate road network information, thereby improving the accuracy of high-precision maps.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a method for generating and updating high-precision maps of an autonomous driving scenario library. Background Art

[0002] With the rapid development of autonomous driving technology, high-precision maps, as a key component of the autonomous driving system, play a vital role in the safety and reliability of autonomous driving due to their accuracy and real-time performance.

[0003] High-precision maps are a core supporting technology for autonomous driving systems. They contain a wealth of road information, such as lane markings, traffic signs, and intersections, enabling precise positioning and navigation for autonomous vehicles. By accurately describing road geometry, traffic signs, lane attributes, and other information, they provide critical data for vehicle positioning, path planning, and decision-making. Especially during the autonomous driving testing phase, a high-precision map format with clear road features is required to efficiently test and optimize autonomous driving algorithms. This ensures that intelligent vehicles meet safety requirements when on the road and select the most appropriate driving strategy based on actual road conditions, ensuring efficient and safe autonomous driving.

[0004] However, existing high-precision map generation technologies still face many technical problems. Traditional methods for identifying small targets (such as traffic markings and signs) are easily limited by image resolution, resulting in insufficient accuracy in identifying road feature elements and prone to missed or false detections. Existing map coordinate system conversion errors are large. When image data collected by drones or vehicle-mounted equipment is converted into the actual map coordinate system, the device posture and the influence of the earth's curvature are not fully considered, resulting in spatial mapping deviations. In many map systems, the connection relationships between road network elements are not fully constructed. The correlation determination of elements such as lanes, intersections, and traffic markings relies on simple rules, and lacks dynamic analysis of traffic sign semantics and lane direction consistency. Existing technologies do not provide a reliable compensation method for missing data, or the compensation effect is poor and inconsistent with the actual road network topology, affecting the reliability of the map.

[0005] These issues result in insufficient accuracy in complex scenarios, making it difficult to meet the high safety requirements of autonomous vehicles. Therefore, a method for efficiently generating and dynamically updating HD maps is urgently needed to improve element recognition accuracy, optimize connection relationship building logic, and implement intelligent loss compensation and exception handling. Summary of the Invention

[0006] The present invention overcomes the problems of low element recognition accuracy, imperfect connection relationship construction, and poor robustness of missing compensation in existing maps. It provides a method for generating and updating high-precision maps of autonomous driving scenario libraries. It uses high-altitude image videos for frame extraction processing and element feature recognition, combines multiple data processing technologies, and integrates multi-source data to obtain more comprehensive and accurate road network information, improve the accuracy of high-precision maps, compensate for missing road network elements and connection relationships, and ensure that map data is consistent with actual road conditions.

[0007] In order to achieve the above object, the present invention adopts the following scheme:

[0008] A method for generating and updating a high-precision map of an autonomous driving scenario library includes the following steps:

[0009] S1: Acquire high-altitude video and perform frame extraction, set a target detection frame on the extracted image, and extract the features of each road network element within the target detection frame, including element attributes and relevance;

[0010] S2: Constructing the connection relationship of road network elements based on the element attributes and associations of each road network element, including the following processes:

[0011] Determine the lane-to-lane connection based on the relative position, direction consistency, and traffic markings of the two lanes;

[0012] Determine the lane-intersection connection relationship based on the lane direction and the direction of the intersection traffic sign;

[0013] Establish a binding relationship between traffic markings and lanes based on the minimum distance matching principle between traffic markings and lanes and the functional rules of traffic markings;

[0014] Determine the traffic sign-lane mapping relationship based on spatial proximity and matching of sign semantic rules with lane attributes;

[0015] S3: Verify the integrity of the connection relationship between road network elements and compensate for missing road network elements and connection relationships, including:

[0016] Compensate for missing lane elements based on adjacent lane information, intersection connectivity, and traffic markings;

[0017] Compensate for missing traffic marking elements based on identified traffic markings and their continuity principles and setting rules;

[0018] Reanalyzes ambiguous lane connections, repairs incorrect lane connections, and compensates for missing lane connections. For newly identified lanes, supplements the connection between the new lane and the already identified lanes.

[0019] S4: Semantize the compensated road network elements and connection relationships and configure element and relationship attributes. Format the geometric expressions of the elements and relationship attributes in XML and output them as high-precision map data.

[0020] Preferably, in step S1, after setting the target detection frame for the image obtained by frame extraction, the step of converting the image coordinate system into the actual map coordinate system is further included, which includes the following specific steps:

[0021] In the image coordinate system, take half of the image width and height pixel values to get the coordinates of the image center, and translate the vertex coordinates of the target detection box to the coordinate system with the image center as the origin;

[0022] Construct the rotation matrix R(θ) based on the device yaw angle θ:

[0023]

[0024] The coordinates after translation (x p ,y p ) to perform rotation transformation:

[0025]

[0026] The rotated coordinates (x r ,y r ) is superimposed on the actual geographic coordinates of the device to complete the spatial position mapping, and the result is corrected by Gauss-Krüger projection to eliminate the error of the earth curvature.

[0027] Preferably, step S1 further includes enhancing recognition of a small target detection frame whose area is smaller than a preset threshold, which specifically includes the following steps:

[0028] For target areas whose area is smaller than the preset pixel threshold, the bicubic interpolation algorithm is used to magnify them by a factor of 2:

[0029]

[0030] Among them, (u, v) is the pixel coordinate of the image after interpolation and amplification, I zoom (u, v) is the pixel value of the amplified image, I(x+i, y+j) is the pixel value of the original image; W(t) is the weight function of bicubic interpolation, and its expression is:

[0031]

[0032] Where a is the bicubic interpolation smoothing coefficient, corresponding to hour, ; corresponds to When ;

[0033] Adaptive contrast enhancement is performed on the enlarged area, the enlarged area is divided into 8×8 sub-blocks, the sub-blocks are histogram-cropped and the histogram grayscale increment is limited to less than or equal to 3;

[0034] The non-local means denoising algorithm is used to denoise the sub-blocks: the search window size is set to 21×21, the similar block size is set to 7×7, and the pixel weight is calculated by comparing similar blocks within the search window;

[0035] The element features of the enhanced image area are re-identified. When the intersection-over-union ratio of the new and old target detection frames exceeds 0.6, the confidence of the new and old target detection frames is calculated. If the confidence of the new target detection frame is at least 0.2 higher than that of the old target detection frame, the element list is updated with the element features of the new target detection frame.

[0036] Preferably, in step S1, the overlap of road network elements is further detected, and the overlapping road network elements are hierarchically filtered to retain valid target detection frames, wherein the hierarchical filtering of the overlapping road network elements includes the following specific steps:

[0037] Set the IoU threshold for traffic sign elements, lane line elements, and intersection elements respectively. When the IoU of the target detection box exceeds the corresponding IoU threshold, perform target detection box conflict analysis:

[0038] The conflict graph model is constructed with the target detection boxes as nodes and the overlapping conflict relationships between target detection boxes as edges. The node weights are calculated as follows:

[0039]

[0040] Among them, w i is the node weight of the i-th node, A i is the area of the i-th target detection box, α is the category priority coefficient, and β is the area influence coefficient; P r (C i ) is the category priority value, which represents the node category C of the i-th node i Importance priority;

[0041] The branch-and-bound algorithm is used to traverse the conflict graph model. During the traversal process, the current optimal solution of the node is solved and updated until the upper bound of the subproblem is less than the current optimal value, and the branch is terminated to obtain the maximum weighted independent set.

[0042] In the maximum weighted independent set, if there is a cross-category conflict between nodes, the target detection box with the larger category priority value in the cross-category conflict is retained; if there is a same-category conflict between nodes, the node weight w is retained. i In the first 10% of target detection frames, delete the target detection frames that are not selected for saving and complete the hierarchical filtering of overlapping road network elements.

[0043] Preferably, in step S2, the lane-to-lane connection relationship is determined by:

[0044] The lane centerline endpoints are extracted to generate direction vectors. The angle between the two lane directions is calculated using the dot product of the vectors. If the angle is less than a predetermined threshold and the minimum centerline distance is less than a distance threshold, the directions are considered consistent.

[0045] A sliding window detection is performed on the gap area between adjacent lanes. The frequency of occurrence of white dashed lines and solid yellow lines is counted. A connection relationship is forced to be established between adjacent lanes where the frequency of white dashed lines is greater than 0.8, and a connection relationship is prohibited between adjacent lanes where the frequency of yellow solid lines is greater than 0.6.

[0046] Preferably, in step S1, the element attributes include lane type, lane direction, and lane speed limit information, and identifying and setting the lane speed limit information attributes includes the following steps:

[0047] Calculate the curvature radius R of the lane centerline and calculate the curve speed limit V based on the curvature radius of the lane centerline c :

[0048]

[0049] Where μ is the tire-road friction coefficient, g is the acceleration due to gravity, and θ c is the transverse slope angle;

[0050] If a speed limit sign value is detected, the lane speed limit information attribute is set to the smaller value of the curve speed limit value and the speed limit sign value; the uphill speed limit on steep slopes is reduced by 2km / h for every 1% slope gradient; and the speed transition is performed using linear interpolation within 50m before and after the speed limit change interval of the lane.

[0051] Preferably, in step S3, lane elements are compensated using the following method:

[0052] Set three control points P0, P1, and P2 in the identified part of the missing lane and calculate the directional derivative α of the adjacent lane:

[0053]

[0054] Among them, (x start ,y start ) is the starting coordinate of the adjacent lane, (x end ,y end ) are the coordinates of the end point of the adjacent lane;

[0055] Calculate the newly added control point P3:

[0056]

[0057] in, is the direction vector, ; k is the expansion coefficient, ;

[0058] Extend the missing lanes based on the cubic Bezier curve and construct the parametric equation:

[0059]

[0060] Set the maximum curvature radius limit of the curve. If the calculated curvature radius exceeds the limit, adjust the position of P3 to , ;

[0061] At the intersection, a uniform B-spline curve is generated with the number of connected lanes as the order, and the node vectors of the curve are distributed in equal proportions; the curvature continuity of the interrupted lane dividing lines is repaired according to the parabolic equation.

[0062] Preferably, in step S3, the method of compensating the traffic marking elements includes the following steps:

[0063] Compensate for missing dashed line segments in traffic marking elements:

[0064] Perform Fourier transform on the solid line segments before and after the dashed line, extract the spectrum features and calculate the phase consistency. If the phase consistency is greater than 0.85, generate a new dashed line in the missing segment according to the standard dashed-to-solid ratio.

[0065] To compensate for missing flow lines:

[0066] The reference radius R is obtained based on the curvature radius R of the lane centerline b :If R≤50m, then the corresponding R b =R; otherwise R b =50+0.2(R-50); the involute equation in the polar coordinate system is superimposed on the sine wave coefficient to generate a natural transition trajectory, and the polar coordinate system is converted to the Cartesian coordinate system. The steering angle is discretely sampled to generate a continuous guide line for compensation.

[0067] Preferably, in step S2, after determining the lane-intersection connection relationship, the intersection is also marked with a driving path priority, including the following steps:

[0068] Assign unique serial numbers to each entrance and exit lane of the intersection and generate all legal driving paths based on traffic rules;

[0069] Count the historical driving times of each driving route, normalize the driving times of the same direction path, and generate the efficiency score E ij :

[0070]

[0071] Among them, C ij is the number of vehicles traveling from the i-th import lane to the j-th exit lane, C max is the maximum number of trips along the same direction;

[0072] For each driving path, calculate the conflict area A between it and the driving paths that can be driven at the same time c and the number of intersections N c , the conflict area and the number of intersections are integrated into the conflict index to generate the safety score S ij :

[0073]

[0074] Where σ is the exponential adjustment coefficient;

[0075] The comprehensive priority score of each driving path is obtained by weighting the efficiency score and safety score by weight coefficient:

[0076]

[0077] Among them, F ij is the comprehensive priority score of the driving path from the i-th import lane to the j-th exit lane, and γ is the normalized weight coefficient.

[0078] Preferably, in step S3, data anomaly processing is also performed in the following manner:

[0079] When the number of compensations for an area exceeds three times or the final confidence level of the area is lower than 0.4, the boundary of the area is extracted and a spiral search path is generated to control the drone to retake the photo. The search path radius is expanded from 10 meters to 50 meters in steps of 5 meters, and the search angle is increased by 30 degrees. The newly collected data is fused with the original data through Kalman filtering. The state equation of Kalman filter fusion adopts a uniform velocity model to construct the observation equation, and the observation matrix is used to directly map the observation position coordinates.

[0080] The present invention has at least the following beneficial effects: (1) Utilizing high-altitude image video for frame extraction processing and element feature recognition, combined with a variety of data processing technologies, more detailed information can be obtained compared to traditional acquisition methods. Through multi-source data fusion, more comprehensive and accurate road network information can be obtained, thereby improving the accuracy of high-precision maps; (2) Compensation for missing road network elements and connection relationships, as well as reshooting and fusion during data anomaly processing, ensure that map data is consistent with actual road conditions, and the dynamic update mechanism can timely reflect changes in the road environment; (3) Reasonable road network element connection relationship construction and attribute setting provide a reliable basis for path planning and decision-making of autonomous driving vehicles, provide accurate lane connection relationships, speed limit information, etc., and ensure safe and efficient driving of vehicles; (4) Enhanced recognition of small target detection frames and hierarchical filtering of overlapping road network elements improve the quality and reliability of data and reduce the impact of erroneous information on maps; (5) Due to the use of efficient image acquisition and processing algorithms, the various element features in the road can be efficiently and accurately identified, and abnormal data can be compensated, making this method highly adaptable and scalable and applicable to different road scenes and autonomous driving systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 This is a principle block diagram of the method for generating and updating high-precision maps for the autonomous driving scenario library provided by the present invention. DETAILED DESCRIPTION

[0082] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0083] like Figure 1 As shown, the method for generating and updating a high-precision map of an autonomous driving scenario library provided by the present invention includes the following steps:

[0084] S1: Acquire high-altitude image video and perform frame extraction processing, perform element feature recognition on the extracted image, and extract the road network element features within the target detection frame; convert the image coordinate system into the actual map coordinate system; enhance the recognition of small target detection frames with an area smaller than a preset threshold, and perform hierarchical filtering on overlapping road network elements to retain valid target detection frames.

[0085] Aerial video frame extraction processing uses drones or satellites to acquire continuous aerial aerial imagery. Fixed-time sampling or keyframe extraction algorithms (such as those based on scene change detection) are used to extract video frames, balancing data volume and processing efficiency. Element feature recognition and extraction utilizes convolutional neural networks (such as a road network element recognition operator model trained on YOLO5) to detect objects such as road markings, traffic signs, and lane boundaries. This generates rectangular object detection boxes with class labels and extracts feature vectors such as texture, color, and geometry. The result data output by the parsing operator contains the bounding box information of the detected road network elements, namely the coordinates of the top-left and bottom-right corners of the bounding box, the corresponding class label, and information such as motorway lanes, intersections, and white dashed lines. This data is serialized into a list, where each element contains the coordinates and class information of the corresponding road network element.

[0086] The two-dimensional pixel coordinate system of the camera image (with the origin in the upper left corner of the image) is converted to the map coordinate system through coordinate system transformation. Lens distortion correction and georeferencing parameters are also taken into account to establish a mapping relationship between the pixel points and the Earth's ellipsoid. The coordinates of the camera on the actual map and the shooting angle are obtained. For each vertex coordinate of the detection frame, its coordinates in the image coordinate system are first subtracted from the image center coordinates to complete the translation operation, moving the coordinate origin to the center of the image. Then, trigonometric functions are used to calculate the rotation based on the angle of the camera. Finally, the rotated coordinates are added to the coordinates of the camera on the actual map to complete the transformation from the image coordinate system to the actual map coordinate system, thereby obtaining the position of the road network element on the actual map.

[0087] The small target enhancement recognition strategy uses super-resolution reconstruction technology to improve recognition rates for small targets such as signs resulting from long-distance photography. It also incorporates an attention mechanism to enhance local feature extraction. The recognition result list is traversed, and the area of each detection box is calculated by subtracting the upper left corner's horizontal coordinate from the lower right corner's horizontal coordinate, and then multiplying the result by the lower right corner's vertical coordinate minus the upper left corner's vertical coordinate. Detection boxes with an area smaller than a preset threshold may correspond to small road network elements, such as small traffic signs. The image regions corresponding to these small targets are extracted and enlarged using the bicubic interpolation algorithm to increase the number of pixels and enhance the clarity of the small targets. This enlarged region is then further enhanced, for example by applying histogram equalization to increase contrast. After this processing is complete, the enhanced image region is re-entered into the recognition operator for recognition, updating the recognition result list and improving the recognition accuracy of small targets. For any two detection boxes in the recognition result list, their intersection area is calculated. During the calculation, the larger of the two minimum horizontal coordinate values and the smaller of the maximum horizontal coordinate values of the two detection frames are taken, and the two are subtracted to obtain the horizontal width of the intersection area. Similarly, the larger of the minimum vertical coordinate values and the smaller of the maximum vertical coordinate values are taken, and the vertical height of the intersection area is subtracted. The horizontal width and vertical height are multiplied together to obtain the intersection area. The union area is calculated by adding the areas of the two detection frames and subtracting the intersection area. The intersection-in-union ratio (IOR) is then calculated: the intersection area divided by the union area. If the IOR exceeds a set threshold, the two detection frames may overlap. In this case, correct detections are retained based on factors such as element category priority (e.g., traffic signs take precedence over lane markings) and area size, while duplicate or incorrect detection frames are removed to ensure accurate recognition results. The hierarchical filtering mechanism sets conflict resolution rules based on the semantic priority of traffic elements (e.g., traffic signs > lane markings > intersections). When detection frames overlap, the higher-priority element is prioritized based on a weighted calculation to ensure the integrity of critical information.

[0088] S2: Constructing the connection relationship between road network elements based on the attributes and relevance of each road network element, including the following processes:

[0089] Lane connectivity is determined based on the relative positions, directional consistency, and traffic markings of two lanes. Topological connections are established by calculating the directional continuity (small angle between direction vectors) and physical connectivity (small spacing) of the lane centerlines, taking into account lane line type (e.g., dashed lines allow lane changes). For any two identified lanes, a sequence of coordinate points along their centerlines is obtained. The distance between each pair of points on these two centerlines is calculated by summing the squared difference between the horizontal and vertical coordinates of the two points and taking the square root. The minimum distance among all point pairs is then found. If the minimum distance is less than a preset distance threshold, the two lanes are considered potentially connected, as a close distance indicates spatial proximity and potential connection. The slope of the lane centerline is calculated by dividing the vertical coordinate difference between the starting and ending points of the lane centerlines by the horizontal coordinate difference, which is then used to determine the lane direction. For two potentially connected lanes, the angle between their directions is calculated using the inverse tangent function based on their slopes. If the angle is less than a preset angle threshold, this further supports the connection, as lanes with similar directions are more likely to be connected. For adjacent lanes that may be connected, the traffic marking type between them is checked. If the traffic marking is a dotted line, it usually means that the lanes are connected, as dotted lines indicate crossing is allowed. If it is a solid line, it means crossing is prohibited, so it is determined that there is no direct connection.

[0090] The lane-intersection connection is determined based on the lane direction and the orientation of the intersection traffic signs. The corresponding relationship between lane arrows and intersection direction signs is analyzed to establish a mapping of lane and intersection access permissions. For example, a left-turn lane can only be connected to a left-turn arrow. The intersection center coordinates and the coordinates of the lane centerline endpoint near the intersection are determined. A vector pointing from this endpoint to the intersection center is calculated by subtracting the lane centerline endpoint coordinates from the intersection center coordinates. Based on the direction of the vector, whether the lane is entering the intersection (the vector points toward the intersection center) or exiting the intersection (the vector points away from the intersection center) is determined, thereby determining the direction of the lane-intersection connection. Traffic signs at the intersection are traversed, such as directional arrow signs. The direction of the arrow indicates the lane's direction of travel at the intersection.

[0091] Traffic marking-lane binding is established based on the principle of minimum distance matching between traffic markings and lanes, as well as the functional rules of traffic markings. A nearest neighbor search algorithm is used to match lane markings with adjacent lanes, and spatial and functional associations are established based on the marking type (for example, guide lines must be bound to specific turning lanes). The coordinates of the lane boundary and traffic marking are obtained, and the minimum horizontal and vertical distances between the traffic marking and lane boundary are calculated. Horizontally, the absolute difference between the horizontal coordinates of the traffic marking and the lane boundary is calculated, and the minimum value is taken. The same process is repeated vertically. If the distance is less than the set distance threshold, the marking and lane are considered to be positionally associated, as a close distance indicates a close relationship. Traffic marking type and function are then matched with lane driving rules. For example, a white dashed line separates lanes traveling in the same direction. If a white dashed line is detected between two lanes traveling in the same direction, it is considered to be functionally associated with both lanes. A solid yellow line separates lanes traveling in opposite directions. If a solid yellow line is detected between two lanes traveling in opposite directions, it is considered to be functionally associated with both lanes.

[0092] Based on spatial proximity and the matching of sign semantic rules with lane attributes, the traffic sign-lane mapping relationship is determined. Traffic sign and lane coordinates are obtained, and the horizontal and vertical distances between the traffic sign and the lane are calculated. Horizontally, the absolute value of the difference between the traffic sign's abscissa and the lane's abscissa is calculated; vertically, the absolute value of the difference between the ordinate and the abscissa is calculated. If both distances are less than a set distance threshold, the traffic sign and lane are considered to have a positional correspondence, indicating that the traffic sign is designated for that lane. The meaning and rules of the traffic sign are then matched with the lane's driving rules. For example, if an arrow sign indicates the lane's driving direction, if the arrow's direction aligns with the lane's established driving direction, the sign is considered to match the lane's driving rules. A speed limit sign specifies the lane's speed limit. If the speed limit on the sign is within a reasonable range with the speed limit inferred for that lane based on other factors (such as road type), the sign is considered to match the lane's driving rules. For example, urban arterial roads typically have higher speed limits. If the speed limit on the speed limit sign falls within the speed limit range for that urban arterial road, the two are considered to match.

[0093] S3: Identify and set attributes for each road network element, including lane type, lane direction, and lane speed limit information. By scanning and identifying speed limit sign values and combining lane marking type (solid / dashed), the lane function (bus lane, tidal lane) is determined. GPS trajectory data can also be integrated to calculate actual lane speeds. In the actual map coordinate system, the lane width is calculated by calculating the sum of the squared difference between the horizontal and vertical coordinates of the two diagonal vertex coordinates based on the lane boundary coordinates. The square root is then taken to determine the lane width. This width is compared with the standard width ranges for different lane types, such as 3-3.75 meters for motor vehicle lanes and 1.5-2.5 meters for non-motor vehicle lanes. Lane type is determined by comparison. Lanes are checked for specific signs, such as bicycle signs on non-motor vehicle lanes and pedestrian signs on sidewalks. If a bicycle sign is detected, the lane is classified as a non-motor vehicle lane; if a pedestrian sign is detected, the lane is classified as a sidewalk. Analyze the width relationship between a lane and its adjacent lanes. If a lane is smaller than its adjacent lane and its width is within the non-motorized vehicle lane width threshold, and the traffic markings between the two lanes meet the rules for separating non-motorized vehicle lanes from motor vehicle lanes, such as a solid white line in the middle, the lane is determined to be a non-motorized vehicle lane.

[0094] Lane direction can be set based on arrow signs: the direction indicated by the arrow is determined as the lane's direction of travel. It can also be based on intersection connections: if the lane is connected to an intersection's entrance, the lane points toward the intersection; if it is connected to an exit, the lane points away from the intersection.

[0095] Lane speed limit setting can be referenced by speed limit sign recognition: Find the detection box corresponding to the speed limit sign category label from the recognition results and obtain the speed limit value within the detection box. Road type can also be determined: By analyzing factors such as the road's location (e.g., city center, suburbs), surrounding building types (e.g., commercial district, residential district), and connections to other roads, road types (e.g., urban arterial, secondary road) can be determined. For example, a road located in the city center, connecting important commercial and residential areas, with wide roads and high traffic volume, may be classified as an urban arterial. Furthermore, analysis can be further performed based on road segment characteristics, including the presence of curves. For curves, an adjusted speed limit is calculated based on the curve radius and vehicle safety standards. Generally speaking, the smaller the curve radius, the lower the speed limit. For example, empirical formulas indicate that the speed limit may be inversely proportional to the square root of the curve radius. For steep slopes, the adjusted speed limit is calculated based on the slope gradient and vehicle safety standards. The steeper the slope, the lower the speed limit. Experimental data shows that the speed limit decreases by a certain percentage with each increase in slope gradient.

[0096] S4: Compensate for missing road network elements and connections, including:

[0097] Compensate for missing lane elements based on adjacent lane information, intersection connectivity, and traffic markings;

[0098] Compensation based on adjacent lane information: When lane elements are incomplete in the image, the direction and length of the adjacent lane are obtained. If the adjacent lane is straight and the direction is clear, the incomplete lane is extended in the direction of the adjacent lane. For example, if the adjacent lane is a straight horizontal line pointing to the right, the incomplete lane is also extended horizontally to the right. Based on the length of the adjacent lane, the length of the missing portion of the incomplete lane is reasonably estimated. For example, if the adjacent lane is 50 meters long and the ratio of the incomplete lane to the adjacent lane can be determined in the identified part, the length of the missing portion is estimated based on this ratio.

[0099] Compensation based on intersection connectivity: Combining the identified lane-intersection connectivity, if a lane is known to be connected to an entrance lane at the intersection and is partially missing in the image, the missing lane's connection method and extension direction at the intersection are determined based on the intersection structure and the conditions of other related lanes. For example, if the intersection is a crossroads and a lane is known to enter the intersection from the east and is partially missing, the correct connection method and extension direction of the missing lane at the intersection are determined based on the directions of other lanes and the intersection's design rules.

[0100] Compensation based on traffic markings: Observe the traffic markings surrounding the incomplete lane, such as lane dividing lines and edge lines. If the lane dividing line suddenly interrupts the incomplete lane, the lane boundary can be determined by reasonably estimating its extension position based on its previous direction and the conditions of adjacent lanes and following the rules of line continuity. For example, if the lane dividing line was previously straight and parallel to the adjacent lane dividing line, its extension position can be estimated based on this parallel relationship.

[0101] Compensate for missing traffic marking elements based on identified traffic markings and their continuity principles and setting rules;

[0102] Compensation based on traffic sign continuity: If part of a traffic marking is missing, the system checks the surrounding identified traffic markings. For continuous markings, such as white dashed lines, the missing dashed lines are supplemented based on the spacing and position of the preceding and following dashed lines, following the principle of marking continuity. For example, if the spacing between the preceding and following dashed lines is 5 meters, the missing dashed lines are supplemented at a 5-meter interval.

[0103] Compensation based on traffic marking rules: Based on common traffic marking rules and the surrounding road network elements, the type and location of missing markings are inferred. For example, at intersections, guide lines are often set at specific locations to guide vehicle direction. If a guide line at that location is partially missing, the missing guide line is supplemented based on the guide line setting rules and information such as surrounding lanes and traffic signs. For example, guide lines are generally set at bends in intersections. Based on the shape of the intersection and the direction of other lanes, the correct position and shape of the guide line are determined and supplemented.

[0104] Ambiguous lane connections are reanalyzed, correcting any errors and compensating for any missing lane connections. When lane connections are uncertain due to missing elements or misidentification, the position, direction, and traffic markings of adjacent lanes are reanalyzed. The lane centerline continuity is rechecked. If the centerline is interrupted at a junction but can be reasonably connected, for example, if the direction and position of the centerlines on both sides of the interruption can be connected by simply extending or adjusting the angle, the connection is repaired. The correct connection is then determined based on information such as traffic markings.

[0105] For newly identified lanes, the system completes the connection between the new lane and the previously identified lanes. If a new lane element is discovered in a certain area, its position, direction, and traffic markings are analyzed relative to the surrounding lanes. Using lane connection relationship reasoning, the system determines the connection between the new and existing lanes and fills in any missing connections. For example, if a newly discovered lane is located close to an existing lane, the system analyzes their directions and the traffic markings between them to determine whether a connection exists and how.

[0106] The system can also compensate for lane-intersection connections, including correcting them. When a lane-intersection connection is incorrect or uncertain, the system reanalyzes the lane orientation, traffic signs, and markings surrounding the intersection. By observing traffic signs, markings, and vehicle trajectories at the intersection, the system can determine the correct lane-intersection connection and correct any errors. For example, if a lane-intersection connection direction was originally misjudged, it can be corrected by re-examining traffic signs and vehicle trajectories.

[0107] If a new lane element is discovered near an intersection, the system analyzes the intersection structure, the locations of new and old lanes, and traffic signs. Using lane-intersection connection reasoning, the system determines the connection between the newly discovered lane and the intersection, supplementing any missing connection information. For example, if a newly discovered lane is located in a corner of the intersection, the system analyzes the intersection structure and its connections to other lanes to determine the correct connection between the lane and the intersection.

[0108] S5: Semantize road network elements and configure element attributes, format the geometric expressions of road network elements in XML, and output them as high-precision map data.

[0109] Encodes semantic road network elements, attribute configurations, connection relationship topology, and geometric features into an OpenDRIVE-compliant XML format, outputting high-precision map data that includes lane-level geometry, speed limit attributes, and topological connections. Map elements include intersections, lanes, lane markings, road arrows, road facilities, and road signs.

[0110] Specific elements under the intersection category include: intersection, safety island, diversion island, traffic island, intersection center circle, intersection edge line, intersection guide line, and intersection mesh line.

[0111] Specific elements under the lane category include: driving lanes, central dividing strips, guide lines, turning areas, and green belts.

[0112] Specific elements under the lane marking category include: solid white line, dashed white line, solid yellow line, dashed yellow line, double solid white line, double dashed white line, double solid yellow line, double dashed yellow line, solid white dashed line, solid white dashed line, solid white dashed line, solid yellow dashed line, solid yellow dashed line, and stop line.

[0113] Specific elements under the road arrow category include: straight arrow, straight left turn arrow, straight right turn arrow, left turn arrow, right turn arrow, left U-turn, left U-turn, left merge arrow, right merge arrow.

[0114] Specific elements under the road facilities category include: pedestrian crossings, cones, parking spaces, construction signs, motor vehicles, traffic bollards, crash barriers, and water barriers.

[0115] Specific elements under the road sign category include: pedestrian waiting area, grid lines, speed reduction markings, longitudinal speed reduction markings, bicycle signs, pedestrian crossing warnings, text, slow down and give way (lines), slow down and give way (arrows), stop and give way (lines), stop and give way (arrows).

[0116] This method uses high-altitude imagery videos for frame extraction and element feature recognition, combined with multiple data processing technologies. Compared with traditional acquisition methods, it can obtain more detailed information. Through multi-source data fusion, it can obtain more comprehensive and accurate road network information, improving the accuracy of high-precision maps; it compensates for missing road network elements and connection relationships to ensure that map data is consistent with actual road conditions, and the dynamic update mechanism can promptly reflect changes in the road environment; reasonable road network element connection relationship construction and attribute setting provide a reliable basis for path planning and decision-making of autonomous driving vehicles, providing accurate lane connection relationships, speed limit information, etc., to ensure safe and efficient vehicle driving; enhanced recognition of small target detection frames and hierarchical filtering of overlapping road network elements improve the quality and reliability of data and reduce the impact of erroneous information on the map; due to the use of efficient image acquisition and processing algorithms, it can efficiently and accurately identify the various element features in the road and compensate for abnormal data, making this method highly adaptable and scalable and applicable to different road scenarios and autonomous driving systems.

[0117] In another technical solution, in step S1, converting the image coordinate system into the actual map coordinate system includes the following specific steps:

[0118] In the image coordinate system, take half of the image width and height pixel values to get the coordinates of the image center, and translate the vertex coordinates of the target detection box to the coordinate system with the image center as the origin;

[0119] Construct the rotation matrix R(θ) based on the device yaw angle θ:

[0120]

[0121] The coordinates after translation (x p ,y p ) to perform rotation transformation:

[0122]

[0123] The rotated coordinates (x r ,y r ) is superimposed on the actual geographic coordinates of the device to complete the spatial position mapping, and the result is corrected by Gauss-Krüger projection to eliminate the error of the earth curvature.

[0124] Coordinate translation quantization eliminates the odd-even pixel bias that can exist in traditional solutions. Using radians for angle parameters avoids the accumulation of angle conversion errors and improves the accuracy of rotational transformation calculations. Dynamically calculating spatial resolution values based on the geometric relationship between flight altitude and the camera's field of view allows for adaptability to diverse aerial photography scenarios compared to fixed resolution parameters. Quadratic and quartic corrections can be added to the traditional Gauss-Krüger projection linear terms, reducing coordinate errors from meters to centimeters. By establishing a rigorous mathematical symbol system and mapping physical meanings, coordinate transformations are made more precise.

[0125] In another technical solution, in step S1, enhancing the recognition of the small target detection frame specifically includes the following steps:

[0126] For target areas whose area is smaller than the preset pixel threshold, the bicubic interpolation algorithm is used to magnify them by a factor of 2:

[0127]

[0128] Among them, (u, v) is the pixel coordinate of the image after interpolation and amplification, I zoom (u, v) is the pixel value of the amplified image, I(x+i, y+j) is the pixel value of the original image; W(t) is the weight function of bicubic interpolation, and its expression is:

[0129]

[0130] Where a is the bicubic interpolation smoothing coefficient, corresponding to hour, ; corresponds to When ;

[0131] Adaptive contrast enhancement is performed on the enlarged area, the enlarged area is divided into 8×8 sub-blocks, the sub-blocks are histogram-cropped and the histogram grayscale increment is limited to less than or equal to 3;

[0132] The non-local means denoising algorithm is used to denoise the sub-blocks: the search window size is set to 21×21, the similar block size is set to 7×7, and the pixel weight is calculated by comparing similar blocks within the search window;

[0133] The element features of the enhanced image area are re-identified. When the intersection-over-union ratio of the new and old target detection frames exceeds 0.6, the confidence of the new and old target detection frames is calculated. If the confidence of the new target detection frame is at least 0.2 higher than that of the old target detection frame, the element list is updated with the element features of the new target detection frame.

[0134] Small object area thresholds are used to dynamically filter small objects, avoiding misjudgments caused by fixed size thresholds. A smoothing coefficient is introduced into the bicubic interpolation formula to achieve a balance between edge sharpening and noise suppression. Experiments have verified that the block size used in CLAHE contrast enhancement achieves the optimal balance between enhancement effect and computational overhead. Non-local means denoising uses a combination of a large search window and a small similarity window to improve texture preservation. A fusion threshold of IoU > 0.6 can be set to prevent interference from duplicate boxes, and a confidence difference threshold of 0.2 prevents low-quality enhancement results from overwriting existing correct detections.

[0135] In step S1, hierarchical filtering of overlapping road network elements includes the following specific steps:

[0136] Set the IoU threshold for traffic sign elements, lane line elements, and intersection elements respectively. When the IoU of the target detection box exceeds the corresponding IoU threshold, perform target detection box conflict analysis:

[0137] The conflict graph model is constructed with the target detection boxes as nodes and the overlapping conflict relationships between target detection boxes as edges. The node weights are calculated as follows:

[0138]

[0139] Among them, w i is the node weight of the i-th node, A i is the area of the i-th target detection box, α is the category priority coefficient, and β is the area influence coefficient; P r (C i ) is the category priority value, which represents the node category C of the i-th node i Importance priority;

[0140] The branch-and-bound algorithm is used to traverse the conflict graph model. During the traversal process, the current optimal solution of the node is solved and updated until the upper bound of the subproblem is less than the current optimal value, and the branch is terminated to obtain the maximum weighted independent set.

[0141] In the maximum weighted independent set, if there is a cross-category conflict between nodes, the target detection box with the larger category priority value in the cross-category conflict is retained; if there is a same-category conflict between nodes, the node weight w is retained. i In the first 10% of target detection frames, the filtered target detection frames are deleted to complete the hierarchical filtering of overlapping road network elements.

[0142] For traffic signs (high priority but prone to dense occurrence), a lower intersection-of-union ratio threshold (e.g., 0.4) is set to avoid missing duplicate frames; for lane markings (low priority but requiring continuous detection), a higher threshold (e.g., 0.6) is set to preserve reasonable overlap. A cross-class conflict retention rule is used to address the issue of class priority being ignored in traditional independent set algorithms. Edge weights reflect the severity of overlap and provide gradient information for node optimization. The weight function is designed to mitigate the advantage of large target areas and balance weight competition between large and small targets. The coefficients α and β are used to quantitatively integrate artificial rules (class priority) with data features (target size). The introduction of a cached subproblem upper bound in the branch-and-bound method can reduce repeated computation by approximately 40%. For conflict graphs with more than 100 nodes, a greedy pre-screening strategy can be enabled.

[0143] In another technical solution, in step S2, the lane-to-lane connection relationship is determined by:

[0144] The lane centerline endpoints are extracted to generate direction vectors. The angle between the two lane directions is calculated using the dot product of the vectors. If the angle is less than a predetermined threshold and the minimum centerline distance is less than a distance threshold, the directions are considered consistent.

[0145] A sliding window detection is performed on the gap area between adjacent lanes. The frequency of occurrence of white dashed lines and solid yellow lines is counted. A connection relationship is forced to be established between adjacent lanes where the frequency of white dashed lines is greater than 0.8, and a connection relationship is prohibited between adjacent lanes where the frequency of yellow solid lines is greater than 0.6.

[0146] Dynamically selecting endpoints and extension points avoids misjudgments caused by fixed sampling points, and interval design balances directional sensitivity and noise tolerance. The directional angle threshold is set to 15°, corresponding to the curvature tolerance of a typical curve radius (50 meters) on urban roads. The distance threshold can be set to 2 meters (57% of the standard lane width of 3.5 meters) to ensure physical lane connectivity. A frequency threshold of 0.8 / 0.6 in the sliding window confidence mechanism effectively filters out occasional detection errors. The window size is set to 1m×1m to match the standard size of traffic markings (dashed line segments are 2-4 meters long). Anti-interference design is also available. For ambiguous situations outside the frequency threshold, a multi-frame verification mechanism is activated (requiring three consecutive frames of detection to meet the standard). When the minimum lane distance is greater than 5 meters, it is directly determined to be unconnected, reducing computational effort.

[0147] In step S3, setting the lane speed limit information attribute includes the following steps:

[0148] Calculate the curvature radius R of the lane centerline and calculate the curve speed limit V based on the curvature radius of the lane centerline c :

[0149]

[0150] Where μ is the tire-road friction coefficient, g is the acceleration due to gravity, and θ c is the transverse slope angle;

[0151] If a speed limit sign value is detected, the lane speed limit information attribute is set to the smaller value of the curve speed limit value and the speed limit sign value; the uphill speed limit on steep slopes is reduced by 2km / h for every 1% slope gradient; and the speed transition is performed using linear interpolation within 50m before and after the speed limit change interval of the lane.

[0152] In the curve frame number, the moving least square method can be used to dynamically fit the curvature radius, which has higher accuracy than the fixed window method. By introducing the lateral slope compensation term tanθ c , which solves the defect of traditional formula ignoring ultra-high design.

[0153] Furthermore, when the calculated value is not greater than the speed limit sign value, the speed limit sign value and the curve model calculated value can be weighted and summed in a 6:4 ratio to obtain the final speed limit value. This respects traffic regulations while taking into account actual physical constraints, preventing abnormal model calculation values.

[0154] Optimization is carried out on steep slopes, using a step function (2km / h decay for every 1% slope) instead of linear decay, which is more in line with the expectations of autonomous driving. A hard lower limit of 20km / h avoids non-physical speed limits on extreme slopes; a smooth transition mechanism is used to match national standards, and linear interpolation ensures the continuity of speed limit changes to prevent the autonomous driving system from accidentally triggering emergency braking.

[0155] In another technical solution, in step S4, lane elements are compensated using the following method:

[0156] Set three control points P0, P1, and P2 in the identified part of the missing lane and calculate the directional derivative α of the adjacent lane:

[0157]

[0158] Among them, (x start ,y start ) is the starting coordinate of the adjacent lane, (x end ,y end ) are the coordinates of the end point of the adjacent lane;

[0159] Calculate the newly added control point P3:

[0160]

[0161] in, is the direction vector, ; k is the expansion coefficient, ;

[0162] Extend the missing lanes based on the cubic Bezier curve and construct the parametric equation:

[0163]

[0164] Set the maximum curvature radius limit of the curve. If the calculated curvature radius exceeds the limit, adjust the position of P3 to , ;

[0165] At the intersection, a uniform B-spline curve is generated with the number of connected lanes as the order, and the node vectors of the curve are distributed in equal proportions; the curvature continuity of the interrupted lane dividing lines is repaired according to the parabolic equation.

[0166] This compensation mechanism implements directional adaptation, dynamically adjusting the extension vector based on the directional angles of adjacent lanes to ensure smooth integration of the compensated lanes into the existing road network. The tan(α) term is introduced into the scaling factor k to automatically shorten the extension distance in sharp bends (where α is large) to avoid path oscillation. The number of connecting lanes n at an intersection determines the spline order, enabling adaptive spline order and accurate topological representation of complex intersections. Bezier curves can be discretized with a fixed step size of Δt=0.01 to balance accuracy and real-time performance. The B-spline basis functions are recursively calculated using the de Boer algorithm.

[0167] In step S4, the method of compensating the traffic marking elements includes the following steps:

[0168] Compensate for missing dashed line segments in traffic marking elements:

[0169] Perform Fourier transform on the solid line segments before and after the dashed line, extract the spectrum features and calculate the phase consistency. If the phase consistency is greater than 0.85, generate a new dashed line in the missing segment according to the standard dashed-to-solid ratio.

[0170] To compensate for missing flow lines:

[0171] The reference radius R is obtained based on the curvature radius R of the lane centerline b :If R≤50m, then the corresponding R b =R; otherwise R b =50+0.2(R-50); the involute equation in the polar coordinate system is superimposed on the sine wave coefficient to generate a natural transition trajectory, and the polar coordinate system is converted to the Cartesian coordinate system. The steering angle is discretely sampled to generate a continuous guide line for compensation.

[0172] The spectrum analysis uses a 256-pixel window adapted to the standard dashed line length (6 meters @ 0.05 meters / pixel) to ensure a spectrum resolution of ±0.1 Hz. The phase consistency threshold is set to 0.85 to filter random noise interference, which can reduce the false positive rate to <5%.

[0173] The original design period of the marking is inferred through the dominant frequency f to achieve dynamic period generation, solving the problem of dotted line deformation caused by zooming in and out of the image; a 2:3 virtual-to-real ratio is used to strictly comply with the national standard for lane dividing lines.

[0174] To achieve guide line deformation control, a radius compression strategy is adopted (sensitivity is reduced when R>50 meters) to avoid excessive distortion of guide lines in large-radius curves; three-cycle sinusoidal modulation is used to generate naturally transitioned diversion trajectories to adapt to complex scenarios such as three-way intersections; the Fourier transform uses the FFT algorithm to support real-time processing; the polar coordinate sampling interval of 0.01rad corresponds to a ground resolution of 0.1 meter, meeting the requirements of high-precision maps.

[0175] In another technical solution, in step S2, after determining the lane-intersection connection relationship, the intersection is also marked with a driving path priority, including the following steps:

[0176] Each entrance and exit lane at an intersection is assigned a unique serial number, and all legal driving paths are generated based on traffic regulations. Entrance and exit lanes, representing lanes entering and exiting the intersection, can use different numbering sequences, with each serial number corresponding to a lane shift. For example, entrance lanes are numbered 1 to M, and exit lanes are numbered 1 to N, but are distinguished by different symbols. An example of a legal driving path setup: a right-turn entrance lane can only connect to an exit lane on the right. If there are two exit lanes on the right, there are two paths for turning right from an entrance lane, which complies with traffic regulations. For entrance lanes that can be both straight and turning, there are more diverse driving paths and directions.

[0177] Count the historical driving times of each driving route, normalize the driving times of the same direction path, and generate the efficiency score E ij :

[0178]

[0179] Among them, C ij is the number of vehicles traveling from the i-th import lane to the j-th exit lane, C max is the maximum number of trips along the same direction;

[0180] For each driving path, calculate the conflict area A between it and the driving paths that can be driven at the same time c and the number of intersections N c , the conflict area and the number of intersections are integrated into the conflict index to generate the safety score S ij :

[0181]

[0182] Where σ is the exponential adjustment coefficient;

[0183] The comprehensive priority score of each driving path is obtained by weighting the efficiency score and safety score by weight coefficient:

[0184]

[0185] Among them, F ij is the comprehensive priority score of the driving path from the i-th import lane to the j-th exit lane, and γ is the normalized weight coefficient.

[0186] Historical driving counts represent a period of time and can be obtained through traffic systems or monitoring statistics, providing complete data for a representative period. A same-direction path is defined as one with identical or adjacent entry lanes and permission to travel in the same direction (e.g., multiple lanes for through traffic), and identical or adjacent exit lanes. For example, if there are three adjacent entry lanes (R1, R2, and R3) and exit lanes (L1, L2, and L3) for through traffic in the same direction, there are nine possible through-travel paths. Traffic flow in different lanes typically varies, which means that lane and route selection can affect intersection efficiency. Therefore, a lane efficiency score is calculated. Considering the collision risk associated with lane changes (e.g., from R1 to L2) and the risk of interference from irregular non-motorized vehicle traffic in motor vehicle lanes adjacent to non-motorized vehicle lanes, the safety of the selected lanes must be considered to calculate a safety score. The efficiency and safety of a driving path are mutually influential. For example, a more dangerous lane may have fewer vehicles, resulting in higher efficiency when it can be used normally. However, this efficiency may also be reduced due to maneuvers such as meeting or avoiding other vehicles. Therefore, a weighted calculation is performed on the efficiency and safety scores to produce a comprehensive score, providing more options for vehicles undergoing braking maneuvers. Choosing a route with a higher comprehensive score allows for a better balance between safety and efficiency. Choosing a route with a higher safety score but a lower efficiency score can maximize driving safety and avoid traffic accidents. Choosing a route with a higher efficiency score but a lower safety score is a crucial challenge for autonomous driving testing. The highly intelligent control methods of autonomous driving systems can respond quickly to various emergencies, seeking to achieve the most efficient driving while also balancing driving safety. By marking the scores of different routes at intersections on high-precision maps, autonomous vehicles can select appropriate routes based on actual needs and their own intelligent control levels, without relying on the vehicle system's real-time calculations and judgments.

[0187] In another technical solution, in step S4, data anomaly processing is also performed in the following manner:

[0188] When the number of compensations for an area exceeds three times or the final confidence level of the area is lower than 0.4, the boundary of the area is extracted and a spiral search path is generated to control the drone to retake the photo. The search path radius is expanded from 10 meters to 50 meters in steps of 5 meters, and the search angle is increased by 30 degrees. The newly collected data is fused with the original data through Kalman filtering. The state equation of Kalman filter fusion adopts a uniform velocity model to construct the observation equation, and the observation matrix is used to directly map the observation position coordinates.

[0189] A spiral search path is generated for optimized coverage. An angular step size of Δθ = π / 6 ensures 30° heading adjustments and 85% overlap for complete blind spot coverage. Dynamic radius expansion (Δρ = 5 meters) accommodates a maximum search radius of 50 meters, balancing efficiency and completeness. Earth curvature correction is used to map the flat spiral to a spherical trajectory, achieving a longitude conversion error of < 0.0001°, ensuring precise mapping of geographic coordinates.

[0190] Using the Kalman filter dual model, the process model is used for the uniform motion assumption, which is suitable for the stable flight state of the UAV; the observation model performs direct position observation, matches GPS / RTK positioning data, and can use adaptive noise covariance to achieve multi-source data fusion.

[0191] It should be noted that although the steps are described above in a specific order, this does not necessarily mean that the steps must be performed in this specific order. In fact, some of these steps can be performed concurrently or even in a different order, as long as the required functions can be achieved. The number of devices and processing scales described here are intended to simplify the description of the present invention. Applications, modifications, and variations of the present invention will be apparent to those skilled in the art.

[0192] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for generating and updating a high-precision map of an autonomous driving scenario library, characterized in that: The following steps are involved: S1: Acquire high-altitude video and perform frame extraction, set a target detection frame on the extracted image, and extract the features of each road network element within the target detection frame, including element attributes and relevance; S2: Constructing the connection relationship of road network elements based on the element attributes and associations of each road network element, including the following processes: Determine the lane-to-lane connection based on the relative position, direction consistency, and traffic markings of the two lanes; Determine the lane-intersection connection relationship based on the lane direction and the direction of the intersection traffic sign; Establish a binding relationship between traffic markings and lanes based on the minimum distance matching principle between traffic markings and lanes and the functional rules of traffic markings; Determine the traffic sign-lane mapping relationship based on spatial proximity and matching of sign semantic rules with lane attributes; S3: Verify the integrity of the connection relationship between road network elements and compensate for missing road network elements and connection relationships, including: Compensate for missing lane elements based on adjacent lane information, intersection connectivity, and traffic markings; Compensate for missing traffic marking elements based on identified traffic markings and their continuity principles and setting rules; Reanalyzes ambiguous lane connections, repairs incorrect lane connections, and compensates for missing lane connections. For newly identified lanes, supplements the connection between the new lane and the already identified lanes. S4: semanticize the compensated road network elements and connection relationships and configure element and relationship attributes. Format the geometric expressions of the elements and relationship attributes in XML and output them as high-precision map data. In step S1, the overlap of road network elements is also detected, and the overlapping road network elements are hierarchically filtered to retain valid target detection frames, wherein the hierarchical filtering of overlapping road network elements includes the following specific steps: Set the IoU threshold for traffic sign elements, lane line elements, and intersection elements respectively. When the IoU of the target detection box exceeds the corresponding IoU threshold, perform target detection box conflict analysis: The conflict graph model is constructed with the target detection boxes as nodes and the overlapping conflict relationships between target detection boxes as edges. The node weights are calculated as follows: Among them, w i is the node weight of the i-th node, A i is the area of the i-th target detection box, α is the category priority coefficient, and β is the area influence coefficient; P r (C i ) is the category priority value, which represents the node category C of the i-th node i Importance priority; The branch-and-bound algorithm is used to traverse the conflict graph model. During the traversal process, the current optimal solution of the node is solved and updated until the upper bound of the subproblem is less than the current optimal value, and the branch is terminated to obtain the maximum weighted independent set. In the maximum weighted independent set, if there is a cross-category conflict between nodes, the target detection box with the larger category priority value in the cross-category conflict is retained; if there is a same-category conflict between nodes, the node weight w is retained. i In the first 10% of target detection frames, delete the target detection frames that are not selected for saving and complete the hierarchical filtering of overlapping road network elements.

2. The method for generating and updating a high-precision map of an autonomous driving scenario library according to claim 1, characterized in that: In step S1, after setting the target detection frame for the image obtained by frame extraction, the step of converting the image coordinate system into the actual map coordinate system is also included, which includes the following specific steps: In the image coordinate system, take half of the image width and height pixel values to get the coordinates of the image center, and translate the vertex coordinates of the target detection box to the coordinate system with the image center as the origin; Construct the rotation matrix R(θ) based on the device yaw angle θ: The coordinates after translation (x p ,y p ) to perform rotation transformation: The rotated coordinates (x r ,y r ) is superimposed on the actual geographic coordinates of the device to complete the spatial position mapping, and the result is corrected by Gauss-Krüger projection to eliminate the error of the earth curvature.

3. The method for generating and updating a high-precision map of an autonomous driving scenario library according to claim 1, characterized in that: Step S1 also includes enhancing the recognition of small target detection frames whose area is smaller than a preset threshold, which specifically includes the following steps: For target areas whose area is smaller than the preset pixel threshold, the bicubic interpolation algorithm is used to magnify them by a factor of 2: Among them, (u, v) is the pixel coordinate of the image after interpolation and amplification, I zoom (u, v) is the pixel value of the amplified image, I(x+i, y+j) is the pixel value of the original image; W(t) is the weight function of bicubic interpolation, and its expression is: Where a is the bicubic interpolation smoothing coefficient, corresponding to hour, ; corresponds to When ; Adaptive contrast enhancement is performed on the enlarged area, the enlarged area is divided into 8×8 sub-blocks, the sub-blocks are histogram-cropped and the histogram grayscale increment is limited to less than or equal to 3; The non-local means denoising algorithm is used to denoise the sub-blocks: the search window size is set to 21×21, the similar block size is set to 7×7, and the pixel weight is calculated by comparing similar blocks within the search window; The element features of the enhanced image area are re-identified. When the intersection-over-union ratio of the new and old target detection frames exceeds 0.6, the confidence of the new and old target detection frames is calculated. If the confidence of the new target detection frame is at least 0.2 higher than that of the old target detection frame, the element list is updated with the element features of the new target detection frame.

4. The method for generating and updating a high-precision map of an autonomous driving scenario library according to claim 1, characterized in that: In step S2, the lane-to-lane connection relationship is determined by: The lane centerline endpoints are extracted to generate direction vectors. The angle between the two lane directions is calculated using the dot product of the vectors. If the angle is less than a predetermined threshold and the minimum centerline distance is less than a distance threshold, the directions are considered consistent. A sliding window detection is performed on the gap area between adjacent lanes. The frequency of occurrence of white dashed lines and solid yellow lines is counted. A connection relationship is forced to be established between adjacent lanes where the frequency of white dashed lines is greater than 0.8, and a connection relationship is prohibited between adjacent lanes where the frequency of yellow solid lines is greater than 0.

6.

5. The method for generating and updating a high-precision map of an autonomous driving scenario library according to claim 1, characterized in that: In step S1, the element attributes include lane type, lane direction, and lane speed limit information. Identifying and setting the lane speed limit information attributes includes the following steps: Calculate the curvature radius R of the lane centerline and calculate the curve speed limit V based on the curvature radius of the lane centerline c : Where μ is the tire-road friction coefficient, g is the acceleration due to gravity, and θ c is the transverse slope angle; If a speed limit sign value is detected, the lane speed limit information attribute is set to the smaller value of the curve speed limit value and the speed limit sign value; the uphill speed limit on steep slopes is reduced by 2km / h for every 1% slope gradient; and the speed transition is performed using linear interpolation within 50m before and after the speed limit change interval of the lane.

6. The method for generating and updating a high-precision map of an autonomous driving scenario library according to claim 1, characterized in that: In step S3, lane elements are compensated using the following method: Set three control points P0, P1, and P2 in the identified part of the missing lane and calculate the directional derivative α of the adjacent lane: Among them, (x start ,y start ) is the starting coordinate of the adjacent lane, (x end ,y end ) are the coordinates of the end point of the adjacent lane; Calculate the newly added control point P3: in, is the direction vector, ; k is the expansion coefficient, ; Extend the missing lanes based on the cubic Bezier curve and construct the parametric equation: Set the maximum curvature radius limit of the curve. If the calculated curvature radius exceeds the limit, adjust the position of P3 to , ; At the intersection, a uniform B-spline curve is generated with the number of connected lanes as the order, and the node vectors of the curve are distributed in equal proportions; the curvature continuity of the interrupted lane dividing lines is repaired according to the parabolic equation.

7. The method for generating and updating a high-precision map of an autonomous driving scenario library according to claim 1, wherein: In step S3, the method of compensating the traffic marking elements includes the following steps: Compensate for missing dashed line segments in traffic marking elements: Perform Fourier transform on the solid line segments before and after the dashed line, extract the spectrum features and calculate the phase consistency. If the phase consistency is greater than 0.85, generate a new dashed line in the missing segment according to the standard dashed-to-solid ratio. To compensate for missing flow lines: The reference radius R is obtained based on the curvature radius R of the lane centerline b :If R≤50m, then the corresponding R b =R; otherwise R b =50+0.2(R-50); the involute equation in the polar coordinate system is superimposed on the sine wave coefficient to generate a natural transition trajectory, and the polar coordinate system is converted to the Cartesian coordinate system. The steering angle is discretely sampled to generate a continuous guide line for compensation.

8. The method for generating and updating a high-precision map of an autonomous driving scenario library according to claim 1, wherein: In step S2, after determining the lane-intersection connection relationship, the intersection is also marked with a driving path priority, including the following steps: Assign unique serial numbers to each entrance and exit lane of the intersection and generate all legal driving paths based on traffic rules; Count the historical driving times of each driving route, normalize the driving times of the same direction path, and generate the efficiency score E ij : Among them, C ij is the number of vehicles traveling from the i-th import lane to the j-th exit lane, C max is the maximum number of trips along the same direction; For each driving path, calculate the conflict area A between it and the driving paths that can be driven at the same time c and the number of intersections N c , the conflict area and the number of intersections are integrated into the conflict index to generate the safety score S ij : Where σ is the exponential adjustment coefficient; The comprehensive priority score of each driving path is obtained by weighting the efficiency score and safety score by weight coefficient: Among them, F ij is the comprehensive priority score of the driving path from the i-th import lane to the j-th exit lane, and γ is the normalized weight coefficient.

9. The method for generating and updating a high-precision map of an autonomous driving scenario library according to claim 1, characterized in that: In step S3, data anomaly processing is also performed in the following manner: When the number of compensations for an area exceeds three times or the final confidence level of the area is lower than 0.4, the boundary of the area is extracted and a spiral search path is generated to control the drone to retake the photo. The search path radius is expanded from 10 meters to 50 meters in steps of 5 meters, and the search angle is increased by 30 degrees. The newly collected data is fused with the original data through Kalman filtering. The state equation of Kalman filter fusion adopts a uniform velocity model to construct the observation equation, and the observation matrix is used to directly map the observation position coordinates.

Images