A method for real-time update of local maps

Abstract

The present application discloses a method and system for real-time update of local maps, which relates to map update. Point cloud data during vehicle driving is acquired; the point cloud data is filtered according to the vehicle speed and vehicle height; according to the filtered point cloud data, road boundaries are extracted through a deep learning network and geometric algorithms to obtain a road boundary point set; the current lane boundary of the vehicle in the map is acquired; according to the road boundary point set and the current lane boundary, a boundary error is calculated; when the boundary error is greater than a preset threshold, the current lane centerline is reconstructed according to the road boundary point set, and the local map is updated; the local map represents a map including the current lane boundary and the current lane centerline. Aiming at the long update period of traditional high-precision maps in the weak communication environment of underground coal mines, the present application adopts a weighted adaptive algorithm to reconstruct the lane centerline and only performs local update when the boundary error exceeds the threshold, improving the map update efficiency.

Description

Technical Field

[0001] This application relates to the field of image data processing, and particularly to a method for real-time updating of a local map. Background Art

[0002] As a key infrastructure for autonomous driving and driverless systems, high-precision maps provide crucial support for the precise positioning, path planning, and safe operation of vehicles. In special industrial environments such as underground mines, the application of high-precision maps is more challenging and necessary. Underground mine areas usually have characteristics such as complex terrain, changing environments, and limited communication conditions, and conventional map updating methods are difficult to meet the actual needs. Especially in a weak communication environment, the traditional cloud map updating mode is severely restricted, and it is impossible to transmit a large amount of point cloud data and map information in a timely manner, resulting in a mismatch between the map data and the actual road environment, thereby affecting vehicle safety and operation efficiency.

[0003] With the development of the trend of mine intelligence and unmanned operation, the transport vehicles and operating equipment in underground mine areas are gradually evolving towards automation, posing higher requirements for the timeliness and accuracy of map data. At the same time, the road conditions in the mining area will change frequently due to reasons such as mining activities, equipment movement, and temporary obstacles. If these changes cannot be reflected in the high-precision map in a timely manner, it will lead to decision-making errors in the autonomous driving system and even cause safety accidents.

[0004] Traditional high-precision map updating technologies mainly rely on professional surveying vehicles to regularly collect data, and generate maps through post-processing, modeling, and other processes. The update cycle is usually in units of weeks or months. This method is acceptable in an open environment with good communication conditions, but it has obvious deficiencies in the weak communication environment of underground mines: on the one hand, it is difficult for professional surveying vehicles to enter the operation area frequently; on the other hand, the limited communication bandwidth cannot support the real-time transmission and processing of large-scale point cloud data.

[0005] For example, the related patent document CN118857271B discloses a method for making a mining area map based on a user-defined mapping tool, which relates to the field of map making, including: obtaining the terrain and vehicle specification parameters of the mining area to generate a road network model of the mining area; then using on-vehicle sensors to collect vehicle driving trajectories, slope information, and point cloud data, and performing data preprocessing; then formulating a mapping standard including the types, expression methods, and usage scenarios of mining area map elements according to the collected data; finally, based on the mapping standard, extracting mining area map elements through a user-defined mapping tool to generate a mining area map that meets the usage requirements of driverless vehicles. However, this solution has the following defects:

[0006] First, this solution still adopts the traditional batch processing mode, which cannot achieve real-time map updates. The update cycle is long, making it difficult to cope with the frequent changes in the road environment of underground coal mines. Second, this solution does not consider the particularity of the weak communication environment in underground coal mines. A large amount of point cloud data needs to be transmitted to the processing center, and due to bandwidth limitations, its practical application is hindered. Summary of the Invention

[0007] Aiming at the long update cycle of traditional high-precision maps in the weak communication environment of underground coal mines in the prior art, this application provides a method and system for real-time update of local maps. The lane centerline is reconstructed using a weighted adaptive algorithm, and only local updates are performed when the boundary error exceeds the threshold, improving the map update efficiency.

[0008] One aspect of this application provides a method for real-time update of local maps, including: obtaining point cloud data during vehicle driving; filtering the point cloud data according to the vehicle speed and vehicle height; extracting road boundaries from the filtered point cloud data through a deep learning network and geometric algorithms to obtain a set of road boundary points; obtaining the current lane boundary of the vehicle in the map; calculating the boundary error according to the set of road boundary points and the current lane boundary; when the boundary error is greater than a preset threshold, reconstructing the current lane centerline according to the set of road boundary points and updating the local map; the local map represents a map including the current lane boundary and the current lane centerline.

[0009] Further, filtering the point cloud data according to the vehicle speed and vehicle height includes: obtaining the current position coordinates of the vehicle and the current vehicle speed v; calculating the maximum effective distance threshold , where k is a preset adjustment coefficient; traversing each point in the point cloud data and calculating the distance between the point and the current position coordinates of the vehicle; removing the points whose distance is greater than the maximum effective distance threshold , and removing the points in the point cloud data whose z coordinate value is greater than the vehicle height to obtain the filtered point cloud data.

[0010] Further, obtaining the set of road boundary points includes: using the pre-trained PV-RCNN deep learning network to extract the filtered point cloud data to obtain a candidate road boundary area; generating a three-dimensional bounding box according to the candidate road boundary area; using geometric algorithms to correct the three-dimensional bounding box to obtain an optimal boundary line; classifying boundary points according to the optimal boundary line to obtain a set of road boundary points; among them, deep learning provides preliminary recognition ability, and geometric algorithms provide precise positioning and correction. This fusion architecture enables boundary recognition to maintain a high accuracy rate under harsh conditions such as complex lighting, dust, and water accumulation in underground coal mines.

[0011] Extract the filtered point cloud data using the pre-trained PV-RCNN deep learning network to obtain candidate road boundary regions, including: Making a dataset: Label the historical point clouds collected. Each labeled data contains , where is the starting point of the boundary, are the length, width, and height of the 3D box respectively, is the lane direction; Data augmentation: Randomly rotate and translate the point clouds and labeled data for data augmentation; Model training: Modify the size of the PV-RCNN anchor boxes to fit the lane boundary boxes, adjust the anchor box size to [5, 0.2, 0.1], and load the network weight file to initialize the network. Start model training. When the training loss converges, stop training to obtain the trained weight file; Load the point clouds and the trained weight file, and output to obtain the 3D boundary boxes containing the boundary points.

[0012] Further, obtain the optimal boundary line, including: Randomly select multiple points from the point clouds of each 3D boundary box to fit the initial straight line; Calculate the distance from each point to the initial straight line. When the distance is less than the preset threshold, mark the corresponding point as an inlier; Calculate the remaining number of iterations N according to the inlier ratio w; After each iteration, calculate the inlier ratio , and update the remaining number of iterations according to the inlier ratio ; When reaching the preset upper limit of the number of iterations or the change in the inlier ratio w is less than the threshold within the preset number of consecutive rounds, stop the iteration, and select the straight line with the most inliers as the optimal boundary line.

[0013] Further, calculate the remaining number of iterations N according to the inlier ratio w. The calculation formula is: , where p represents the expected probability of randomly sampling and selecting all inliers; n is the minimum number of points required to fit the straight line.

[0014] Further, according to the optimal boundary line, perform boundary point classification to obtain the road boundary point set, including: Obtain the line segment of the optimal boundary line within the 3D boundary box as the valid line segment; Generate a boundary point at a preset distance interval according to the valid line segment to obtain the boundary point set ; Obtain the current coordinate point of the vehicle and the vehicle driving direction vector V; For each boundary point , calculate the vector pointing from the current coordinate point of the vehicle to the boundary point , and the cross product result with the vector V; When the cross product result is greater than zero, classify the corresponding boundary point into the left boundary point set , when the cross product result When it is less than or equal to zero, the corresponding boundary point is classified into the right boundary point set ; Combine the left boundary point set and the right boundary point set to form a complete set of road boundary points.

[0015] Furthermore, obtain the current lane boundary of the vehicle in the map, including: obtaining the lane data of the map; storing the lane data using the R-Tree data structure; starting from the root node of the R-Tree data structure, traversing the R-Tree data structure to obtain the minimum bounding rectangle MBR of each node, where the lower left corner coordinates of the minimum bounding rectangle MBR , and the upper right corner coordinates ; Determine the current coordinate point of the vehicle , and determine whether it satisfies and . If it satisfies, then determine that the current coordinate point C of the vehicle is located in the current MBR; traverse each lane in the MBR where the vehicle is located, and calculate the sum of the angles formed by connecting the current coordinate point of the vehicle to each point on the lane boundary. If the sum of the angles is 360°, it is determined that the vehicle is in the current lane, and obtain the left lane boundary and the right lane boundary of the lane as the current lane boundary.

[0016] Furthermore, according to the set of road boundary points and the current lane boundary, calculate the boundary error, including: calculating the distance from each point in the left boundary point set to the left lane boundary to obtain the left boundary point distance set ; Calculate the distance from each point in the right boundary point set to the right lane boundary to obtain the right boundary point distance set ; Calculate the average value of the distances in the left boundary point distance set and the right boundary point distance set ; According to the average value of each distance , calculate the total boundary error error: .

[0017] Furthermore, when the boundary error is greater than the preset threshold, reconstruct the current lane centerline according to the set of road boundary points and update the local map, including: according to the left boundary point set and the right boundary point set , along their respective boundary line directions, resample according to the segmentation distance , where is the segmentation adjustment coefficient; obtain the resampled left boundary point set and the resampled right boundary point set ; According to the left boundary point distance set ​, calculate the average left boundary error ; According to the right boundary point distance set , calculate the average right boundary error ; According to the average left boundary error and the average right boundary error , set the boundary credibility weight coefficients and ; According to the vehicle driving direction vector V, for the resampled and points, establish the corresponding relationship between the left and right boundary points to form multiple horizontal connection line segments; for each horizontal connection line segment, calculate the weighted center point , where and are the coordinates of the left and right endpoints of the corresponding line segment respectively; in the vehicle driving direction, sequentially connect all the weighted center points to form an initial center point sequence; perform smoothing processing on the initial center point sequence to generate a reconstructed lane center line; update the local map according to the reconstructed lane center line and the obtained current lane boundary.

[0018] Another aspect of the present application also provides a local map real-time update system for executing a local map real-time update method of the present application.

[0019] Compared with the prior art, the advantages of the present application are as follows:

[0020] (1) In the weak communication environment of underground mines, due to limited bandwidth and unstable data transmission, it is difficult to process point cloud data. The prior art generally uses a fixed distance threshold for point cloud filtering, but there are defects such as the filtering effect not being dynamically adjusted according to the driving state and waste of computing resources. The present application introduces an adaptive point cloud filtering mechanism based on vehicle speed. Through the design of "maximum effective distance threshold " associated with vehicle speed, it realizes the intelligent adjustment of paying attention to more distant information when driving at high speed and focusing on the close-range environment when driving at low speed, improving the point cloud processing efficiency and resource utilization rate; in addition, the present application adopts splitting distance formula to make the resampling density dynamically associated with vehicle speed, appropriately reducing the sampling density when driving at high speed and increasing the sampling density when driving at low speed, which not only ensures the positioning accuracy but also optimizes the computing load, and is especially suitable for the real-time processing requirements in the complex road conditions of underground mines.

[0021] (2) For the high-precision map update in the underground mine environment, the prior art generally reconstructs the lane center line by using equal weight calculation or simply taking the midpoint of the left and right boundaries, but there are defects such as being unable to cope with inaccurate recognition of unilateral boundaries, being easily affected by light changes and interference from special obstacles in the mining area. The boundary credibility weight coefficients and introduced by the present application, Dynamically adjust based on boundary error to make the centerline reconstruction process more intelligent.

[0022] In addition, the local on-demand update strategy based on the boundary error threshold proposed in this application, combined with the weighted reconstruction algorithm, effectively solves the problem of inaccurate maps caused by frequent changes in road boundaries in underground mine areas, and at the same time reduces the data transmission pressure in weak communication environments. The entire reconstruction process combines segmented processing, dynamic weights, and smoothing algorithms to solve the problem of large positioning deviations of the centerline in complex scenarios such as irregular roads and unilateral fuzzy boundaries in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] This application will be further described by way of exemplary embodiments, which will be described in detail through the accompanying drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where:

[0024] Figure 1 is a schematic diagram of an exemplary application scenario of an xx system shown in some embodiments of the present application;

[0025] Figure 2 is a schematic diagram of point cloud preprocessing shown in some embodiments of the present application;

[0026] Figure 3 is a schematic diagram of an R-Tree spatial index structure shown in some embodiments of the present application;

[0027] Figure 4 is a schematic diagram of outlier filtering shown in some embodiments of the present application.

[0028] Description of the reference numerals in the drawings: DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] The methods and systems provided in the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0030] As Figure 1 shown, obtain the point cloud data during the vehicle's driving; filter the point cloud data according to the vehicle's speed and vehicle height; extract the road boundary through a deep learning network and a geometric algorithm based on the filtered point cloud data to obtain a road boundary point set; obtain the current lane boundary of the vehicle in the map; calculate the boundary error according to the road boundary point set and the current lane boundary; when the boundary error is greater than a preset threshold, reconstruct the current lane centerline according to the road boundary point set and update the local map; the local map represents a map including the current lane boundary and the current lane centerline.

[0031] During the autonomous driving process of the vehicle, real-time point cloud collection is performed, and the collected point cloud is output to the point cloud processor for filtering the point cloud.

[0032] As shown Figure 2 in the figure, the point cloud data is filtered according to the vehicle speed and vehicle height of the vehicle. First, the system obtains the current position coordinates (x, y, z) of the vehicle through the on-vehicle GPS / IMU integrated navigation system, and obtains the current vehicle speed v (unit: m / s) from the CAN bus.

[0033] Calculate the maximum effective distance threshold according to the current vehicle speed v , where the adjustment coefficient k can be adjusted according to different areas of the underground mine, and the general value range is 2.5 to 3.5. For example, in a straight and spacious roadway, k can take a smaller value of 2.5, while in a complex intersection, k can take a larger value of 3.5.

[0034] Let the current position coordinates of the vehicle be , and for each point in the point cloud, calculate the Euclidean distance: ; when is satisfied, eliminate this point; at the same time, eliminate the points where z > H, where H is the vehicle height (the height of mine vehicles is usually 2.8 to 3.2 meters). The filtered point cloud data is retained for subsequent processing.

[0035] Among them, compared with the fixed distance threshold filtering, the adaptive filtering mechanism based on vehicle speed can intelligently adjust the perception range, focus on close-range detail processing (such as obstacle avoidance, precise positioning) at low speeds, and expand the perception range at high speeds to ensure safety. By effectively filtering the point cloud data in non-critical areas, the system can concentrate limited computing resources on truly important areas, which is especially suitable for in-vehicle computing platforms in the weak communication environment of underground mines.

[0036] This system uses a pre-trained PV-RCNN deep learning network to process the filtered point cloud data and extract candidate road boundary regions. First, collect the point cloud data of different regions in the underground mine (including main roadways, auxiliary roadways, intersections, loading areas, etc.), with at least 100 frames for each region. Subsequently, perform manual annotation and mark 7-dimensional parameters for each boundary data , where represents the starting coordinates of the boundary, respectively represent the length, width, and height of the 3D bounding box, represents the lane direction angle.

[0037] The annotation data specifically considers the characteristics of the underground mine environment, such as insufficient lighting, road surface dust, irregular walls, etc., to ensure the adaptability of the model. For example, when annotating in the main roadway area, usually takes 5 to 8 meters, while in the intersection area, can be shortened to 3 to 5 meters to adapt to the complex environment.

[0038] To enhance the model's generalization ability, the original point cloud and annotation data are enhanced: Random rotation: The point cloud is randomly rotated within the range of [-π / 8, π / 8]; Random translation: Randomly translated within the range of ±0.5 meters in the horizontal direction; Random scaling: Randomly scaled within the range of [0.95, 1.05]; Random sampling: Randomly discard 10% - 15% of the point cloud data; During model training, the PV-RCNN network is optimized according to the characteristics of the road boundaries in underground coal mines. The anchor box size is adjusted from the general setting to [5, 0.2, 0.1] to adapt to the narrow and low-variation-height mine road boundaries. The training uses the Adam optimizer, with the initial learning rate set to 0.01, and decays to 0.8 times the original every 50 epochs. When the training loss changes less than 1% for 10 consecutive epochs, the training stops and the weight file is exported.

[0039] In the inference stage, the system performs boundary detection every 200 ms. After loading the filtered point cloud data and the training weight file, the PV-RCNN network outputs the 3D bounding box prediction results, including the bounding box position, size, orientation, and confidence score. Since deep learning methods may produce inaccurate boundaries in some complex scenarios (such as sudden changes in lighting, water accumulation areas), this application introduces a geometric algorithm for boundary correction.

[0040] For each candidate road boundary region, the system extracts the subset of the point cloud therein and optimizes the 3D bounding box according to the point cloud density and spatial distribution characteristics. In the context of underground coal mines, the system pays special attention to the point cloud within the height range of 5 to 15 cm near the ground, which usually best represents the actual road boundary.

[0041] This application uses an improved RANSAC algorithm to accurately fit the boundary: Randomly select 2 points (referred to as sample points) from the point cloud within the bounding box to fit the initial straight line equation; Calculate the distance from all points within the bounding box to this straight line. When the distance is less than the threshold ε (usually set to 5 - 10 cm), mark the point as an inlier; Calculate the current inlier ratio w (number of inliers / total number of points); Dynamically calculate the required number of iterations N based on the inlier ratio w: , where p is the expected probability (usually set to 0.99), and n is the minimum number of points required to fit the straight line (here it is 2); Update the inlier ratio and the number of iterations after each round of iteration ; Stop the iteration when any of the following conditions is met: reaching the preset maximum number of iterations (usually 200 times), the change in the inlier ratio w is less than 0.01 within 5 consecutive rounds, the inlier ratio w exceeds 0.9. Select the line with the most inliers from all iterations, refit using all inliers, and obtain the optimal boundary line. This dynamic iteration mechanism is particularly suitable for the underground mine environment. For example, at a clear straight boundary, the number of iterations may only be 20 to 30 times, while at a fuzzy boundary, it may require 150 to 180 times of iteration, achieving efficient utilization of computing resources.

[0042] Finally, the system classifies the optimal boundary line to obtain a structured set of road boundary points. The system extracts the part of the optimal boundary line located within the three-dimensional bounding box as the valid line segment. In the complex roadways of underground mines, this step can effectively filter out the misdetected parts caused by lighting or obstacles.

[0043] Generate a boundary point every d meters (d is usually set to 0.5 - 1 meter) along the valid line segment to form a set of boundary points . In the area of curves or intersections, the system automatically reduces the d value (down to 0.3 to 0.5 meters) to improve the boundary expression accuracy.

[0044] Obtain the current coordinate point of the vehicle and the driving direction vector V, and perform the following operations: for each boundary point , calculate the vector pointing from the vehicle to this point ; Calculate the cross product result of and V ; Classify according to the cross product result: when , classify as the left boundary point set ; when , classify as the right boundary point set

[0045] This direction-aware classification method enables the system to correctly distinguish the left and right boundaries in the complex road network of underground mines, and can work reliably even in irregular roadways or Y-shaped fork intersections. and Merge to form a complete set of road boundary points, and perform post-processing optimization: remove isolated points (points with an abnormally large distance from adjacent points); smooth processing (apply the weighted moving average algorithm); density equalization (ensure that the density of boundary points in each section is similar).

[0046] Obtain the current lane boundary of the vehicle in the map. First, load the complete lane data from the high-precision map database, including the center line, left and right boundary lines, attribute information, etc. of each lane. For the complex road network environment of underground mines, use the spatial index technology R-Tree data structure to organize the lane data. As Figure 3As shown, the R-Tree is a balanced tree structure optimized for multi-dimensional spatial data retrieval. Its core idea is to cluster similar spatial objects into the same node to form a hierarchical index.

[0047] In this embodiment, the system divides the underground mine area into multiple overlapping rectangular regions, and each lane segment is associated with a corresponding rectangular region. The leaf nodes of the R-Tree store the lane segment ID and the coordinates of its corresponding minimum bounding rectangle (MBR), and the non-leaf nodes store the MBR aggregation information of the child nodes. This structure makes the spatial query complexity close to , greatly improving the vehicle positioning efficiency.

[0048] The lane where the vehicle is located is queried every 100 ms. After obtaining the current coordinate point of the vehicle , the following query process is executed: starting from the root node of the R-Tree, check whether the MBR of the child node contains the current position of the vehicle. For each node, the system obtains its minimum bounding rectangle MBR, which is uniquely determined by the lower left corner coordinates and the upper right corner coordinates .

[0049] Judge whether the vehicle coordinate point meets the conditions and . Meeting the conditions means that the vehicle may be in a certain lane under the jurisdiction of this node, and the system will continue to traverse the child nodes of this node downward.

[0050] After reaching the leaf node, the system obtains all the lane segments associated with this node. Usually in the underground mine environment, a leaf node contains 2 - 5 lane segments.

[0051] For each candidate lane, the system connects its left and right boundaries to form a closed polygon, and then uses the angle sum method to determine whether the vehicle is inside this polygon: calculate the vectors from the vehicle coordinate point C to each vertex of the polygon; calculate the angles between adjacent vectors; accumulate all the angles. If the sum is close to 360° (allowing an error of ±0.1°), it is determined that the vehicle is in this lane.

[0052] In special areas such as intersections, it may occur that the vehicle matches multiple lanes at the same time. At this time, the system will select the most reasonable lane as the current lane according to information such as the vehicle's forward direction and historical position.

[0053] After determining the lane where the vehicle is located, the system extracts the complete boundary information of this lane. Each lane boundary is represented as an ordered point sequence in the high-precision map, usually sampling a point every 0.5 meters. The system obtains the left and right boundary point sequences within 50 meters before and after the vehicle as the current lane boundary for subsequent boundary error calculation.

[0054] In the narrow tunnels of underground mines, the system can adaptively adjust the extraction range according to the road width. For example, in a narrow tunnel with a width of 2.5 meters, it can be reduced to 30 meters in front and back to improve processing efficiency.

[0055] After obtaining the current lane boundary and the boundary point set detected in real time, the system performs accurate boundary error calculation. Calculate the left boundary distance: Each point in , calculate the shortest distance to the left edge of the lane in the map The calculation process uses the point-to-line shortest distance algorithm: for each pair of adjacent points A and B on the left border of the map, a line segment AB is formed; the point The shortest distance to line segment AB; take the minimum distance among all line segments as the distance from pi to the left boundary ; All The values ​​​​compose the left boundary point distance set .

[0056] Calculate the right boundary distance: Similarly, calculate the right boundary point set The distance from each point to the right edge of the map , forming the right boundary point distance set .

[0057] right and Filter outliers in the , and eliminate possible noise effects. Figure 4 As shown, the system uses the 3σ criterion to eliminate distance values ​​that are significantly deviated from the average value to ensure the accuracy of boundary error calculation.

[0058] Based on the filtered distance set, calculate the statistical features: the average error of the left boundary: , where n is the left boundary point set The number of points in is the distance from the i-th left boundary point to the left boundary of the map. The average error of the right boundary is: , where m is the right boundary point set The number of points in is the distance from the jth right boundary point to the right boundary of the map.

[0059] Calculate the distance set of left boundary points The distance set from the right boundary point The average value of the distances in ;

[0060] Considering that the boundary of underground mining roads may be locally deformed or damaged, the system divides the boundary points into several segments according to their locations. Each segment contains k consecutive points (usually k is 5-10), and the local average distance of each segment is calculated:

[0061] Average distance of the s-th segment of the left boundary: 。Average distance of the t-th segment of the right boundary: 。These segment averages and constitute the set of average values of each distance 。

[0062] Based on the average value of each distance , calculate the total boundary error error: , this weighting ensures a reasonable error assessment even when there are fewer boundary points on one side (for example, one side is blocked by an obstacle).

[0063] When the system detects that the boundary error exceeds the preset threshold (usually 0.25 meters for the main road, 0.3 meters for the working area, and 0.4 meters for the loading area), the boundary point resampling operation is first performed. This system adopts an innovative vehicle speed adaptive resampling mechanism, and the specific implementation is as follows: Obtain the current speed v of the vehicle (unit: m / s), and calculate the segmentation distance according to the formula , where is the segmentation adjustment coefficient. In different areas of the underground mine, takes dynamic adjustment: Straight section: , for example, when the vehicle speed is 5 m / s, the sampling interval is 0.5 meters; Curve section: , for example, when the vehicle speed is 5 m / s, the sampling interval is reduced to 0.25 meters; Intersection: , providing a more refined boundary expression; This adaptive mechanism ensures that the number of sampling points is appropriately reduced during high-speed driving, reducing the computational burden; When driving at low speed, the sampling point density increases, improving the boundary accuracy.

[0064] Perform parametric processing on the original boundary point set and : First, assign a cumulative distance parameter to each boundary point along the driving direction; Use cubic spline interpolation to construct a continuous boundary curve function; According to the calculated segmentation distance d, uniformly sample on the parametric curve; Generate the resampled boundary point set and ; In the narrow curve area of the underground mine, this technology ensures the accurate expression of the curve boundary, and the typical value range of the sampling point interval is 0.2 to 0.8 meters.

[0065] To solve the common problem of unilateral boundary ambiguity in the underground mine environment (such as ore accumulation, water accumulation, insufficient lighting, etc. on one side), the system introduces a boundary credibility weight mechanism: Reuse the calculated left and right boundary point distance sets, and calculate the average error: Average error of the left boundary: ; Average error of the right boundary: 。

[0066] This application adopts the principle that "the higher the credibility of the boundary with smaller error". The system sets weights through the following formulas: Weight of the left boundary: ; Weight of the right boundary: ; This design ensures that , and the smaller the boundary error, the greater its weight. In practical applications, when the left boundary is blurred due to ore accumulation (such as meters, meters), the system will calculate , , making the reconstructed centerline more biased towards the reliable right boundary.

[0067] Establish a correspondence based on the direction vector, obtain the vehicle driving direction vector V, and establish the correspondence between the left and right boundary points: Establish a local coordinate system at the vehicle position, with the driving direction as the y-axis; for each left boundary point , calculate its abscissa in the local coordinate system ; perform the same operation on the right boundary point to obtain the abscissa ; establish the correspondence between the left and right boundary points based on the proximity of the abscissas (the principle of minimum distance); form n horizontal connection line segments ; in the case of mismatched numbers of points, the system uses the dynamic programming algorithm to find the optimal match and avoid cross-matching.

[0068] For each horizontal connection line segment, calculate the weighted center point by applying the credibility weight: , this weighting mechanism enables the centerline to remain accurate even when one-sided boundaries are blurred. Specifically, this application solves the problem of difficult determination of the correspondence between the left and right boundary points in irregular road sections; introduces direction information, making the boundary matching directional and avoiding incorrect matching. In addition, this application breaks through the limitations of the traditional simple midpoint-taking method; automatically "tilts" towards the more reliable boundary through the weight mechanism, improving the accuracy of the centerline; and can still maintain the accuracy of the centerline in the case of poor recognition of one-sided boundaries.

[0069] Connect all the weighted center points in the order of the vehicle driving direction , forming an initial center point sequence . To eliminate local noise and discontinuity, the system applies a third-order smoothing algorithm: First-level smoothing: Use a 5-point weighted moving average to filter high-frequency noise; Second-level smoothing: Apply B-spline fitting with a relaxation factor λ = 0.5; Third-level smoothing: Curvature limit processing to ensure that the curvature change conforms to the vehicle dynamics constraints; After smoothing, the curvature change of each point on the centerline is limited within 0.1 / meter, meeting the requirements of the steering dynamics of mining area vehicles.

[0070] Determine the local update range based on the vehicle's position and driving direction: In front of the vehicle: v × 10 seconds (at least 50 meters); Behind the vehicle: fixed 30 meters; Horizontally: 1.5 times the width of the current lane; This dynamic range setting ensures that the update area adjusts with the vehicle speed, and the update range in front is longer when driving at high speed.

[0071] To avoid discontinuity between the update area and the original map, the system applies transitional fusion at the boundaries of the update area: Set the length of the transition zone to 5 meters; Within the transition zone, the old and new map data are linearly fused according to the distance weight; Ensure the continuity of the position, direction, and curvature of the center line and boundaries at the connection.

[0072] The above schematically describes the present invention and its implementation manners. This description is not restrictive. Without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Any reference signs in the claims should not limit the claims involved. Therefore, if those of ordinary skill in the art are inspired by it and, without departing from the purpose of this creation, design structurally similar ways and embodiments to this technical solution without creative efforts, they should all fall within the protection scope of this application. In addition, the term "including" does not exclude other elements or steps, and the term "a" before an element does not exclude including "a plurality of" such elements. The multiple elements stated in the product claims can also be implemented by one element through software or hardware. First, second, etc. are used to indicate names and do not indicate any specific order.

Claims

1. A method for real-time update of local maps, characterized in that, Including: Obtain point cloud data during vehicle driving; Filter the point cloud data according to the vehicle speed and vehicle height; Extract the road boundary from the filtered point cloud data through a deep learning network and a geometric algorithm to obtain a road boundary point set; Obtain the current lane boundary of the vehicle in the map; Calculate the boundary error according to the road boundary point set and the current lane boundary; When the boundary error is greater than a preset threshold, reconstruct the current lane center line according to the road boundary point set and update the local map; The local map represents a map including the current lane boundary and the current lane center line; S6, reconstruct the current lane center line according to the road boundary point set and update the local map, including: According to the left boundary point set and the right boundary point set , along the respective boundary line directions, resample according to the segmentation distance , where is the segmentation adjustment coefficient; obtain the resampled left boundary point set and the resampled right boundary point set ; among them, straight line segment: ; curved road area: ; intersection: ; According to the left boundary point distance set , calculate the left boundary average error ; According to the right boundary point distance set , calculate the right boundary average error ; According to the average error of the left boundary and the average error of the right boundary , set the boundary credibility weight coefficients and ; According to the vehicle driving direction vector V, for the points in the resampled and , establish the corresponding relationship of the left and right boundary points to form multiple horizontal connection line segments; establish a local coordinate system at the vehicle position, with the driving direction as the y-axis; for each left boundary point , calculate its abscissa in the local coordinate system; perform the same operation on the right boundary points to obtain the abscissa ; establish the corresponding relationship of the left and right boundary points based on the similarity of the abscissas; form n horizontal connection line segments ; For each horizontal connecting line segment, calculate the weighted center point , where and are the coordinates of the left and right endpoints of the corresponding line segment, respectively; Connect all the weighted center points in sequence according to the vehicle driving direction , to form an initial center point sequence; Smooth the initial center point sequence to generate a reconstructed lane center line; Update the local map according to the reconstructed lane center line and the obtained current lane boundary.

2. The local map real-time update method according to claim 1, characterized in that: S2, filter the point cloud data according to the vehicle speed and vehicle height, including: Obtain the current position coordinates of the vehicle and the current vehicle speed v; Calculate the maximum effective distance threshold , where k is a preset adjustment coefficient; Traverse each point in the point cloud data and calculate the distance between the point and the current position coordinates of the vehicle; Filter out the points with a distance greater than the maximum effective distance threshold and filter out the points in the point cloud data with a z coordinate value greater than the vehicle height to obtain the filtered point cloud data.

3. The local map real-time update method according to claim 1, characterized in that: S3, obtain the road boundary point set, including: Use the pre-trained PV-RCNN deep learning network to extract the filtered point cloud data to obtain a candidate road boundary area; Generate a three-dimensional bounding box according to the candidate road boundary area; Use a geometric algorithm to correct the three-dimensional bounding box to obtain an optimal boundary line; Classify the boundary points according to the optimal boundary line to obtain a road boundary point set.

4. The local map real-time update method according to claim 3, characterized in that: Obtain the optimal boundary line, including: Randomly select multiple points from the point clouds of each three-dimensional bounding box to fit an initial straight line; Calculate the distance from each point to the initial straight line. When the distance is less than a preset threshold, mark the corresponding point as an inlier; Calculate the remaining number of iterations N according to the inlier ratio w; After each iteration, calculate the inlier ratio , and update the remaining number of iterations according to the inlier ratio ; When the preset iteration upper limit is reached or the change in the inlier ratio w within a continuous preset number of rounds is less than the threshold, stop the iteration and select the straight line with the most inliers as the optimal boundary line.

5. The local map real-time update method according to claim 4, characterized in that: Calculate the remaining number of iterations N according to the inlier ratio w, and the calculation formula is: where p represents the expected probability of randomly sampling and selecting all inliers; n is the minimum number of points required to fit a straight line.

6. The local map real-time update method according to claim 3, characterized in that: Classify the boundary points according to the optimal boundary line to obtain a road boundary point set, including: Obtain the line segment of the optimal boundary line located within the three-dimensional bounding box as the effective line segment; Generate a boundary point at a preset distance interval according to the valid line segments to obtain a set of boundary points ; Obtain the current coordinate point of the vehicle and the vehicle driving direction vector V; For each boundary point , calculate the cross product result of the vector pointing from the current coordinate point of the vehicle to the boundary point , and the vector V ; When the cross product result is greater than zero, the corresponding boundary points are classified into the left boundary point set , when the cross product result is less than or equal to zero, the corresponding boundary points are classified into the right boundary point set ; Merge the left boundary point set and the right boundary point set to form a complete road boundary point set.

7. The local map real-time update method according to any one of claims 2 to 6, characterized in that: S4, obtain the current lane boundary of the vehicle in the map, including: Obtain the lane data of the map; Use the R-Tree data structure to store the lane data; Starting from the root node of the R-Tree data structure, traverse the R-Tree data structure to obtain the minimum bounding rectangle (MBR) of each node. Among them, the lower left corner coordinates of the minimum bounding rectangle MBR , and the upper right corner coordinates ; Determine the current coordinate point of the vehicle , whether it satisfies and . If it satisfies, then determine that the current coordinate point C of the vehicle is located in the current MBR; Traverse each lane in the MBR where the vehicle is located and calculate the current coordinate point of the vehicle The sum of the angles formed by connecting the current coordinate point of the vehicle with each point on the lane boundary. If the sum of the angles is 360°, it is determined that the vehicle is within the current lane, and the left boundary and right boundary of the lane are obtained as the current lane boundaries.

8. The local map real-time update method according to claim 7, characterized in that: S5. Calculate the boundary error based on the road boundary point set and the current lane boundary, including: Calculate the distance from each point in the left boundary point set to the left boundary of the lane to obtain the left boundary point distance set ; Calculate the distance from each point in the right boundary point set to the right boundary of the lane to obtain the right boundary point distance set ; Calculate the average value of the distances in the left boundary point distance set and the right boundary point distance set ; ; Based on the average value of each distance , calculate the total boundary error error: .

9. A local map real-time update system, characterized in that: It includes: At least one processing unit; configured to execute instructions to implement the local map real-time update method according to any one of claims 1 to 8.

Images