Dynamic boundary expansion method applied to airport automatic driving tractor

Abstract

The invention relates to a dynamic boundary expansion method applied to an airport automatic driving tractor, and belongs to the technical field of intelligent transportation, and the method comprises the steps: initial point cloud map construction: generating an airport point cloud map through a laser radar SLAM technology; real-time sensing and comparison: registering the current point cloud with a locally cached airport point cloud map, and extracting a difference region based on a grid map; and dynamically updating the map: receiving the difference areas reported by the plurality of vehicles, and dynamically updating the airport point cloud map by the cloud server according to the updating strategy. According to the method, through innovative designs such as space-time dimension dual verification, multi-modal data fusion and hierarchical progressive updating, the sub-second-level response capability to the dynamic environment of the airport is realized while the centimeter-level positioning precision is ensured, and through a sliding time window and a confidence coefficient accumulation mechanism, the positioning accuracy is improved. And the problem of invalid obstacles possibly caused by a single sensing error is effectively solved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent transportation, and in particular to a dynamic boundary expansion method applied to an airport automatic driving tractor. BACKGROUND

[0002] In the automatic driving technology, the dynamic boundary expansion refers to the ability of a vehicle to improve safety and driving efficiency by perceiving the environment in real time and dynamically adjusting the "perception-planning" boundary. The traditional airport automatic driving relies on a static high-precision map, but cannot adapt to dynamic obstacles (such as temporary maintenance areas and mobile fences). The existing technology has a long map updating cycle, resulting in poor real-time performance.

[0003] The single vehicle perception ability is insufficient, and the vehicle-mounted radar / camera cannot establish the global spatial relationship alone, it is difficult to distinguish temporary obstacles from permanent structures, and it is easy to cause path planning errors. The point cloud data utilization defects, the existing scheme does not compare real-time point cloud with historical point cloud in multiple dimensions, and lacks confidence evaluation and incremental updating mechanism for dynamic obstacles. SUMMARY

[0004] To solve the above-mentioned problems in the prior art, the application provides a dynamic boundary expansion method applied to an airport automatic driving tractor.

[0005] A dynamic boundary expansion method applied to an airport automatic driving tractor, comprising the following steps:

[0006] Initial point cloud map construction: generate an airport point cloud map through laser radar SLAM technology, and upload the cloud server;

[0007] Real-time perception and comparison: during the driving of the vehicle, download the airport point cloud map from the cloud server to the local vehicle system, register the current point cloud with the local cached airport point cloud map, extract the difference area based on the grid map, and verify through the semantic camera auxiliary verification;

[0008] Dynamic map updating: receive the difference area reported by multiple vehicles, perform multi-source data fusion by using D-S evidence theory, and dynamically update the airport point cloud map according to the updating strategy of the cloud server.

[0009] Further, the registration of the current point cloud with the local cached airport point cloud map is specifically:

[0010] Point cloud preprocessing, including downsampling processing and ground segmentation;

[0011] ICP registration, in the registration process, a feature constraint term is introduced: E = a E ICP + b E feature, E is the total optimization target energy value, a and b are weight coefficients, a + b = 1, E ICP is the alignment error of the traditional ICP algorithm, and E feature is the feature matching constraint error.

[0012] Further, the difference region extraction is specifically:

[0013] A difference matrix is constructed, diff_matrix = (current_voxel-baseline_voxel) / density_threshold, wherein current_voxel is the point density value of the current scanning point cloud in the grid map, baseline_voxel is the point density value of the corresponding grid of the baseline high-precision map, density_threshold is the minimum effective point density threshold, and diff_matrix is the normalized density difference matrix.

[0014] Region clustering, adjacent difference points are merged by using the DBSCAN algorithm, and noise regions with an area of less than 1 square meter are filtered.

[0015] Further, the downsampling processing is specifically: the point cloud is subjected to downsampling processing by voxel grid filtering to reduce the data size and uniformize the density.

[0016] Further, the ground segmentation is specifically: the ground plane is segmented based on the RANSAC algorithm to remove redundant ground points, and the non-ground point cloud input for ICP registration is provided, which is denoised, lightweight and retains geometric features.

[0017] Further, the update strategy is specifically:

[0018] New obstacle detection: identify new obstacles in the fused data;

[0019] Generate temporary boundaries: generate temporary boundaries based on the positions of the obstacles;

[0020] Airport point cloud map update: verify the temporary boundaries through data of different vehicles, and when the verification is passed, update the airport point cloud map of the cloud server.

[0021] Further, the process of verifying the temporary boundaries through data of different vehicles is specifically:

[0022] Set a sliding time window W(t): W(t) = [t-△t, t], the right boundary is the current time point t, and △t is the window length, and the window dynamically slides with the system time;

[0023] Confidence calculation: C(t) = a x K(t) + b x N(t) + c x S(t), C(t) is the confidence of the temporary boundary at the current time point, K(t) is the number of vehicles in the window W(t) reporting data overlapping the temporary boundary, N(t) is the time in the window W(t) reporting data overlapping the temporary boundary, S(t) is the visual semantic matching degree, which is obtained by YOLO v7 target detection to identify the similar matching degree of obstacles and obstacle signs, a, b and c are weight coefficients, a + b + c = 1;

[0024] Update trigger condition: if the cumulative duration of the confidence exceeding the threshold value m reaches T, that is, it satisfies Where χ {C(t)≥m} is an indicator function, 1 when C(t) > m, the integral condition is established, and the airport point cloud map is automatically updated; otherwise, 0.

[0025] Further, the updating process includes: global point cloud resampling; construction sign identification based on YOLO v7, generating a polygon no-entry area, and dynamically adjusting the path network graph.

[0026] The beneficial effects of the present application are: through the innovative design of spatiotemporal dimension double verification, multi-modal data fusion, hierarchical progressive update, etc., while ensuring centimeter-level positioning accuracy, the present application realizes sub-second response capability to the dynamic environment of the airport, and the map update delay is reduced from hours to minutes. Through the multi-vehicle cross-verification of the sliding time window and the confidence accumulation mechanism, the false positive rate is reduced to below 0.1%, effectively solving the problem of invalid obstacles caused by single sensing error. At the same time, through incremental updating, the data transmission amount is reduced, and the resources are optimized. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0029] In the present embodiment: referring to Figure 1 The method for dynamically expanding the boundary of the automatic driving tractor applied to the airport of the present application, comprising:

[0030] I. Initial point cloud map construction: generate an airport point cloud map through laser radar SLAM technology, and upload to a cloud server;

[0031] II. Real-time perception and comparison: During vehicle driving, the vehicle-mounted system performs the following every 30 seconds:

[0032] a. Download the airport point cloud map from the cloud server to the local vehicle-mounted system, and the vehicle-mounted system will register the current point cloud with the locally cached airport point cloud map,

[0033] b. Difference area extraction based on grid map; (0.1m x 0.1m resolution)

[0034] c. Verification by semantic camera. (Recognize construction signs, cone barrels, etc.)

[0035] Wherein, the registration of the current point cloud with the locally cached airport point cloud map specifically comprises:

[0036] 1. Point cloud preprocessing:

[0037] (1) Down-sampling processing: The point cloud is down-sampled by voxel grid filtering to reduce the data size and uniformize the density, and the specific process is as follows:

[0038] Voxel space division: The point cloud space is divided into uniform three-dimensional cubic grid (voxel), and the voxel edge length is set according to the requirement (such as 0.05m).

[0039] Voxel point aggregation: All points falling into the same voxel are regarded as a local point group.

[0040] Centroid / representative point calculation: Calculate the geometric centroid (or randomly select a representative point) of the points in each voxel, and replace the original dense points.

[0041] Sparse output: Only the centroids or representative points are retained to form the down-sampled point cloud.

[0042] (2) Ground segmentation; Based on RANSAC algorithm, the ground plane is segmented to remove redundant ground points, providing a non-ground point cloud input for ICP registration that is denoised, lightweight and retains geometric features, thereby improving registration efficiency and accuracy, and the specific process is as follows:

[0043] Model assumption: Assume that the ground is a plane model, and the parameters are the plane equation ax+by+cz+d=0.

[0044] Random sampling: Randomly select 3 points from the point cloud to calculate the plane parameters of the fitting.

[0045] Inlier selection: Count the points with a distance less than a threshold (such as 0.05m) from the plane as inliers.

[0046] Iterative optimization: Repeat the above steps, and keep the plane with the most inliers as the optimal ground model.

[0047] Point cloud segmentation: according to the optimal plane equation, the point cloud is divided into ground points (inliers) and non-ground points (outliers).

[0048] 2. ICP registration:

[0049] (1) Feature extraction.

[0050] (2) Bidirectional matching constraint: introduce feature constraint term in the registration process: E = a E ICP + b E feature, where E is the total optimization target energy value, a and b are weight coefficients (usually a + b = 1), E ICP is the alignment error of traditional ICP (iterative closest point) algorithm, E feature is the feature matching constraint error.

[0051] (3) Robust optimization, using RANSAC+LM hybrid optimization.

[0052] (4) Dynamic weight evaluation, special processing for airport flat ground, as follows:

[0053] Call is_flat_region() to judge the point cloud region characteristics: calculate the standard deviation of the point cloud normal vector, <0.1 is determined as a flat region;

[0054] Flat area sets weight a = 0.8, b = 0.2 (enhance ground constraint), feature-rich area sets a = 0.3, b = 0.7;

[0055] Execute apply_ground_constraint(): limit the rotation freedom degree of Z axis of transformation matrix, force horizontal alignment (accuracy improved to <5cm).

[0056] Among them, the difference area extraction is specifically:

[0057] Construct difference matrix, diff_matrix = (current_voxel-baseline_voxel) / density_threshold, where current_voxel is the point density value of the current scanning point cloud in the grid map, baseline_voxel is the point density value of the corresponding grid of the reference high-precision map, density_threshold is the minimum effective point density threshold, and diff_matrix is the normalized density difference matrix.

[0058] Region clustering, using DBSCAN algorithm to merge adjacent difference points, and filtering noise areas with area <1㎡.

[0059] III. Dynamic updating of the map: receiving the difference areas reported by multiple vehicles, using D-S evidence theory for multi-source data fusion, and updating the airport point cloud map on the cloud server according to the updating strategy.

[0060] In this embodiment, the updating strategy is specifically:

[0061] New obstacle detection: identifying new obstacles in the fused data;

[0062] Temporary boundary generation: generating a temporary boundary based on the position of the obstacle;

[0063] Airport point cloud map updating: verifying the temporary boundary through data from different vehicles, and updating the airport point cloud map on the cloud server when the verification is passed.

[0064] Specifically, the process of verifying the temporary boundary through data from different vehicles is as follows:

[0065] Setting a sliding time window W(t): W(t) = [t-△t, t], the right boundary is the current time point t, and△t is the window length, and the window slides dynamically with the system time;

[0066] Confidence calculation: C(t) = a x K(t) + b x N(t) + c x S(t), C(t) is the confidence of the temporary boundary at the current time point, K(t) is the number of vehicles reporting data overlapping with the temporary boundary within the window W(t), N(t) is the time of reporting data overlapping with the temporary boundary within the window W(t), S(t) is the visual semantic matching degree, which is obtained by identifying the similarity matching degree of the obstacle and the obstacle identifier through YOLO v7 target detection, and a, b and c are weight coefficients, a + b + c = 1;

[0067] Update trigger condition: if the confidence exceeds the threshold m for a cumulative duration of T, i.e. where X {C(t)≥m} is an indicator function, which takes 1 when C(t) > m, and the integral condition is satisfied, and the airport point cloud map is automatically updated; otherwise, it takes 0.

[0068] The process of updating includes:

[0069] Structural layer updating: global point cloud resampling;

[0070] Semantic layer updating: construction sign recognition based on YOLO v7, generating a polygon no-entry zone; (buffer distance ≥ 0.5 m)

[0071] Topological layer updating: dynamically adjusting the path network graph.

[0072] In the present application, the hierarchical processing mechanism of vehicle-mounted end (feature extraction) -> edge server (change confirmation) -> cloud end (global update).

[0073] The present application realizes sub-second response capability to the dynamic environment of the airport while ensuring centimeter-level positioning accuracy through the innovative design of spatiotemporal dimension double verification, multi-modal data fusion, hierarchical progressive update, etc. The map update delay is reduced from hours to minutes. Through the multi-vehicle cross-verification of the sliding time window and the confidence accumulation mechanism, the false positive rate is reduced to below 0.1%, effectively solving the problem of invalid obstacles caused by single sensing error. At the same time, through incremental update, the data transmission volume is reduced by 70% (only transmit change data (difference grid coordinates, temporary no-entry area vector, path weight change), using Delta compression encoding; the daily average traffic of a single vehicle is reduced from 801MB to 30.1MB (96% saved), and the conservative evaluation of the overall transmission volume is reduced by 70%), realizing the optimization of resources.

[0074] In the description of the embodiments of the present application, the terms "first", "second", "third", "fourth" are only used for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second", "third", "fourth" can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0075] In the description of the embodiments of the present application, the term "and / or" herein only describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " herein generally represents an "or" relationship between the front and rear associated objects.

[0076] Although the embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations of the embodiments can be made without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic boundary expansion of an autonomous driving tractor at an airport, characterized in that: The following steps are involved: Initial point cloud map construction: Generate an airport point cloud map using LiDAR SLAM technology and upload it to the cloud server; Real-time perception and comparison: While the vehicle is driving, the airport point cloud map is downloaded from the cloud server and cached locally on the vehicle system. The vehicle system then aligns the current point cloud with the locally cached airport point cloud map, extracts difference areas based on the grid map, and uses semantic camera assistance for verification. Dynamic map update: Receive difference areas reported by multiple vehicles, use DS evidence theory to fuse multi-source data, and the cloud server dynamically updates the airport point cloud map according to the update strategy.

2. The method for dynamic boundary expansion of an automatic driving tractor for an airport according to claim 1, characterized in that: The specific steps of registering the current point cloud with the locally cached airport point cloud map are as follows: Point cloud preprocessing, including downsampling and ground segmentation; ICP registration introduces a feature constraint term in the registration process: E = α·E_{ICP}+β·E_{feature}, where E is the total optimization target energy value, α and β are weight coefficients, α+β=1, E_{ICP} is the alignment error of the traditional ICP algorithm, and E_{feature} is the feature matching constraint error.

3. The method for dynamic boundary expansion of an autonomous driving tractor for an airport according to claim 1, characterized in that: The specific steps of extracting the difference regions are as follows: Construct the difference matrix, diff_matrix = (current_voxel-baseline_voxel) / density_threshold, where current_voxel is the point density value of the current scanned point cloud in the raster map, baseline_voxel is the point density value of the grid corresponding to the benchmark high-precision map, density_threshold is the minimum valid point density threshold, and diff_matrix is ​​the standardized density difference matrix. For regional clustering, the DBSCAN algorithm is used to merge adjacent difference points and filter out noise areas with an area of ​​<1㎡.

4. The method for dynamic boundary expansion of an airport autonomous driving tractor according to claim 1, characterized in that: The downsampling process is specifically: downsampling the point cloud by voxel grid filtering to reduce the data size and uniformize the density.

5. The method for dynamic boundary expansion of an automatic driving tractor at an airport according to claim 1, characterized in that: The ground segmentation specifically includes: segmenting the ground plane based on the RANSAC algorithm to remove redundant ground points, and providing a denoised, lightweight and geometrically feature-preserving non-ground point cloud input for ICP registration.

6. The method for dynamic boundary expansion of an automatic driving tractor at an airport according to claim 1, characterized in that: The update strategy is specifically as follows: New obstacle detection: Identify new obstacles in the fused data; Generate temporary boundaries: Generate temporary boundaries based on the location of obstacles; Airport point cloud map update: The temporary boundary is verified through data from different vehicles. When the verification is passed, the airport point cloud map on the cloud server is updated.

7. The method for dynamic boundary expansion of an automatic driving tractor at an airport according to claim 6, characterized in that: The specific process of verifying the temporary boundary using data from different vehicles is as follows: Set the sliding time window W(t): W(t) = [t-△t, t], where the right boundary is the current time point t, △t is the window length, and the window slides dynamically with the system time; Confidence calculation: C(t) = a×K(t)+b×N(t)+c×S(t), where C(t) is the confidence of the temporary boundary at the current time point, K(t) is the number of vehicles whose reported data overlaps with the temporary boundary in the window W(t), N(t) is the time that the reported data in the window W(t) overlaps with the temporary boundary, S(t) is the visual semantic matching degree, a, b, and c are weight coefficients, and a+b+c=1; Update trigger condition: If the cumulative time that the confidence exceeds the threshold m reaches T, it is satisfied Among them, χ{C(t)≥m} is the indicator function. When C(t)>m, it takes 1, the integration condition is met, and the airport point cloud map is automatically updated; otherwise, it takes 0.

8. The method for dynamic boundary expansion of an automatic driving tractor at an airport according to claim 6, characterized in that: The updating process includes: global point cloud re-collection; construction sign recognition based on YOLO v7, generation of polygonal no-entry zones, and dynamic adjustment of the path network diagram.

Images