Unmanned obstacle avoidance method and system applied to forage harvesting

Abstract

The application provides an unmanned obstacle avoidance method and system applied to forage grass harvesting, relates to the technical field of agricultural machinery automation, and first adjusts the cooperative adaptation relationship of an obstacle avoidance perception component and a harvesting component according to the topographic features of a forage grass harvesting area and unmanned operation parameters, integrates to form a cooperative adaptation parameter, then starts synchronous collection of multiple perception components according to the cooperative adaptation parameter, obtains obstacle feature information, forage grass growth information and ground state information, integrates to form a three-dimensional perception dataset, then builds a target coupling model, divides an avoidance space, marks the space coordinates to form coupling model data, then plans and optimizes a dynamic obstacle avoidance trajectory according to the coupling model data and the cooperative adaptation parameter, and finally converts the optimized obstacle avoidance trajectory into an unmanned device execution instruction to adjust the device.

Description

Technical Field

[0001] This invention relates to the field of agricultural machinery automation technology, and more specifically, to an unmanned obstacle avoidance method and system for forage harvesting. Background Technology

[0002] In the field of forage harvesting, traditional methods primarily rely on manual operation of machinery. However, manual operation has many limitations. On the one hand, manual operation makes it difficult to fully and accurately perceive the complex terrain features of the forage harvesting area, such as ground softness, undulation, and flatness. For different terrain conditions, it is difficult for humans to quickly and appropriately adjust harvesting parameters, such as the lateral coverage area of ​​a single harvest, travel speed, and the working position of the harvesting components. This can lead to low harvesting efficiency and even damage to the equipment. On the other hand, when facing various obstacles in the forage harvesting area, manual operation can only rely on experience and visual observation for obstacle avoidance. It is difficult to accurately obtain detailed characteristic information of obstacles in advance, such as their outline size, material hardness, and movement status, and it is also impossible to effectively plan obstacle avoidance trajectories. This can easily lead to untimely or excessive obstacle avoidance, affecting the continuity and quality of the harvesting operation. Furthermore, traditional methods lack comprehensive consideration of forage growth information and ground condition information, making it difficult to achieve precise forage harvesting operations and failing to meet the demands of modern forage harvesting for high efficiency, precision, and intelligence. Summary of the Invention

[0003] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an unmanned obstacle avoidance method for forage harvesting, the method comprising: By combining the terrain features of the forage harvesting area with the operating parameters of the unmanned operation, the cooperative adaptation relationship between the obstacle avoidance perception component and the harvesting component is adjusted and integrated to form cooperative adaptation parameters. The terrain features include ground softness, ground undulation, and ground flatness. The operating parameters include the lateral coverage size of the equipment in a single harvest, the driving speed, and the operating position of the harvesting component. Based on the collaborative adaptation parameters, the synchronous acquisition process of multiple sensing components is initiated to capture obstacle feature information, pasture growth information, and ground condition information in the pasture harvesting area, and integrate them to form a three-dimensional sensing dataset. The obstacle feature information includes outline size, material hardness, and motion state; the pasture growth information includes plant distribution, coverage area, and plant height; and the ground condition information includes soil compaction and pothole distribution. A target coupling model was built based on the 3D perception dataset. In the target coupling model, the avoidance space corresponding to different types of obstacles was divided. The spatial coordinates of the core harvesting area and the avoidable area of ​​pasture harvesting were marked and integrated to form the coupling model data. By combining coupled model data and collaborative adaptation parameters, a dynamic obstacle avoidance trajectory is planned. The overall direction of the dynamic obstacle avoidance trajectory is adjusted to match the adaptation value of the harvesting component's operating position, so that the direction of the dynamic obstacle avoidance trajectory fits the obstacle avoidance space and the lateral coverage size of the equipment's single harvest. This integration yields an optimized obstacle avoidance trajectory. Based on the optimized obstacle avoidance trajectory, the unmanned driving equipment executes commands. The driving direction and speed of the equipment are adjusted according to the optimized obstacle avoidance trajectory. The working position and harvesting frequency of the harvesting components are adjusted according to the adaptation value of the harvesting component's working position. The unmanned driving equipment executes commands in an integrated manner.

[0004] Furthermore, embodiments of the present invention also provide an unmanned obstacle avoidance system for forage harvesting, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the aforementioned unmanned obstacle avoidance method for hay harvesting by executing the machine-executable instructions.

[0005] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions stored in a computer-readable storage medium, a processor of an unmanned obstacle avoidance system applied to hay harvesting reads the machine-executable instructions from the computer-readable storage medium, the processor executes the machine-executable instructions, causing the unmanned obstacle avoidance system applied to hay harvesting to perform the above-described unmanned obstacle avoidance method applied to hay harvesting.

[0006] Based on the above, by combining the terrain features of the forage harvesting area with the operational parameters of unmanned driving, the cooperative adaptation relationship between the obstacle avoidance perception components and the harvesting components is adjusted and integrated to form cooperative adaptation parameters. This effectively improves the equipment's adaptability to different operating environments. Based on the cooperative adaptation parameters, a multi-sensor component synchronous acquisition process is initiated to obtain a comprehensive and detailed three-dimensional perception dataset, covering obstacle feature information, forage growth information, and ground condition information. A target coupling model is built, and avoidance spaces corresponding to different types of obstacles are divided. The spatial coordinates of the core harvesting area and avoidable areas are marked. By combining the coupling model data and cooperative adaptation parameters, a dynamic obstacle avoidance trajectory is planned and optimized, so that the obstacle avoidance trajectory not only conforms to the obstacle avoidance space but also meets the equipment's operational parameter requirements. This ensures that the equipment can efficiently complete the harvesting operation while avoiding obstacles. Finally, based on the optimized obstacle avoidance trajectory, the unmanned driving equipment executes commands, achieving precise adjustment of the equipment's driving direction, speed, and the operating position and frequency of the harvesting components. This greatly improves the automation, accuracy, and efficiency of forage harvesting operations, while reducing labor costs and operational risks. Attached Figure Description

[0007] Figure 1This is a schematic diagram of the execution flow of an unmanned obstacle avoidance method for forage harvesting provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of exemplary hardware and software components of an unmanned obstacle avoidance system for forage harvesting provided in an embodiment of the present invention. Detailed Implementation

[0009] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating an unmanned obstacle avoidance method for hay harvesting provided in one embodiment of the present invention. The following is a detailed description of this unmanned obstacle avoidance method for hay harvesting.

[0010] Step S110: Combine the terrain features of the forage harvesting area with the operating parameters of the unmanned operation, adjust the cooperative adaptation relationship between the obstacle avoidance perception component and the harvesting component, and integrate them to form cooperative adaptation parameters. The terrain features include ground softness, ground undulation amplitude, and ground flatness. The operating parameters include the lateral coverage size of the equipment in a single harvest, the driving speed, and the operating position of the harvesting component.

[0011] In this embodiment, unmanned harvesting operations are carried out in an open pasture planting area. First, soil samples are collected from different locations in the area using soil sampling equipment to obtain ground softness data. Using soil compaction meter readings as a quantitative indicator, a two-dimensional array G is formed. Each element G[i][j] in the array corresponds to the softness value at coordinate (i, j), in megapascals (MPa). A lidar scanner is used to scan the entire area, dividing it into 1m × 1m grids. The height difference between the highest and lowest points in each grid is calculated to obtain ground undulation data H, where H[i][j] represents the undulation of the grid at coordinate (i, j), in centimeters. By calculating the standard deviation of all height values ​​in each grid, ground flatness data P is obtained, where P[i][j] represents the flatness of the grid at coordinate (i, j), in centimeters. Regarding operating parameters, the lateral coverage dimension W of a single harvest is set as a fixed value based on the width of the harvester's header, in meters; the travel speed V is set to an adjustable range from A meters / second to B meters / second; and the operating position of the harvesting component is indicated by its height from the ground, L, in centimeters.

[0012] Next, a correlation model between terrain features and operational parameters is established to adjust the collaborative adaptation relationship between the obstacle avoidance sensing component and the harvesting component. For example, when the detected ground softness G[i][j] is greater than the threshold G0, the lidar scanning angle of the obstacle avoidance sensing component is adjusted downward by θ degrees, while the operating position L of the harvesting component is raised by C centimeters. The weights of each parameter are determined using the analytic hierarchy process (AHP), and the ground softness, ground undulation amplitude, ground flatness, lateral coverage size, driving speed, and harvesting component operating position are weighted and integrated to form a collaborative adaptation parameter. This collaborative adaptation parameter is a set of 8 sub-parameters, namely, lidar scanning angle α, camera focal length f, lateral coverage size correction coefficient k1, driving speed correction coefficient k2, harvesting component height correction coefficient k3, sensing component sampling frequency f_s, data fusion weight matrix W_m, and decision threshold matrix T.

[0013] Step S120: Based on the collaborative adaptation parameters, start the synchronous acquisition process of the multi-sensor components to capture obstacle feature information, grass growth information, and ground condition information in the grass harvesting area, and integrate them to form a three-dimensional perception dataset. The obstacle feature information includes outline size, material hardness, and movement state. The grass growth information includes plant distribution, coverage area, and plant height. The ground condition information includes soil compaction and pothole distribution.

[0014] Step S121: Activate the visual perception component, terrain perception component, material perception component, and soil perception component according to the collaborative adaptation parameters. Set the collection range, collection frequency, and data accuracy of each perception component according to the requirements of forage harvesting operations. Start the synchronous collection action of each perception component to obtain visual collection data, terrain collection data, material collection data, and soil collection data respectively.

[0015] Based on the collaborative adaptation parameters obtained in step S110, the visual perception component (high-definition camera), terrain perception component (LiDAR), material perception component (infrared sensor), and soil perception component (soil compaction sensor) are activated. The visual perception component's acquisition range is set to a rectangular area centered on the harvester's current position, covering a horizontal area of ​​W×k1 meters and a vertical area of ​​D meters. The acquisition frequency is set to f_s frames / second, and the data precision is set to a resolution of 1920×1080 pixels and a color depth of 24 bits. The terrain perception component's acquisition range covers a fan-shaped area E meters in front of the harvester and F meters on each side, with a scanning angle of α degrees and a point cloud density of G points per square meter. The acquisition frequency is synchronized with that of the visual perception component. The material perception component's acquisition range focuses on a rectangular area H meters in front of the harvester with a width of W×k1 meters. The spectral response range is set to 1 nanometer to J nanometers, the acquisition frequency is f_s times / second, and the material recognition accuracy is required to reach K% or higher. The soil sensing component's collection range is the area traversed by the harvester's tracks. The sampling interval is set to L centimeters, and the collection frequency is dynamically adjusted according to the travel speed V×k2. The measurement error is controlled within ±M MPa.

[0016] After the synchronous acquisition actions of each sensing component are initiated, the visual sensing component continuously outputs RGB image sequences to form visual acquisition data. The data format is that each frame contains 1920×1080×3 pixel values ​​and is stored as a JPEG file. The terrain sensing component outputs 3D point cloud data, with each point containing X, Y, and Z 3D coordinates and reflectance intensity values ​​to form terrain acquisition data, stored as a PCD file. The material sensing component outputs the infrared reflectance spectrum curve of each sampling point to form material acquisition data, with each spectrum curve containing reflectance values ​​at N wavelengths. The soil sensing component outputs soil compaction measurements to form soil acquisition data, with each data point containing location coordinates and the corresponding compaction value in megapascals (MPa).

[0017] Step S122: Import the visual acquisition data, terrain acquisition data, material acquisition data, and soil acquisition data into the multi-source data processing flow. Compare the data content point by point according to the spatial coordinates of the harvesting area, eliminate conflicting data segments from different data sources, retain the core information of obstacle outline, ground undulation, material hardness, soil compaction, and pasture distribution, and integrate them into a preliminary fusion dataset.

[0018] The four types of collected data obtained in step S121 are imported into a multi-source data processing system. This system first establishes a unified spatial coordinate system, with the initial position of the harvester as the origin, the X-axis along the harvesting direction, the Y-axis perpendicular to the harvesting direction, and the Z-axis perpendicular to the ground. Distortion correction and perspective transformation are performed on the visually collected data to convert the image pixel coordinates into three-dimensional coordinates in the spatial coordinate system. Noise reduction processing is performed on the terrain collected data to remove outliers and duplicates. Spectral normalization processing is performed on the material collected data to eliminate the influence of illumination variations. Interpolation processing is performed on the soil collected data to obtain a continuous distribution of soil compaction.

[0019] Information from different data sources is compared point-by-point according to spatial coordinates (X, Y). For a point at coordinates (X0, Y0), if visual data identifies it as an obstacle but terrain data does not detect a change in height, or if material data shows it as metallic but soil data shows it as high compaction, then a data conflict is identified. A voting method is used to eliminate conflicting data segments; if three or more data sources are consistent, the data point is retained; otherwise, it is marked as suspicious data. Core information including obstacle outlines (areas consistent between visual and terrain data), ground undulations (height changes in terrain data), material hardness (reflectivity characteristics of material data), soil compaction (measured values ​​in soil data), and pasture distribution (green vegetation areas in visual data) is retained. This information is organized into a three-dimensional array according to spatial coordinates. Each element of the array contains attributes such as coordinates (X, Y, Z), material category, soil compaction value, and vegetation cover, forming a preliminary fused dataset.

[0020] Step S123: Extract obstacle contour information from the preliminary fusion dataset, perform contour completion operation on the blurred edge area of ​​the visual acquisition data caused by grass occlusion, calculate the planar projection offset value of the obstacle position in the terrain acquisition data by combining the terrain slope and sensor view parameters, supplement the offset value to the corresponding spatial coordinate position of the obstacle contour information, and integrate to obtain the contour correction dataset.

[0021] Step S1231: Extract obstacle contour information and visual acquisition data from the preliminary fusion dataset, mark the blurred edge areas caused by grass occlusion in the visual acquisition data by spatial coordinates, record the spatial range and corresponding coordinate information of the blurred areas, integrate the marking results of the blurred areas according to the obstacle type, and form a blurred marking dataset.

[0022] Obstacle contour information is extracted from the preliminary fusion dataset and stored as a binary image, where the pixel value of the obstacle region is 1 and the background region is 0. Simultaneously, RGB images from the corresponding visual acquisition data are extracted, and the Canny edge detection algorithm is used to extract image edges and calculate edge gradient values. When the edge gradient value of a certain region is lower than a threshold T1 and the region is located within a known pasture distribution range, it is determined to be an edge blurring region caused by pasture occlusion. These blurring regions are labeled with spatial coordinates (X, Y), and the coordinates (Xmin, Ymin, Xmax, Ymax) of the minimum bounding rectangle of each blurring region are recorded. The blurring regions are then classified according to the obstacle type attribute (rigid obstacle / flexible obstacle) in the preliminary fusion dataset, forming a blurring labeled dataset. This blurring labeled dataset contains information such as the coordinate range, obstacle type, and blur degree (mean gradient value) of each blurring region.

[0023] Step S1232: Match the fuzzy marker dataset with the material acquisition data, and use the contour extension method to perform a completion operation on the obstacle contour in the fuzzy area. Extend the contour shape according to the different shapes of rigid and flexible obstacles. The completed contour fits the actual material characteristics of the obstacle and is integrated into a contour extension dataset.

[0024] The fuzzy labeled dataset obtained in step S1231 is spatially matched with the material acquisition data to obtain the material reflectance characteristics of each point within the fuzzy region. For rigid obstacles (such as rocks and metal piles), their material reflectance characteristics are characterized by high reflectance and steep spectral curves. A straight-line extension method is used to complete the contour: two feature points A(Xa, Ya) and B(Xb, Yb) are selected at the boundary of the fuzzy region, the direction vector of the line connecting the two points is calculated, and this direction is extended to the edge of the fuzzy region to form a straight contour segment. For flexible obstacles (such as plastic sheeting and haystacks), their material reflectance characteristics are characterized by low reflectance and gentle spectral curves. A curve extension method is used to complete the contour: three feature points C(Xc, Yc), D(Xd, Yd), and E(Xe, Ye) are selected at the boundary of the fuzzy region, and a quadratic Bézier curve is used to fit the three points to form a smooth curve, which is then extended to the edge of the fuzzy region. The completed contour matches the material characteristics of the obstacle; for example, the contour lines of metal obstacles are sharp, while the contour lines of fabric obstacles are soft. All the completed contour information is integrated according to spatial coordinates to form a contour extension dataset. This contour extension dataset contains information such as the coordinate sequence of the completed obstacle contour, contour type (straight line / curve), and material matching degree.

[0025] Step S1233: Extract terrain acquisition data and ground undulation information from the preliminary fusion dataset. Combine terrain slope and sensor view parameters to calculate the planar projection offset value of the obstacle position in the perception data due to ground undulation. Record the direction and specific amount of the offset. Integrate the offset calculation results according to the obstacle coordinates to form an offset calculation dataset.

[0026] The 3D point cloud information and ground undulation amplitude data H of the terrain acquisition data are extracted from the preliminary fusion dataset. For each obstacle, three points P1(X1, Y1, Z1), P2(X2, Y2, Z2), and P3(X3, Y3, Z3) at its bottom edge are selected. The terrain slope α is calculated by fitting the three points, with the formula α=arctan(ΔZ / ΔL), where ΔZ is the average height difference of the three points and ΔL is the average horizontal distance. Sensor view parameters include the installation height H_s of the lidar, the horizontal scanning angle θ_h, and the vertical scanning angle θ_v. A transformation model between the sensor coordinate system and the ground coordinate system is established based on these parameters. When there is a slope α on the ground, the actual position (X, Y, Z) of the obstacle will be offset by a planar projection in the sensor acquisition data. The offset direction is along the slope direction, and the offset amount Δd=Z×tan(α)×cos(θ_h), where Z is the obstacle height. The offset direction (expressed in angle, with 0° being the positive X-axis and increasing clockwise) and offset amount Δd (in meters) are calculated for each obstacle. The calculation results are integrated according to the obstacle coordinates (X, Y) to form an offset calculation dataset. This offset calculation dataset contains information such as obstacle ID, offset direction angle, offset amount Δd, and terrain slope α.

[0027] Step S1234: Import the contour extension dataset and offset calculation dataset into the obstacle information correction process, replace the original obstacle information in the visual acquisition data with the completed obstacle contour and the corrected position coordinates, and re-integrate all the information of the obstacle contour according to the spatial coordinates to form a preliminary correction dataset.

[0028] The contour extension dataset obtained in step S1232 and the offset calculation dataset obtained in step S1233 are associated through obstacle IDs. For each obstacle, the contour coordinates (X, Y) in the contour extension dataset are first corrected to (X + Δd × cos(θ), Y + Δd × sin(θ)) based on the offset direction angle and offset amount Δd in the offset calculation dataset, where θ is the offset direction angle. Then, the original obstacle contour information of the corresponding area in the visual acquisition data is replaced with the corrected contour coordinates, while retaining the height information Z in the terrain acquisition data. The contour coordinate sequence, height value, material characteristics, offset correction amount, and other information of the obstacle contour are reorganized according to the spatial coordinates (X, Y) to form a preliminary corrected dataset. This preliminary corrected dataset is a three-dimensional array, where each element corresponds to the obstacle attribute of a spatial coordinate point.

[0029] Step S1235: Extract hardness feature information from the material acquisition data, match the hardness feature information with the preliminary correction dataset, verify the fit between the corrected obstacle contour shape and the material hardness, fine-tune the detailed coordinates of the contour for areas that do not fit, so that the contour shape fits the material hardness, and form a verification correction dataset.

[0030] Hardness feature information is extracted from the material acquisition data. The material hardness index HI is calculated using the infrared reflectance spectrum curve: HI = ∑(λ_i × R_i) / ∑λ_i, where λ_i is the wavelength and R_i is the reflectance at the corresponding wavelength. HI values ​​are divided into five levels: HI ≥ 0.8 for ultra-hard materials, 0.6 ≤ HI < 0.8 for hard materials, 0.4 ≤ HI < 0.6 for medium-hard materials, 0.2 ≤ HI < 0.4 for soft materials, and HI < 0.2 for ultra-soft materials. The hardness feature information is matched with the preliminary correction dataset using spatial coordinates. For hard material obstacles, the profile should have sharp angles and straight edges; for soft material obstacles, the profile should have smooth curved edges. The fit C between the profile shape and the material hardness is calculated: C = (number of matched profile segments / total number of profile segments) × 100%. When C < 80%, the contour coordinates of the misfit regions are fine-tuned: the curved edges of hard material obstacles are replaced with straight edges, and the sharp angles of soft material obstacles are replaced with rounded transitions (the radius R is set according to the HI value; the smaller the HI, the larger the R). The fine-tuned contour shape matches the material hardness characteristics, forming a verification and correction dataset. This verification and correction dataset adds two attributes: hardness level and fit C.

[0031] Step S1236: Extract the contour information of all obstacles from the verification and correction dataset, mark the corner positions and dimensions of rigid obstacles one by one, mark the edge curvature of flexible obstacles one by one, supplement various detailed feature information of obstacle contours, integrate the marking results of detailed features according to obstacle coordinates, and form a detailed supplementary dataset.

[0032] The contour coordinate sequence of each obstacle is extracted from the validation and correction dataset. For rigid obstacles (hardness level of ultra-hard or hard material), corner detection algorithms (such as the Harris algorithm) are used to identify the positions of the corners on the contour, and the corner coordinates (Xc, Yc) and corner angle θ_c (the angle between two adjacent contour segments) are recorded. At the same time, the protrusion dimension d_c of the corner (the maximum distance from the corner point to the edge of the contour) is measured. For flexible obstacles (hardness level of soft or ultra-soft material), a curve fitting algorithm is used to calculate the radian r of the contour edge, r=(Δs^3) / (6×Δx), where Δs is the arc length of the curve segment and Δx is the chord length of the line connecting the two endpoints of the curve segment. Supplementing the detailed feature information of the obstacle contour also includes the roughness of the contour edge (represented by the standard deviation of the distance from the point on the contour line to the fitted line) and the surface texture direction (calculated by the gray-level co-occurrence matrix). These detailed features are integrated according to the obstacle coordinates (X, Y) to form a detailed supplementary dataset. This detailed supplementary dataset contains information such as the list of edges (rigid obstacles) or radii (flexible obstacles) for each obstacle, roughness values, and texture orientation angles.

[0033] Step S1237: Merge the detailed supplementary dataset with the preliminary correction dataset, add detailed features to the corrected obstacle contour information, improve the morphological description of the obstacle contour, and reintegrate all the information of the obstacle contour according to spatial coordinates to form the contour optimization dataset.

[0034] The detailed supplementary dataset obtained in step S1236 is fused with the preliminary corrected dataset obtained in step S1234. Detailed features (edge ​​position, curvature, roughness, etc.) are added to the corresponding obstacle contour information using spatial coordinates. For example, in the contour coordinate sequence of a rigid obstacle, edge angle and prominent size attributes are added to the edge point coordinates; in the contour coordinate sequence of a flexible obstacle, curvature value attributes are added to the curve segments. Simultaneously, the smoothness of the contour line is adjusted according to the roughness value; a higher roughness value retains more detail fluctuations, while a lower roughness value further smooths the contour line. The fused obstacle contour information is reorganized according to spatial coordinates (X, Y), including the contour coordinate sequence, height Z, material hardness grade, edge / curvature parameters, roughness, texture direction, etc., to form a contour optimization dataset. This contour optimization dataset provides a more refined and accurate description of the obstacle contour.

[0035] Step S1238: Extract color feature information from the visual acquisition data, match the color feature information with the contour optimization dataset, correct the color deviation in the contour optimization dataset, enhance the color contrast between obstacles and grass, so that the visual presentation of the obstacle contour is clearer, and integrate them into a color correction dataset.

[0036] Color feature information is extracted from the RGB images of the visual acquisition data, and the HSV color space parameters (hue (H), saturation (S), and lightness (V)) of each pixel are calculated. The color feature information is matched with the contour optimization dataset according to spatial coordinates to obtain the mean H, S, and V values ​​of the obstacle region and the pasture region. For the obstacle contour edge region, if there is a color deviation due to lighting or pasture occlusion (e.g., the obstacle color is similar to the pasture color), the H value (shifted by ΔH to increase the difference in H value between the obstacle and the pasture), S value (increased by ΔS to improve color vividness), and V value (adjusted by ΔV to distinguish the obstacle region's brightness from the background) are adjusted. For example, when the obstacle is brown rock (H=30°) and the pasture is green (H=120°), the H value of the rock edge is adjusted to 30°, and the S value is increased to enhance the color contrast between the rock and the pasture. The corrected contour optimization dataset contains the adjusted HSV color parameters, forming a color correction dataset that makes the obstacle contour visually clearer and more distinguishable.

[0037] Step S1239: Import the color correction dataset and the preliminary fusion dataset into the data integration process, retain the pasture growth information and ground condition information in the preliminary fusion dataset, remove redundant and duplicate data generated during the contour correction process, and integrate all information according to spatial coordinates to form the contour correction base dataset.

[0038] The color-corrected dataset obtained in step S1238 is integrated with the preliminary fusion dataset obtained in step S122. A data fusion algorithm is used to retain the pasture growth information (plant distribution, coverage, plant height) and ground condition information (soil compaction, pothole distribution) from the preliminary fusion dataset. Simultaneously, obstacle contour information (optimized contour coordinates, detailed features, and color parameters) from the color-corrected dataset is added to the corresponding spatial coordinate positions. During the integration process, duplicate records of the same obstacle information and redundant intermediate calculation results (such as uncorrected original contours) are removed by comparing data timestamps and spatial coordinates. The integrated information is organized according to spatial coordinates (X, Y), with each coordinate point containing obstacle attributes (if present), pasture growth attributes, and ground condition attributes, forming a contour correction base dataset. This contour correction base dataset optimizes and supplements the preliminary fusion dataset, focusing on improving the accuracy of obstacle contour information.

[0039] Step S12310: Integrate all information of the corrected obstacle contour information, position coordinates, and detailed features in the contour correction base dataset, unify the data format and the correspondence between spatial coordinates, and form a contour correction dataset.

[0040] The contour correction dataset obtained in step S1239 is standardized by converting all data into a unified JSON format. Obstacle contour information is represented as an array of coordinate points [{"X":x1, "Y":y1, "Z":z1}, ...], and detail features are represented as key-value pairs {"corner_points":[{"X":xc1, "Y":yc1, "angle":θ1, "size":d1}, ...], "roughness":r, ...}. The location coordinates use a unified spatial coordinate system (consistent with step S122). Data associations are established through spatial coordinate indexing to ensure accurate correspondence between the contour information, location coordinates, and detail features (edges / radians, roughness, texture, color) of each obstacle. A data integrity check is performed, and missing coordinate points and attribute values ​​are supplemented (using nearest neighbor interpolation), ultimately forming the contour correction dataset.

[0041] Step S124: Extract pasture growth-related information from the contour correction dataset, perform spatial labeling operation on the pasture growth area, divide the pasture growth area according to the coordinate range of plant distribution, record the spatial coordinates of pasture plant height and coverage area in each area, and integrate all labeling results according to the spatial block of the harvest area to form a pasture partition dataset.

[0042] From the contour correction dataset obtained in step S12310, extract pasture growth-related information, including a binarized image of plant distribution (1 indicates pasture presence, 0 indicates no pasture), polygon boundary coordinates of the coverage area, and 3D point cloud data of plant height. A region growing algorithm is used to perform spatial labeling on the pasture growth area: starting from the seed point (the coordinates of a known pasture area), adjacent pixels that satisfy the plant height threshold (>H_min) and vegetation index threshold (NDVI>N_min) are merged into a connected region. Each connected region is considered an independent pasture growth area, and its minimum bounding rectangle coordinates (Xmin, Ymin, Xmax, Ymax) and polygon boundary coordinate sequence are recorded.

[0043] For each pasture growth area, plant height is calculated using point cloud data. The maximum Z-coordinate of all points within the area is taken as the average plant height H_avg for that area. Simultaneously, the standard deviation of plant height σ_h is calculated to reflect the uniformity of plant height. Coverage is quantified by calculating the area S and perimeter L of the polygon boundary coordinates. All labeled results are integrated by spatially dividing the harvested area into blocks (e.g., a 10m × 10m grid). Each block contains information such as the ID, coordinate range, average plant height, standard deviation of plant height, coverage area, and perimeter of all pasture growth areas within that block, forming a pasture partition dataset. This pasture partition dataset uses a grid index structure, allowing for quick querying of pasture growth attributes at any spatial location.

[0044] Step S125: Capture moving obstacles within the pasture harvesting area based on the contour correction dataset, continuously record the movement trajectory segments and real-time movement status of the moving obstacles, match the material acquisition data with the spatial coordinates of the moving obstacles, label the corresponding material hardness attribute for each moving obstacle, and integrate all the information of the moving obstacles according to the time series to form a moving obstacle trajectory dataset.

[0045] From the contour correction dataset obtained in step S12310, moving obstacles are captured using the continuous frame difference method: the obstacle contour information of two adjacent frames is differentially calculated, and when the displacement Δs of the center coordinates (Xc, Yc) of an obstacle within the time interval Δt is greater than Δs_min, it is determined to be a moving obstacle. A unique ID is assigned to each moving obstacle, and its motion trajectory segments are continuously recorded. The center coordinates (Xc(t), Yc(t), Zc(t)) are recorded once every Δt interval to form a sequence of trajectory points.

[0046] The real-time motion state is calculated as follows: velocity V(t) = Δs / Δt, direction angle θ(t) = arctan[(Yc(t) - Yc(t-1)) / (Xc(t) - Xc(t-1))], and acceleration a(t) = (V(t) - V(t-1)) / Δt. The material acquisition data is matched with the real-time spatial coordinates (Xc(t), Yc(t)) of the moving obstacle to obtain the material hardness index HI at the corresponding location, which is labeled as the material hardness attribute of the moving obstacle at time t. Information such as the ID, trajectory point sequence, velocity, direction angle, acceleration, and material hardness index of each moving obstacle is integrated according to the time series t to form a moving obstacle trajectory dataset. This moving obstacle trajectory dataset uses a timestamp index, allowing for the traceability of the moving obstacle's state at any given time.

[0047] Step S126: Associate the pasture partition dataset with the acquisition parameters of each sensing component. Adjust the acquisition frequency of the visual sensing component, terrain sensing component, material sensing component, and soil sensing component according to the plant distribution coordinate range in the pasture partition dataset. Increase the acquisition frequency in the core area of ​​pasture plant distribution and the area around moving obstacles, and decrease the acquisition frequency in the edge area of ​​pasture plant distribution, and integrate them into a frequency-optimized dataset.

[0048] Associate the forage partition dataset obtained in step S124 with the current acquisition frequency parameters of each sensing component. Define the core area of the forage as the area where the average plant height H_avg > H_core and the covered area S > S_core, and the edge area as the area where H_avg < H_edge or S < S_edge (H_core > H_edge, S_core > S_edge). The area around the moving obstacle is defined as a circular area with the center coordinates (Xc(t), Yc(t)) of the moving obstacle as the center and a radius R, and R is dynamically adjusted according to the speed V(t) of the moving obstacle. The larger V(t) is, the larger R is.

[0049] For the visual sensing component, the acquisition frequencies in the core area and the area around the moving obstacle are increased to f_s × k_high, and decreased to f_s × k_low in the edge area (k_high > 1, k_low < 1); the point cloud density of the terrain sensing component is increased to G × k_high in the core area and the surrounding area, and decreased to G × k_low in the edge area; the acquisition frequencies of the material sensing component and the soil sensing component are adjusted similarly. Integrate the acquisition frequency parameters, area coordinate ranges, effective times, etc. of each sensing component in different areas after adjustment to form a frequency optimization dataset. This frequency optimization dataset dynamically optimizes the acquisition strategy of the sensing components through real-time updated area information, reducing the overall data volume while ensuring the data quality in key areas.

[0050] Step S127: Extract the distribution information of soil compactness from the soil acquisition data, mark the spatial range where the equipment is prone to sink in the forage harvesting area according to the distribution characteristics of soil compactness, extract the spatial coordinates of the pothole distribution from the terrain acquisition data and match them with the area prone to sink, supplement the terrain feature information of the area prone to sink, and integrate to obtain the ground risk dataset.

[0051] Extract the soil compactness values from the soil acquisition data, and obtain the soil compactness distribution raster data of the entire harvesting area through Kriging interpolation. The raster cell size is 0.5 m × 0.5 m. Set the soil compactness threshold C_sink. When the compactness value of a certain raster < C_sink, mark it as an area prone to sink, and record its polygon boundary coordinates. Extract the pothole distribution information from the terrain acquisition data, and identify the pothole area through regional minimum filtering: when the Z coordinate of a certain point is lower than the average Z coordinate of its neighborhood (3×3 raster) by more than the threshold ΔZ_pit, it is determined as a pothole point, and the pothole area is formed through region growing method, and record its boundary coordinates and depth value D_pit (the difference between the lowest point of the pothole and the average height of the neighborhood).

[0052] Spatial overlay analysis was performed on subsidence-prone areas and pothole distribution areas to calculate the area ratio of the overlapping region. Areas with a ratio greater than 50% were identified as high-risk areas. Pothole depths (D_pit) and soil compaction values ​​(C) were also recorded. Additional topographic features of subsidence-prone areas included slope α, aspect β, and surface roughness. Finally, the coordinate range, soil compaction, pothole depth, and topographic parameters of all subsidence-prone areas were integrated to form a ground risk dataset.

[0053] Step S128: Import the contour correction dataset, pasture partition dataset, moving obstacle trajectory dataset, frequency optimization dataset, and ground risk dataset into the data integration process, unify the data format, data dimensions, and spatial coordinate correspondence of all datasets, and integrate all data content according to the spatial block division of pasture harvesting areas to form an integrated perception dataset.

[0054] The contour correction dataset from step S12310, the pasture zoning dataset from step S124, the mobile obstacle trajectory dataset from step S125, the frequency optimization dataset from step S126, and the ground risk dataset from step S127 are imported into the data integration system. First, the data format is unified to JSON, with spatial coordinates (X, Y) as the primary key and time dimension (t) as the secondary key (for mobile data). Coordinate transformation is used to map all datasets to the same spatial coordinate system, and nearest neighbor interpolation is used to handle resolution differences between the different datasets.

[0055] Data is integrated according to the spatial division of the forage harvesting area (a 10m × 10m grid consistent with step S124). Each block contains obstacle information (static + dynamic), forage growth information, ground risk information, and sensing frequency parameters. The data is checked for consistency, duplicate records are deleted, and missing values ​​are added (using the mean within the block or the value of neighboring blocks), ultimately forming an integrated sensing dataset.

[0056] Step S129: Capture obstacles that are obscured by grass and located in potholes from the integrated perception dataset, extract visual acquisition data, terrain acquisition data, and material acquisition data of the corresponding area of ​​the obstacle, combine the feature information of various types of data to infer the outline size and material hardness attributes of the corresponding obstacle, and integrate all the information of the corresponding obstacle according to spatial coordinates to form a hidden obstacle supplementary dataset.

[0057] Step S1291: Extract the pasture zoning dataset and ground risk dataset from the integrated sensing dataset, mark the spatial boundary coordinates of the core area of ​​pasture plant distribution and the area of ​​pothole distribution, record the relevant data of pasture plant height and pothole depth in each area, integrate the marking results according to spatial blocks, and form a hidden area marking dataset.

[0058] The core region boundary coordinates (Xmin, Ymin, Xmax, Ymax) and average plant height H_avg from the integrated sensing dataset, and the boundary coordinates and pit depth D_pit from the ground risk dataset, are extracted. These regions are defined as areas where hidden obstacles may exist (obscured by pasture or located within pits). For each region, the H_avg, plant height standard deviation σ_h, and coverage area S of the pasture core region, and the D_pit, slope α, and boundary coordinate sequence of the pit region are recorded. This labeling information is integrated into 10m × 10m spatial blocks. Each block contains the type (pasture core region / pit region), boundary coordinates, and feature parameters (H_avg, D_pit, etc.) of all hidden regions within that block, forming a hidden region labeling dataset.

[0059] Step S1292: Match the integrated perception dataset with the hidden area labeling dataset, filter out the visual acquisition data, terrain acquisition data, and material acquisition data corresponding to the core distribution area of ​​pasture plants and the distribution area of ​​potholes, and integrate the contents of various acquisition data according to the region to form the hidden area acquisition dataset.

[0060] Based on the region boundary coordinates in the hidden region labeling dataset, visual acquisition data (RGB images), terrain acquisition data (point clouds), and material acquisition data (reflectance spectra) within the corresponding spatial range are selected from the integrated perceptual dataset. For the pasture core region, pixel values ​​of the corresponding region in the image, Z-coordinate data in the point cloud, and reflectance curves in the spectrum are extracted; for the pitted region, pixel values ​​within the pit boundaries in the image, depth data in the point cloud, and material features in the spectrum are extracted. These acquired data are categorized and integrated according to region type (pasture core region / pitted region) and region ID, with each region corresponding to a data subset containing multi-source original acquisition data for that region, forming the hidden region acquisition dataset.

[0061] Step S1293: Extract terrain data from the hidden area data collection dataset, combine it with the numerical characteristics of ground elevation changes, infer the location of completely obscured obstacles in the core area of ​​pasture vegetation distribution, remove the elevation change interference caused by the height of the pasture itself, mark the spatial coordinates of the inferred obstacles, and form an obscured obstacle location dataset.

[0062] Topographic point cloud data of the core pasture area was extracted from the hidden area dataset, and the ground elevation datum Z_base (using the mean Z coordinates of non-vegetated points within the area) was calculated. For each point (X, Y, Z) in the point cloud, the elevation difference ΔZ = Z - Z_base was calculated. When ΔZ > H_avg + 3σ_h (considering the maximum possible value of pasture height), it was identified as a potential obstacle point (excluding interference from the height of the pasture itself). A clustering algorithm (such as DBSCAN) was used to aggregate these potential obstacle points into obstacle candidate regions, and the center coordinates (Xc, Yc) of each region were used as the predicted obstacle location coordinates. These coordinates were labeled, and the corresponding elevation difference ΔZ and the number of cluster points (reflecting the size of the obstacle) were recorded to form an occlusion obstacle location dataset.

[0063] Step S1294: Match the occlusion obstacle location dataset with the hidden area acquisition dataset, combine the hardness features of the material acquisition data with the color residue information of the visual acquisition data, infer the outline size and material hardness attributes of the obstacles occluded by the pasture at each location, and integrate the inference results by coordinate to form an occlusion obstacle feature dataset.

[0064] The inferred coordinates (Xc, Yc) of the occluded obstacle location dataset are matched with the material acquisition data and visual acquisition data in the hidden region acquisition dataset to obtain the infrared reflectance spectrum curve and RGB pixel values ​​at that location. The material hardness index HI is calculated using the spectrum curve (method as in steps S1235), and the material hardness attribute (extra-hard / hard / medium-hard / soft / extra-soft) is determined based on the HI value. Color residue information at that location is extracted from the visual acquisition data (even if occluded, there may be edge pixels or color through gaps), and the contour shape is inferred by combining it with the material hardness attribute: the contour of a hard material obstacle may be a regular polygon, while the contour of a soft material obstacle may be an irregular shape.

[0065] The contour dimensions are calculated using the bounding boxes of the clustered point cloud: minimum X-coordinate Xmin, maximum X-coordinate Xmax, minimum Y-coordinate Ymin, and maximum Y-coordinate Ymax. The contour width W = Xmax - Xmin, the length L = Ymax - Ymin, and the height H = ΔZ. The inferred contour dimensions (W, L, H), material hardness attributes, and coordinates (Xc, Yc) are integrated position by position to form an occlusion obstacle feature dataset.

[0066] Step S1295: Extract terrain data from the hidden area data collection dataset, combine the spatial coordinate features of the pothole morphology to infer the actual location of obstacles within the pothole distribution area, distinguish the terrain feature differences between obstacles and pothole edges, mark the inferred spatial coordinates of obstacles, and form a pothole obstacle location dataset.

[0067] Topographic point cloud data of pothole areas were extracted from the hidden region dataset, given the boundary coordinates and depth D_pit of the pothole areas. The relative elevation ΔZ' = Z - Z_pit_min (where Z_pit_min is the Z-coordinate of the lowest point in the pothole) was calculated for each point within the pothole area. Points were considered potential obstacles when ΔZ' > ΔZ_pit_th (the obstacle detection threshold within the pothole, set according to the pothole depth D_pit, typically D_pit × 0.3). Morphological filtering was used to distinguish obstacles from pothole edges: the elevation changes at pothole edges are continuous and distributed along the pothole boundary, while the elevation changes of obstacles are concentrated and unrelated to the pothole boundary. A region growing method was used to extract obstacle regions from potential obstacle points, and their center coordinates (Xc, Yc) were labeled to form a pothole obstacle location dataset.

[0068] Step S1296: Match the pit obstacle location dataset with the hidden area collection dataset, supplement the hardness attribute of the material collection data with the compaction difference information of the soil collection data, infer the material hardness attribute of the obstacle in the pit distribution area at each location, integrate the inference results by coordinate, and form a pit obstacle feature dataset.

[0069] The coordinates (Xc, Yc) of the pothole obstacle location dataset are matched with the material and soil data from the hidden region dataset to obtain the infrared reflectance spectrum (HI value) and soil compaction value C at that location. If C > C_sink (the sinking threshold), the high ΔZ' at that location is more likely an obstacle than a loose soil pile. The material hardness attribute is determined by combining the HI and C values: when HI > 0.6 and C > C_sink, it is determined to be a hard material obstacle; when 0.2 ≤ HI ≤ 0.6 and C > C_sink, it is determined to be a medium-hard material obstacle; when HI < 0.2 or C ≤ C_sink, it may be a soft obstacle or a loose soil pile (further confirmation with visual data is required). The inferred material hardness attribute, contour size (calculated from point cloud bounding boxes), and coordinates (Xc, Yc) are recorded for each location to form a pothole obstacle feature dataset.

[0070] Step S1297: Import the occlusion obstacle feature dataset and the pothole obstacle feature dataset into the data integration process, integrate the location coordinates, outline size, and material hardness attributes of all hidden obstacles, and integrate all information by dividing it into blocks according to the spatial coordinates of the pasture harvesting area to form the basic dataset of hidden obstacles.

[0071] The occlusion obstacle feature dataset from step S1294 and the pothole obstacle feature dataset from step S1296 are integrated and standardized in format, including hidden obstacle ID, type (occlusion / pothole), location coordinates (Xc, Yc, Zc), contour dimensions (W, L, H), material hardness grade, and confidence level (based on data integrity score, 0-100). The data is organized into 10m × 10m spatial blocks, with each block containing information on all hidden obstacles within that block, forming the basic hidden obstacle dataset.

[0072] Step S1298: Extract other obstacle information from the integrated perception dataset within the working area, match the other obstacle information with the hidden obstacle base dataset, adjust the outline size and position coordinates of the obstacles in the hidden obstacle base dataset according to the spatial distribution of other obstacles, avoid conflicts with the spatial distribution of other obstacles, and form a hidden obstacle correction dataset.

[0073] Extract identified obstacle information (non-hidden obstacles) from the integrated perception dataset, including their location coordinates and outline dimensions. Perform spatial conflict detection on this obstacle information and the basic hidden obstacle dataset: calculate the center distance d between the hidden obstacle and the known obstacle; if d < (W1 / 2 + W2 / 2) (W1 and W2 are the widths of the two obstacles), a spatial conflict is determined. For conflicting hidden obstacles, adjust their location coordinates according to the outline of the known obstacle (move Δd away from the known obstacle, Δd = (W1 / 2 + W2 / 2) - d + 0.5 meters) or reduce their outline dimensions (proportionally reduce W and L, keeping H unchanged). The adjusted hidden obstacle has no spatial overlap with the known obstacle, forming a corrected hidden obstacle dataset.

[0074] Step S1299: Extract all inferred obstacle information from the hidden obstacle correction dataset, label the data source for the inferred obstacle outline size and material hardness attributes, mark the spatial coordinates of the uncertain areas in the data inference, integrate the labeling results by coordinates, and form a hidden obstacle labeling dataset.

[0075] The contour dimensions (W, L, H) and material hardness attributes of each hidden obstacle are extracted from the hidden obstacle correction dataset, and their data sources are labeled: contour dimensions are from terrain point cloud data, and material hardness attributes are from material reflectance spectrum data. For regions with incomplete data (such as those with missing point clouds leading to large errors in contour dimension calculation), uncertain regions are marked with dashed contours, and the coordinate range and error range of these regions are recorded (e.g., error ±ΔW for W, error ±ΔH for H). These labeled information are then integrated according to spatial coordinates to form a hidden obstacle labeled dataset.

[0076] Step S12910: Merge the hidden obstacle correction dataset and the hidden obstacle labeling dataset, integrate the location coordinates, outline dimensions, material hardness attributes, and data source information of all hidden obstacles, standardize the data format and correspondence with spatial coordinates, and obtain the hidden obstacle supplementary dataset.

[0077] The obstacle attribute information from the hidden obstacle correction dataset is fused with the source and uncertainty markers from the hidden obstacle label dataset to form a supplementary hidden obstacle dataset. This supplementary dataset contains complete information for each hidden obstacle: ID, type, location coordinates, precise contour dimensions (solid line portion), uncertain contour range (dashed line portion), material hardness grade, data source (point cloud / spectral / soil data), and confidence score. The data format is consistent with the integrated sensing dataset, with one-to-one spatial coordinate correspondence, allowing it to be directly added to the corresponding location in the integrated sensing dataset.

[0078] Step S1210: Import the integrated perception dataset and the hidden obstacle supplementary dataset into the spatial coordinate matching process, match the obstacle feature information, pasture growth information and ground condition information of the same spatial location point by point, supplement the hidden obstacle information to the corresponding position of the integrated perception dataset, and integrate to obtain the three-dimensional perception dataset.

[0079] The integrated sensing dataset obtained in step S128 is spatially matched with the hidden obstacle supplementary dataset obtained in step S12910. For each spatial coordinate point (X, Y), if obstacle information exists in the hidden obstacle supplementary dataset, it is added to the corresponding position in the integrated sensing dataset to supplement obstacle feature information (outline, material, hardness, etc.). Simultaneously, it is ensured that the pasture growth information (plant height, coverage) and ground condition information (soil compaction, potholes) at the same location are consistent with the newly added obstacle information (e.g., the pasture coverage at the obstacle location is set to 0). The matched dataset is then validated to ensure no data conflicts or omissions, ultimately forming a 3D sensing dataset. This 3D sensing dataset is a complete 3D environment model containing obstacles (visible + hidden), pasture, and ground condition.

[0080] Step S130: Based on the 3D perception dataset, build a target coupling model, divide the avoidance space corresponding to different types of obstacles in the target coupling model, mark the spatial coordinates of the core harvesting area and the avoidable area of ​​pasture harvesting, and integrate them to form coupling model data.

[0081] Step S131: Extract obstacle feature information from the 3D perception dataset, classify obstacles into rigid obstacles and flexible obstacles according to material hardness and contour size, record the contour size and edge distribution of rigid obstacles for each obstacle, record the contour boundary and deformation buffer space requirements of flexible obstacles for each obstacle, integrate all information according to obstacle type to form an obstacle classification dataset.

[0082] Extract the feature information of all obstacles from the three-dimensional perception dataset obtained in step S1210, including the material hardness index HI, the contour dimensions (W, L, H), the edge angle distribution (for rigid obstacles), and the contour boundary curve (for flexible obstacles). Classify the obstacle types according to the material hardness and contour dimensions: When HI ≥ 0.6 and W ≥ W_rigid or H ≥ H_rigid, it is classified as a rigid obstacle; when HI < 0.6 or W < W_rigid and H < H_rigid, it is classified as a flexible obstacle (W_rigid and H_rigid are the determination thresholds for rigid obstacles, which are set according to the device parameters).

[0083] For rigid obstacles, record the contour dimensions (accurate to centimeters), the position coordinates of the edge angle distribution (Xc_i, Yc_i), and the edge angle θ_i for each obstacle; for flexible obstacles, record the curve equation parameters of the contour boundary (such as the control point coordinates of the Bezier curve) and the deformation buffer space requirements (set according to the material hardness, the smaller the HI, the greater the requirement, for example, the buffer space radius for soft material obstacles is R_soft = 0.5 meters, and for ultra-soft materials is R_ultrasoft = 1.0 meters). Integrate this information according to the obstacle type (rigid / flexible) and ID to form an obstacle classification dataset.

[0084] Step S132: Extract the forage growth information from the three-dimensional perception dataset, screen the core forage growth areas within the forage harvesting area according to the plant distribution and coverage range, mark the spatial coordinates of the core forage growth areas region by region and draw complete boundary lines, determine the overall spatial range of the core harvesting area, and integrate the information of the core harvesting area by spatial partitioning to form a core area dataset.

[0085] Extract the forage growth information from the three-dimensional perception dataset, including the plant distribution density ρ (plants per square meter), the coverage polygon, and the average plant height H_avg. Set the screening conditions for the core forage growth areas: ρ > ρ_core, H_avg > H_core, and the coverage area S > S_core (ρ_core, H_core, and S_core are the determination thresholds for the core area). Merge the forage areas that meet the conditions (merge when the distance between adjacent areas < d_merge) to form the core forage growth areas.

[0086] Mark the spatial coordinates of the core areas region by region, draw complete boundary lines by fitting the polygon boundary points (simplify the boundary using the Douglas-Peucker algorithm), calculate the total area, perimeter, and centroid coordinates of the core areas. Take the union of all core areas as the overall spatial range of the core harvesting area, and record the area proportion, average plant height, density, etc. of the core areas within each 10 m × 10 m spatial block to form a core area dataset.

[0087] Step S133: Extract ground condition information from the three-dimensional perception dataset, match the distribution information of soil compaction with the information of ground undulation amplitude, label the bearing capacity information of the ground in the pasture harvesting area with spatial coordinates, draw the distribution boundary of ground bearing capacity and mark the corresponding spatial coordinates, and integrate them into a ground bearing capacity dataset.

[0088] Soil compaction C and ground undulation amplitude H (H[i][j] in step S110) are extracted from the 3D perception dataset and matched according to spatial coordinates. The ground bearing capacity index BCI is defined as 0.6×C_norm+0.4×(1-H_norm), where C_norm is the normalized value of soil compaction (C / C_max), H_norm is the normalized value of ground undulation amplitude (H / H_max), and the BCI value ranges from 0 to 1, with a larger value indicating stronger bearing capacity.

[0089] The bearing capacity index (BCI) is calculated for each spatial coordinate (X, Y). The ground surface is divided into five bearing capacity levels based on the BCI: BCI ≥ 0.8 is extremely strong, 0.6 ≤ BCI < 0.8 is strong, 0.4 ≤ BCI < 0.6 is moderate, 0.2 ≤ BCI < 0.4 is weak, and BCI < 0.2 is extremely weak. The distribution boundaries of each bearing capacity level are plotted (using contour lines), boundary coordinates are marked, and the area, proportion, and spatial coordinate range of each level are integrated to form a ground bearing capacity dataset.

[0090] Step S134: Based on the obstacle classification dataset, divide the avoidance space corresponding to rigid obstacles and flexible obstacles. Expand the fixed avoidance distance outward according to the outline size of the rigid obstacle and draw the boundary coordinates of the avoidance space. Set the minimum avoidance distance according to the deformation buffer space requirements of the flexible obstacle and draw the boundary coordinates of the avoidance space. Integrate the avoidance space information for each obstacle to form an obstacle avoidance dataset.

[0091] Step S1341: Extract all relevant information about rigid obstacles from the obstacle classification dataset, record the length, width, and height of each rigid obstacle, accurately mark the spatial coordinates of the outermost contour point of the rigid obstacle, and integrate the feature information of the rigid obstacles according to the obstacle number to form a rigid contour dataset.

[0092] Extract the ID, contour dimensions (length L, width W, height H), and corner distribution coordinates of rigid obstacles from the obstacle classification dataset obtained in step S131. For each rigid obstacle, find the outermost contour points through its contour coordinate sequence, i.e., the maximum / minimum points in the X direction (Xmax, Xmin) and the maximum / minimum points in the Y direction (Ymax, Ymin). These points constitute the minimum bounding rectangle boundary of the obstacle. Record the spatial coordinates (Xmax, Ymax, Z), (Xmin, Ymin, Z), etc. of these outermost contour points for each obstacle, as well as the specific values ​​of L, W, and H, and integrate them according to the obstacle number to form a rigid contour dataset.

[0093] Step S1342: Based on the rigid contour dataset, extend a fixed avoidance distance outward for each rigid obstacle according to the operational requirements of the unmanned driving equipment, draw the boundary coordinates of the avoidance space for each obstacle, record the range, area and shape of the avoidance space for each rigid obstacle, integrate the information according to the obstacle number, and form a rigid avoidance dataset.

[0094] Based on the safety parameters of the autonomous driving equipment, a fixed avoidance distance D_rigid (e.g., 1.5 meters) is set for rigid obstacles. For each obstacle in the rigid contour dataset, its minimum bounding rectangle boundary is extended outward by D_rigid to form a new rectangular boundary, which is the avoidance space. If the obstacle is irregularly shaped (non-rectangular), a buffer algorithm (offset) is used to offset the contour boundary outward by a distance of D_rigid to form a polygonal avoidance space.

[0095] Draw the boundary coordinate sequence of the avoidance space, and calculate its area (using the polygon area formula) and shape parameters (such as rectangularity and perimeter). Record the boundary coordinates, area, shape parameters, and corresponding obstacle ID of the avoidance space for each obstacle to form a rigid avoidance dataset.

[0096] Step S1343: Extract all relevant information about flexible obstacles from the obstacle classification dataset, record the contour boundary and spatial requirements of deformation buffer for each flexible obstacle, mark the specific spatial range of deformation buffer of flexible obstacles in combination with material hardness, and integrate the feature information of flexible obstacles by obstacle number to form a flexible characteristic dataset.

[0097] Extract the ID, contour boundary curve parameters, material hardness index HI, and deformation buffer space requirement of flexible obstacles from the obstacle classification dataset obtained in step S131. Determine the deformation buffer space range based on the HI value: when 0.4 ≤ HI < 0.6 (medium hard), the buffer radius R = 0.3 meters; when 0.2 ≤ HI < 0.4 (soft), R = 0.5 meters; when HI < 0.2 (ultra-soft), R = 0.8 meters. Record the contour boundary curve equation (e.g., quadratic curve parameters a, b, c), R value, and HI value for each obstacle, and integrate them by obstacle number to form a flexible characteristic dataset.

[0098] Step S1344: Based on the flexible characteristic dataset, set a minimum avoidance distance for each flexible obstacle. The set minimum avoidance distance covers the deformation buffer space of the flexible obstacle. Draw the boundary coordinates of the avoidance space for each obstacle, record the range area of ​​the avoidance space for each flexible obstacle, and integrate the information according to the obstacle number to form a flexible avoidance dataset.

[0099] For each obstacle in the flexible characteristic dataset, a minimum avoidance distance D_flex = R + 0.2 meters is set (0.2 meters is the equipment safety margin). A buffer algorithm is used to offset the outline boundary of the flexible obstacle outward by a distance D_flex, forming an avoidance space (usually an irregular polygon). The boundary coordinate sequence of this avoidance space is plotted, the area is calculated, and the boundary coordinates, area, D_flex value, and corresponding obstacle ID are recorded for each obstacle, forming a flexible avoidance dataset.

[0100] Step S1345: Import the rigid obstacle avoidance dataset and the flexible obstacle avoidance dataset into the spatial comparison process, mark the coordinates of the overlapping area of ​​the rigid obstacle avoidance space and the flexible obstacle avoidance space within the working area, retain the original range of the rigid obstacle avoidance space, and adjust the boundary coordinates of the flexible obstacle avoidance space according to the overlapping area to form an overlapping adjustment dataset.

[0101] Import the boundary coordinates of the obstacle avoidance datasets from both the rigid and flexible obstacle avoidance datasets into spatial analysis software. Use polygon overlay operations to calculate the overlapping region. For the overlapping region, retain the boundary of the rigid obstacle avoidance space unchanged, and adjust the boundary of the flexible obstacle avoidance space: within the overlapping region, shrink the boundary of the flexible obstacle avoidance space away from the rigid obstacle avoidance space by half the overlap depth, ensuring that the adjusted flexible obstacle avoidance space does not overlap with the rigid obstacle avoidance space. Record the adjusted flexible obstacle avoidance space boundary coordinates, overlapping region coordinates, and shrinkage distance to form the overlap adjustment dataset.

[0102] Step S1346: Extract all information on ground bearing capacity from the ground bearing capacity dataset, match the ground bearing capacity information with the rigid obstacle avoidance dataset, expand the range of rigid obstacle avoidance space in soft soil sections and pothole-prone sections, and maintain the original range of rigid obstacle avoidance space in sections with good ground bearing capacity, thus forming a ground-adaptive avoidance dataset.

[0103] Ground bearing capacity level information is extracted from the ground bearing capacity dataset obtained in step S133 and matched with the avoidance space coordinates of the rigid obstacle avoidance dataset. When the ground bearing capacity level of the area where the rigid obstacle avoidance space is located is "weak" or "very weak" (BCI < 0.4), the fixed avoidance distance D_rigid is increased by ΔD (e.g., ΔD = 0.5 meters), and the avoidance space boundary is recalculated; when the bearing capacity level is "medium" or above, the original D_rigid is maintained. The adjusted rigid obstacle avoidance space boundary coordinates, bearing capacity level, and ΔD value are recorded to form a ground-adaptive avoidance dataset.

[0104] Step S1347: Extract all information on material hardness from the obstacle classification dataset, match the material hardness information with the ground adaptation avoidance dataset, verify the fit between avoidance distance and material hardness, adjust the boundary coordinates of the avoidance space according to the material hardness so that the avoidance distance fits the material hardness, and form a hardness adaptation dataset.

[0105] Extract the HI values ​​of rigid obstacles from the obstacle classification dataset obtained in step S131 and match them with the avoidance distance D_rigid in the ground-adaptive avoidance dataset. For ultra-hard materials (HI≥0.8), verify whether D_rigid is ≥1.8 meters (higher safety requirement); if insufficient, increase it to 1.8 meters. For hard materials (0.6≤HI<0.8), maintain D_rigid=1.5 meters. Adjust the corresponding avoidance space boundary coordinates and calculate the adjusted fit (the ratio of actual D_rigid to the theoretical requirement), ensuring a fit ≥90%. Record the adjusted boundary coordinates, HI values, and fit to form a hardness-adaptive dataset.

[0106] Step S1348: Extract the boundary coordinates of the core harvesting area from the core area dataset, match the boundary coordinates of the core harvesting area with the hardness adaptation dataset, mark the intersection coordinates of the obstacle avoidance space boundary and the core harvesting area boundary, record the number and distribution of the intersection points, integrate the intersection point information according to the spatial coordinates, and form an intersection point marking dataset.

[0107] Extract the boundary coordinate sequence of the core harvesting area from the core area dataset obtained in step S132, and perform polygon intersection calculations with the obstacle avoidance spatial boundary coordinates in the hardness adaptation dataset to find the coordinates (X_i, Y_i) of all intersection points. Record the obstacle ID, core area ID, and intersection type (entry / exit) of each intersection point, and count the number of intersection points and their distribution on the boundary (e.g., left boundary, right boundary, top boundary, etc.). Sort and integrate these intersection point information according to spatial coordinates to form an intersection point labeling dataset.

[0108] Step S1349: Import the rigid avoidance dataset, flexible avoidance dataset, overlap adjustment dataset, ground adaptation avoidance dataset, hardness adaptation dataset, and intersection mark dataset into the data integration process. Integrate all relevant information of obstacle avoidance space according to obstacle type, unify the data format and spatial coordinate correspondence, and form an integrated avoidance space dataset.

[0109] The rigid obstacle avoidance dataset from step S1342, the flexible obstacle avoidance dataset from step S1344, the overlap adjustment dataset from step S1345, the ground-adaptive obstacle avoidance dataset from step S1346, the stiffness-adaptive dataset from step S1347, and the intersection marker dataset from step S1348 are integrated. The data is organized by obstacle type (rigid / flexible), and the obstacle avoidance space information for each obstacle includes the original boundary, the adjusted boundary (considering overlap, ground, and stiffness factors), intersection coordinates, area, and safety distance parameters. The unified data format is JSON, and the spatial coordinates adopt the same coordinate system as the 3D perception dataset, forming an integrated obstacle avoidance space dataset.

[0110] Step S13410: Standardize the boundary coordinates, area, and adaptation conditions of all obstacle avoidance spaces in the obstacle avoidance dataset to form an obstacle avoidance dataset.

[0111] The obstacle avoidance space dataset is standardized: boundary coordinates are retained to two decimal places, area calculations are accurate to 0.1 square meters, and adaptation conditions (such as ground load-bearing capacity and material hardness) are clearly marked. Redundant intermediate adjustment data is removed, retaining only the final obstacle avoidance space parameters. Topology checks ensure that boundary coordinates are closed and do not self-intersect, ultimately forming an obstacle avoidance dataset containing the final obstacle avoidance space information for all obstacles.

[0112] Step S135: Extract the spatial coordinates of the core harvesting area from the core area dataset, mark the actual location of all obstacles within the spatial range of the core harvesting area, match the ground bearing capacity dataset with the location of the obstacles, label the bearing capacity information of the surrounding ground for each obstacle, and integrate the obstacle information in the core harvesting area according to the spatial coordinates to form a regional obstacle dataset.

[0113] For example, step S1351: Extract the boundary coordinates and overall spatial range of the core harvesting area from the core area dataset, further divide the core harvesting area into sub-regions according to the coordinate characteristics of plant distribution, mark the coordinates of each sub-region and record the information of plant distribution and plant height to form a core area subdivision dataset.

[0114] Extract the boundary coordinates (Xmin, Ymin, Xmax, Ymax) and overall spatial extent of the core harvesting area from the core area dataset obtained in step S132. Divide the core area into sub-regions using a 5m × 5m grid, and label the center coordinates (Xc, Yc) and boundary coordinates of each sub-region. Record the plant distribution density ρ, average plant height H_avg, and standard deviation of plant height σ_h for each sub-region to form a detailed dataset of the core area.

[0115] Step S1352: Extract the location coordinates and feature information of all obstacles from the obstacle classification dataset, filter out the obstacles within the core harvesting area, record the outline size, material hardness, motion state and location coordinates of each obstacle, organize the information according to the obstacle type, and output the obstacle dataset within the core area.

[0116] Extract the location coordinates (Xc, Yc), contour dimensions (W, L, H), material hardness (HI), and motion state (static / dynamic) of all obstacles from the obstacle classification dataset obtained in step S131. Filter out obstacles within the core harvesting area by spatial inclusion judgment (whether the obstacle's center coordinates are within the boundary of the core harvesting area), record the above information for each obstacle, and classify them according to rigid / flexible and static / dynamic categories to form an obstacle dataset within the core area.

[0117] Step S1353: Extract ground bearing capacity information from the ground bearing capacity dataset in the core harvesting area, match the ground bearing capacity information with the obstacle dataset in the core area, label the soil compaction and ground undulation of the surrounding ground for each obstacle, organize the results according to the obstacle coordinates, and obtain the ground dataset around the obstacle.

[0118] Extract the BCI value, soil compaction C, and ground undulation amplitude H from the ground bearing capacity dataset obtained in step S133 within the core harvesting area. Draw a circle with the center coordinates (Xc, Yc) of each obstacle and a radius R_ground (e.g., 2 meters), and extract the ground bearing capacity information within the circle: average BCI value, minimum C value, and maximum H value. Label this information for each obstacle, organize it according to the obstacle coordinates, and form a ground bearing capacity dataset around the obstacles.

[0119] Step S1354: Match the obstacle dataset within the core area with the ground dataset around the obstacle, filter out fixed obstacles whose outline size is within the equipment detour threshold and whose surrounding ground bearing capacity is suitable for equipment detour, record the minimum space range required for detour of each obstacle, organize the information by coordinates, and organize it into a detourable obstacle dataset.

[0120] Set equipment detour thresholds: width W ≤ W_turn (minimum turning width of the equipment), length L ≤ L_turn (minimum turning length of the equipment). For static rigid barriers within the core area, if W ≤ W_turn and L ≤ L_turn, and the surrounding ground BCI ≥ 0.4 (medium load-bearing capacity), then it is determined to be a detourable barrier. Calculate the minimum space required for detour: Based on the barrier avoidance space, expand the equipment turning radius R_turn outward to form a detour channel. Record the boundary coordinates and width of the channel, organize them according to the barrier coordinates, and form a detourable barrier dataset.

[0121] Step S1355: Match the obstacle dataset within the core area with the ground dataset around the obstacle, filter out fixed obstacles whose outline size exceeds the equipment detour threshold and whose surrounding ground bearing capacity is suitable for the equipment to cross, record the height and width parameters required to cross each obstacle, organize the information by coordinates, and output the dataset of obstacles to be crossed.

[0122] For static rigid barriers within the core area, if W > W_turn or L > L_turn, but height H ≤ H_cross (maximum crossing height of the equipment) and the surrounding ground BCI ≥ 0.6 (strong bearing capacity), then it is determined to be a barrier that needs to be crossed. Record the actual height H, top width W_top, and bottom width W_bottom of the barrier, calculate the adjustment parameters of the equipment suspension system required for crossing, organize them according to the barrier coordinates, and form a dataset of barriers that need to be crossed.

[0123] Step S1356: Extract all information of moving obstacles from the obstacle dataset within the core area, record the motion trajectory segments and speed characteristics of all moving obstacles one by one, reserve buffer space for the equipment to bypass the corresponding obstacles, mark the coordinate range of the buffer space, and form a dataset of moving and bypassable obstacles.

[0124] Extract the ID, motion trajectory segment (coordinate sequence), velocity V, and orientation angle θ of dynamic obstacles (moving obstacles) from the obstacle dataset within the core area. Reserve a buffer space for each moving obstacle: a rectangular buffer space centered on the current position, with a length of V×t_safe (t_safe is the safe reaction time) along the direction of motion and a vertical width of D_rigid+0.5 meters. Mark the boundary coordinates of the buffer space, organize them by obstacle ID, and form a dataset of movable, dodgeable obstacles.

[0125] Step S1357: Extract information on areas with poor ground bearing capacity from the ground data set around the obstacle, match the corresponding information with the obstacle dataset to be crossed, adjust the obstacles to be crossed corresponding to the areas around the obstacles with poor ground bearing capacity in the core harvesting area to bypass obstacles, re-mark the position coordinates of the corresponding obstacles, and obtain the obstacle classification adjustment dataset.

[0126] Regions with a BCI < 0.6 (indicating poor load-bearing capacity) are selected from the ground data set surrounding obstacles. These regions are then matched with the obstacle locations in the obstacle dataset that needs to be crossed. If the ground BCI surrounding an obstacle that needs to be crossed is < 0.6, it is reclassified from an obstacle that needs to be crossed to a detourable obstacle (even if its outline size exceeds the detour threshold, a detour path must still be found), and its location coordinates and detour priority are relabeled (higher than the original detourable obstacle), forming an obstacle classification adjustment dataset.

[0127] Step S1358: Match the bypass obstacle dataset with the core area subdivision dataset, plan the spatial range of bypass paths for each bypass obstacle so that the bypass paths are within the spatial range of the core harvesting area, mark the boundary coordinates of the bypass paths, and organize them into a bypass range dataset.

[0128] The obstacle locations in the bypass obstacle dataset are matched with the sub-regions in the core area subdivision dataset. Detour paths are planned within each sub-region: the path starts at the sub-region's entrance point and ends at the exit point. The path must avoid obstacle clearance space and be as close as possible to the center of the core area. The A* algorithm is used to calculate the optimal detour path, and the path's coordinate sequence and boundary coordinates (the path's centerline extends to both sides by half the width of the equipment) are recorded, forming a detour range dataset.

[0129] Step S1359: Extract information on all obstacles to be crossed from the obstacle-crossing dataset, mark the spatial coordinates of the crossing point for each obstacle, set the lifting value of the component during crossing based on the parameters of the harvesting component's operating position, record the specific parameters of the lifting value of the component, and output the crossing parameter dataset.

[0130] For obstacles that need to be crossed, the center coordinates of the flat area at the top of the obstacle are selected as the crossing point (X_cross, Y_cross). The lifting value of the harvesting component is set according to the obstacle height H: ΔL = L0 + H + ΔH (L0 is the normal operating height, ΔH is the safety margin). The coordinates of the crossing point, the value of ΔL, and the crossing speed limit V_cross (below the normal driving speed) are recorded to form a crossing parameter dataset.

[0131] Step S13510: Import the bypassable obstacle dataset, the obstacle to be crossed dataset, the movable bypassable obstacle dataset, the obstacle classification adjustment dataset, the bypass range dataset, and the crossing parameter dataset into the data integration process. Organize the processing information of all obstacles according to the spatial coordinates of the core harvesting area to obtain the regional obstacle dataset.

[0132] The six datasets were integrated according to the coordinates of 5m x 5m sub-regions within the core harvesting area. Each sub-region included the type of all obstacles within that area (obstacles that can be bypassed / crossed / moved), the handling method (obstacle path / crossing parameters / buffer space), and related parameters (path coordinates / lift value / speed limit). After unifying the data format, a regional obstacle dataset was formed.

[0133] Step S136: Import the obstacle classification dataset, core area dataset, ground bearing capacity dataset, obstacle avoidance dataset, and regional obstacle dataset into the model building process. Map each type of dataset to the corresponding spatial location of the target coupled model according to the actual spatial coordinates of the pasture harvesting area, and integrate the various datasets to obtain the basic data of the model.

[0134] The obstacle classification dataset from step S131, the core area dataset from step S132, the ground bearing capacity dataset from step S133, the obstacle avoidance dataset from step S134, and the regional obstacle dataset from step S135 are imported into the target coupling model building platform. This platform uses a three-dimensional mesh model with a mesh cell size of 0.1m × 0.1m × 0.1m. Each mesh cell is labeled with attributes based on the aforementioned datasets: obstacle type (rigid / flexible / hidden), pasture growth status (core area / non-core area / no pasture), ground bearing capacity level, avoidance space markers, obstacle handling method, etc. The information from each dataset is mapped to the corresponding mesh cell according to the actual spatial coordinates (X, Y, Z) of the pasture harvesting area, forming the basic data for the model.

[0135] Step S137: Import the basic model data into the target coupling model for computation, calculate the impact range of obstacles on forage harvesting operations within the core harvesting area for each obstacle, mark the spatial coordinates of the overlapping parts of the obstacle avoidance space and the core harvesting area, integrate the relevant information of the impact range according to the obstacle type, and form the impact range dataset.

[0136] Import the model base data into the target coupling model, which uses the spatial overlay analysis algorithm to calculate the obstacle influence range: for each obstacle, its influence range is the intersection of the obstacle avoidance space and the core harvesting area. Calculate the spatial coordinates of the overlapping part through polygon intersection, calculate the influence area (intersection area) and the influence ratio (influence area / core area). Count the number, total area, and average influence ratio of the influence ranges by obstacle type (rigid / flexible / moving / hidden) to form an influence range data set, which quantifies the interference degree of the obstacles on the harvesting operation.

[0137] Step S138: Based on the influence range data set, adjust the boundary coordinates of the core harvesting area and the boundary coordinates of the obstacle avoidance space, maintain the safety range of the obstacle avoidance space, reduce the occupancy range of the obstacle avoidance space on the core harvesting area, redraw the adjusted area boundary and mark the coordinates, and integrate them into a boundary adjustment data set.

[0138] When the overlapping area between the obstacle avoidance space and the core harvesting area is large (such as the influence ratio > 10%), adjust the core area boundary: near the overlapping area, contract the core area boundary away from the obstacle by half of the overlapping depth to ensure that the core area still contains most of the forage growth area. At the same time, check whether the area of the adjusted core area meets the harvesting requirements (not less than 90% of the original area), and if not, appropriately expand the core area boundary of the non-overlapping area. Redraw the boundary coordinates of the core area and the obstacle avoidance space, record the coordinate changes and area changes before and after the adjustment, and form a boundary adjustment data set.

[0139] Step S139: Set dynamic adjustment nodes in the target coupling model for the local differences between the motion state and the ground state of the moving obstacle, mark the spatial coordinates of the dynamic adjustment nodes one by one, record the trigger conditions and data update ranges of each node, and integrate all the information of the dynamic adjustment nodes according to the spatial position to form a model node data set.

[0140] Predict the sub-areas of the core area that the moving obstacle may enter in the future according to the motion trajectory segment of the moving obstacle (step S125), and set dynamic adjustment nodes in these sub-areas. The spatial coordinate of each node is the center of the sub-area (Xn, Yn), the trigger condition is that the distance d between the moving obstacle and the node < d_trigger (trigger distance), and the data update range is the area around the node R_node (such as 5 meters). When the trigger condition is met, the model will recalculate the obstacle avoidance space and path planning within the update range. Record the coordinates, trigger conditions, update ranges, and associated moving obstacle IDs one by one to form a model node data set.

[0141] Step S1310: Supplement the influence range dataset, boundary adjustment dataset, and model node dataset to the corresponding spatial location of the target coupled model, and integrate the information of the basic model data and various adjustment datasets to form coupled model data.

[0142] The influence range dataset from step S137, the boundary adjustment dataset from step S138, and the model node dataset from step S139 are added to the 3D mesh of the target coupled model: the influence range is labeled in the "Influence Level" attribute of the corresponding mesh, the boundary adjustment is updated in the "Core Region Marker" attribute of the mesh, and the model node information is stored in the "Dynamic Node" attribute of the mesh. After integrating all the data, each mesh cell of the model contains complete environmental and obstacle information, forming coupled model data.

[0143] Step S140: Combine the coupled model data and the collaborative adaptation parameters to plan the dynamic obstacle avoidance trajectory, adjust the overall direction of the dynamic obstacle avoidance trajectory and the adaptation value of the harvesting component's working position, so that the direction of the dynamic obstacle avoidance trajectory fits the obstacle avoidance space and the lateral coverage size of the equipment's single harvest, and integrate to obtain the optimized obstacle avoidance trajectory.

[0144] Step S141: Import the coupled model data and collaborative adaptation parameters into the trajectory planning basic processing flow. Extract information on obstacle avoidance space and core harvesting area boundary from the coupled model data. Extract information on the lateral coverage size, travel speed, and harvesting component operation position of the equipment in a single harvest from the collaborative adaptation parameters. Integrate these into trajectory planning basic data according to the operation requirements to form a trajectory planning basic dataset.

[0145] The coupled model data obtained in step S1310 and the collaborative adaptation parameters obtained in step S110 are imported into the trajectory planning system. From the coupled model data, the boundary coordinates of all obstacle avoidance spaces, the adjusted boundary coordinates of the core harvesting area, and the ground bearing capacity distribution are extracted. From the collaborative adaptation parameters, the lateral coverage dimension W×k1, the driving speed range A×k2 to B×k2, and the baseline operating position of the harvesting component L×k3 are extracted. According to operational requirements (such as prioritizing coverage of the core area and avoiding high-risk ground), the above information is integrated into trajectory planning basic data, including: core area polygons, obstacle avoidance space list, equipment parameters (W, k1, V_min, V_max, L0), and a list of ground risk areas, forming the trajectory planning basic dataset.

[0146] Step S142: Based on the trajectory planning basic dataset, draw the initial obstacle avoidance trajectory within the work area. Plan the overall direction of the initial obstacle avoidance trajectory according to the coordinates of the obstacle avoidance space so that the entire path of the initial obstacle avoidance trajectory avoids all obstacle avoidance spaces, while covering the entire spatial range of the core harvesting area. Integrate the information of the initial obstacle avoidance trajectory according to the coordinate sequence to form the initial trajectory dataset.

[0147] Based on the trajectory planning dataset, an improved RRT* algorithm is used to draw the initial obstacle avoidance trajectory for the work area. The sampling space of the algorithm is the core harvesting area, the target point is the area exit, and the obstacles are all obstacle avoidance spaces. During the path search process, it is ensured that the distance between the trajectory and the obstacle avoidance space is greater than or equal to the safe distance (D_rigid or D_flex), and the trajectory passes through all core sub-regions (coverage ≥ 95%). The resulting initial obstacle avoidance trajectory is a polyline composed of a series of continuous coordinate points (X0, Y0) → (X1, Y1) → ... → (Xn, Yn). The position, cumulative distance, and orientation angle of each point are recorded according to the coordinate sequence to form the initial trajectory dataset.

[0148] Step S143: Extract the spatial coordinates of the turning points of the initial obstacle avoidance trajectory from the initial trajectory dataset, perform smoothing processing on all the turning points of the initial obstacle avoidance trajectory, adjust the coordinate position of the turning points to reduce the sudden angle of the device's driving direction, so that the adjusted trajectory conforms to the motion parameter constraints of the autonomous driving device, and integrate them into a smooth trajectory dataset.

[0149] Extract the coordinates (Xt, Yt) of the turning points where the change in direction angle Δθ > θ_max (maximum steering angle of the equipment) from the initial trajectory dataset. For each turning point, use a Bézier curve for smoothing: take two adjacent points before and after the turning point as control points to generate a smooth curve to replace the original polyline. Adjust the curve parameters so that the rate of change in direction angle ≤ θ_max / Δs (Δs is the distance between adjacent points), ensuring that the equipment's steering mechanism can follow smoothly. Record the smoothed trajectory coordinate sequence, curve parameters for each segment, and the rate of change in direction angle to form a smoothed trajectory dataset.

[0150] Step S144: Extract the dynamically adjusted nodes from the model node dataset from the coupled model data, set the data update nodes at the corresponding spatial locations in the smooth trajectory dataset, reserve a time window for trajectory correction for each data update node according to the motion state of the moving obstacle, mark the coordinates and triggering conditions of the data update nodes, and form a node-adapted trajectory dataset.

[0151] Extract the dynamically adjusted node coordinates (Xn, Yn) from the coupled model data dataset. Find the trajectory point (Xp, Yp) closest to (Xn, Yn) in the smooth trajectory dataset and set this point as the data update node. Calculate the time window Δt = d_trigger / V based on the speed V of the moving obstacle and the trigger distance d_trigger. That is, the model starts the data update process Δt time before the device arrives at the data update node. Label the coordinates, associated moving obstacle ID, trigger time window, and data update range of each data update node to form a node-adapted trajectory dataset.

[0152] Step S145: Extract the driving speed-related data from the collaborative adaptation parameters, adjust the driving speed adaptation values for the sections near the obstacle avoidance space according to the section division of the node adaptation trajectory dataset, set the constant driving speed adaptation values for the sections within the core harvesting area, and mark the speed adaptation values section by section to form the speed adaptation trajectory dataset.

[0153] Extract the driving speed correction coefficient k2 from the collaborative adaptation parameters to obtain the actual speed range V_min = A × k2 and V_max = B × k2. Divide the trajectories of the node adaptation trajectory dataset into sections according to the data update nodes. For the sections near the obstacle avoidance space (distance < D_slow, such as 3 meters), set the speed as V_slow = V_min + 0.3 × (V_max - V_min); for the obstacle-free sections within the core harvesting area, set the speed as V_cruise = V_max; for the sections near the data update nodes, set the speed as V_update = V_min + 0.5 × (V_max - V_min). Mark the speed adaptation values, section start / end coordinates, and lengths section by section to form the speed adaptation trajectory dataset.

[0154] Step S146: Extract the position coordinates of the flexible obstacle from the coupling model data, finely adjust the trajectory directions around the flexible obstacle in the speed adaptation trajectory dataset, reduce the occupied range of the flexible obstacle avoidance space, and retain the operation space of the core harvesting area, and integrate it into the flexible adaptation trajectory dataset.

[0155] Extract the boundary coordinates of the avoidance space of the flexible obstacle from the coupling model data and calculate its shortest distance d from the speed adaptation trajectory. If d > D_flex (flexible obstacle avoidance distance), no adjustment is required; if d ≤ D_flex, finely adjust the trajectory direction: offset by Δd = D_flex - d + 0.1 meters in the direction away from the flexible obstacle, and ensure that the trajectory remains within the core harvesting area and does not approach other obstacles during the offset process. The adjusted trajectory reduces the occupancy of the flexible obstacle avoidance space in the core area, records the adjusted trajectory coordinate sequence and offset amount, and forms the flexible adaptation trajectory dataset.

[0156] Step S147: Extract the ground bearing dataset from the coupling model data, match the coordinates of the soil compactness and pothole distribution with the coordinate sequence of the flexible adaptation trajectory, adjust the operation position adaptation values of the harvesting components for the soft soil sections and pothole distribution sections, and increase the operation position adaptation values of the harvesting components for the corresponding sections to form the height adaptation trajectory dataset.

[0157] Step S1471: Extract the ground bearing data set from the coupled model data. Divide the forage harvesting operation area into soft soil sections, pitted sections, and flat ground sections according to the distribution of soil compactness and the coordinates of the pitted distribution. Mark the boundary coordinates section by section, record the soil compactness values and pitted depth parameters, and form the ground partition data set.

[0158] Extract the BCI value, soil compactness C, and pitted depth D_pit of the ground bearing data set from the coupled model data. Divide the area passed by the trajectory into sections: when C < C_sink, it is a soft soil section; when D_pit > D_pit_th, it is a pitted section; when BCI ≥ 0.6 and D_pit ≤ D_pit_th, it is a flat ground section. Mark the starting / ending coordinates, length, average C value, and maximum D_pit value section by section, and form the ground partition data set.

[0159] Step S1472: Extract the complete coordinate sequence of the trajectory from the flexible adaptation trajectory data set. Accurately match the trajectory coordinate sequence with the spatial coordinates of the ground partition data set, mark the coordinates of each section of the trajectory passing through the soft soil section, pitted section, and flat ground section, and integrate the matching results section by section to form the trajectory-ground association data set.

[0160] Match the coordinate sequence (X0, Y0)...(Xn, Yn) of the flexible adaptation trajectory with the section boundary coordinates of the ground partition data set to determine the section type (soft / pitted / flat) to which each trajectory point belongs. Integrate the trajectory coordinates according to the section type. For example, the soft soil section contains coordinate points (Xa, Ya) to (Xb, Yb), record the section type, starting / ending index, and coordinate sequence, and form the trajectory-ground association data set.

[0161] Step S1473: Extract the part of the trajectory passing through the soft soil section from the trajectory-ground association data set. Match the soil compactness values with the coordinates of the corresponding part of the section, adjust the operation position adaptation value of the harvesting component according to the soil compactness, increase the operation position adaptation value of the harvesting component for the corresponding part of the section according to the softness of the soil, and mark the adaptation value coordinate by coordinate to form the soft area height data set.

[0162] For the soft soil section in the trajectory-ground association data set, extract the average soil compactness C_avg of this section, calculate the harvesting component lifting value ΔL_soft = k_soft × (C_sink - C_avg), where k_soft is the lifting coefficient for the soft soil section. The operation position adaptation value L_soft = L0 × k3 + ΔL_soft, and mark the L_soft value coordinate by coordinate (linear transition within the same section), and form the soft area height data set.

[0163] Step S1474: Extract the portion of the track that passes through the pothole distribution section from the trajectory ground association dataset, match the pothole depth and width parameters with the coordinates of the corresponding section, adjust the harvester operation position adaptation value according to the pothole parameters, increase the harvester operation position adaptation value of the corresponding section according to the pothole depth, reserve buffer space for the pothole edge, and label the adaptation value for each coordinate to form a pothole area height dataset.

[0164] For road sections with pothole distribution, the maximum pothole depth D_pit_max is extracted, and the elevation value ΔL_pit = k_pit × D_pit_max is calculated, where k_pit is the pothole elevation coefficient. Within a buffer space of Δs_buffer (e.g., 1 meter) before and after the pothole edge, the adaptation value linearly transitions from L0 × k3 to L_pit = L0 × k3 + ΔL_pit, while maintaining L_pit in the intermediate region. The adaptation value is labeled coordinate by coordinate to form a pothole area height dataset.

[0165] Step S1475: Extract the portion of the trajectory that passes through the flat road section from the trajectory ground association dataset, extract the original adaptation value of the harvesting component operation position from the cooperative adaptation parameters, use the original adaptation value in the corresponding road section, and label the adaptation value of the harvesting component operation position coordinate by coordinate to form a flat area height dataset.

[0166] For flat road sections, the baseline operating position L0×k3 of the harvesting component in the collaborative adaptation parameters is directly used as the adaptation value, and this value is marked coordinate by coordinate to form a flat area height dataset.

[0167] Step S1476: Match the soft area height dataset, the pothole area height dataset, the flat area height dataset with the coordinate sequence of the flexible adaptation trajectory point by point, label the corresponding harvesting component operation position adaptation value for each trajectory coordinate point, and integrate the information of all adaptation values ​​according to the coordinate sequence to form the height adaptation basic dataset.

[0168] The adaptation values ​​of the three height datasets are matched according to the trajectory coordinate point index, and each coordinate point (Xi, Yi) corresponds to a L_i (job location adaptation value). All adaptation values ​​are integrated according to the coordinate sequence (X0, Y0, L0)...(Xn, Yn, Ln) to form the height adaptation base dataset.

[0169] Step S1477: Extract the coordinates of the switching nodes for the adaptation values ​​of the harvesting component operation position from the highly adapted base dataset, adjust the spacing between the switching nodes, label the gradient of the adaptation value changes, and form a highly smooth dataset.

[0170] Identify the switching nodes where the L_i in the basic dataset changes highly adaptively (such as the point switching from L_soft to L_pit), and calculate the spacing Δs between adjacent switching nodes. If Δs < Δs_min (the minimum switching spacing), merge adjacent nodes and adjust the gradient (slope) of the adaptation value change to make the change smoother. Mark the coordinates of each switching node, the adaptation values before and after, and the change gradient to form a highly smoothed dataset.

[0171] Step S1478: Extract the obstacle position information from the coupled model data, match the obstacle position information with the highly smoothed dataset, and finely adjust the adaptation value of the harvesting component operation position of the trajectory coordinate points around the obstacle in the highly smoothed dataset, and increase the adaptation value of the corresponding coordinate points to avoid obstacles, forming an obstacle adaptation height dataset. <6000356>

[0172] Extract the position coordinates and height H of all obstacles from the coupled model data, and calculate the vertical distance d_z between the trajectory coordinate points and the obstacles. When d_z < H + ΔH (ΔH is the safety margin), finely adjust the adaptation value L_i = L_i + (H + ΔH - d_z) of this coordinate point to ensure that the harvesting component does not collide with the obstacle. Record the adjusted L_i value and the corresponding obstacle ID to form an obstacle adaptation height dataset.

[0173] Step S1479: Extract the spatial coordinates of all harvesting component operation position adjustment nodes from the obstacle adaptation height dataset, record the adaptation value and switching speed of the harvesting component operation position of each adjustment node, and standardize the format of all parameters according to the component movement parameter limit specifications of the unmanned device to form a height node marking dataset.

[0174] Extract the adjustment node coordinates of L_i in the obstacle adaptation height dataset, record the L value of each node and the required switching time Δt (calculated according to the maximum lifting speed V_lift of the component, Δt = |L_new - L_old| / V_lift). Ensure that Δt ≥ Δt_min (the minimum response time of the component), and if not satisfied, reduce the switching speed. Integrate this information in the order of node coordinates to form a height node marking dataset.

[0175] Step S14710: Integrate the height node marking dataset with the flexible adaptation trajectory dataset, supplement the adaptation value of the harvesting component operation position and the information of the height adjustment node to the corresponding positions of the flexible adaptation trajectory, and replace the harvesting component operation position parameters in the original trajectory to form a height adaptation trajectory dataset.

[0176] Add the L value and adjustment node information of the height node marking dataset to the corresponding coordinate points of the flexible adaptation trajectory dataset, and replace the original default operation position parameters. Each trajectory point now contains complete parameters such as (X, Y, L, speed) to form a height adaptation trajectory dataset.

[0177] Step S148: Extract all coordinate sequences of the trajectory from the highly adapted trajectory dataset, divide the trajectory into core harvesting segments and obstacle avoidance segments, mark the spatial coordinates of the connection points between the core harvesting segments and obstacle avoidance segments segment by segment, set the switching value of driving speed for each connection point, integrate all information of segment division and connection points to form a segment-marked trajectory dataset.

[0178] Based on obstacle information and core area boundaries in the highly adapted trajectory dataset, the trajectory is divided into core harvesting segments (located within the core area and without obstacle avoidance) and obstacle avoidance segments (segments that bypass or cross obstacles). The coordinates (Xj, Yj) of the connection point between the two types of segments are identified, and speed switching values ​​are set at the connection point: when switching from the core segment speed V_cruise to the obstacle avoidance segment speed V_slow, the deceleration time Δt_decel = (V_cruise - V_slow) / a_max (where a_max is the maximum deceleration); conversely, the acceleration time Δt_accel = (V_cruise - V_slow) / a_max. The connection point coordinates, the speeds before and after the switch, and the switch time are marked to form a segment-marked trajectory dataset.

[0179] Step S149: Extract the motion state information of the moving obstacle from the coupled model data, adjust the direction of the obstacle avoidance road segment in the road segment marker trajectory dataset according to the motion trend of the moving obstacle, reserve buffer space for the motion of the moving obstacle, adjust the coordinate sequence of the obstacle avoidance road segment to fit the motion characteristics of the moving obstacle, and form a dynamically corrected trajectory dataset.

[0180] For example, step S1491: Extract the motion state information of moving obstacles from the coupled model data, record the motion trajectory segments, real-time speed and motion direction of each moving obstacle, infer the subsequent motion path and location range according to the motion characteristics of the moving obstacles, mark the coordinates of the inferred path, and obtain the moving obstacle trend dataset.

[0181] The ID, trajectory segment (coordinate sequence), velocity V, and orientation angle θ of the moving obstacle are extracted from the coupled model data. A Kalman filter algorithm is used to predict the movement path within the next t_pred (e.g., 5 seconds): X_pred(t) = X_current + V × cosθ × t, Y_pred(t) = Y_current + V × sinθ × t, where t ∈ [0, t_pred]. Considering the possible range of velocity and orientation changes (±ΔV, ±Δθ), a predicted location range (95% confidence interval) is generated. The center and boundary coordinates of the predicted path are marked to form a moving obstacle trend dataset.

[0182] Step S1492: Extract all relevant information of the obstacle avoidance section from the road segment marked trajectory dataset, mark the positions of moving obstacles, avoidance spaces, and path directions corresponding to the obstacle avoidance section, classify and organize all data of the obstacle avoidance section according to the numbers of moving obstacles, and obtain the associated dataset of the obstacle avoidance section.

[0183] Extract the coordinate sequence, associated moving obstacle ID, and current avoidance space coordinates of the obstacle avoidance section from the road segment marked trajectory dataset. Classify and organize according to the moving obstacle ID, and each ID corresponds to a subset of the obstacle avoidance section, including road segment coordinates, length, current speed, and distance from the obstacle, to form the associated dataset of the obstacle avoidance section.

[0184] Step S1493: Match the moving obstacle trend dataset with the associated dataset of the obstacle avoidance section, adjust the overall direction of the obstacle avoidance section according to the speculated movement path of the moving obstacle, make the adjusted obstacle avoidance section maintain an appropriate spatial distance from the speculated movement path, and re-plan the coordinate sequence of the obstacle avoidance section to obtain the dataset of adjusted path direction.

[0185] For each moving obstacle, calculate the spatial distance between its speculated movement path and the corresponding obstacle avoidance section. If the minimum distance d < D_safe (dynamic safety distance, D_safe = V × t_safe + D_rigid), then adjust the direction of the obstacle avoidance section: offset by Δd = D_safe - d along the direction perpendicular to the obstacle movement direction, and re-plan the road segment coordinate sequence (using local re-planning with the A* algorithm). Record the adjusted coordinate sequence, offset direction, and distance to form the dataset of adjusted path direction.

[0186] Step S1494: Import the dataset of adjusted path direction and the moving obstacle trend dataset into the intersection prediction process, calculate the intersection possibility between the speculated movement path of the moving obstacle and the adjusted obstacle avoidance section, mark the coordinate and time node of the possible intersection area, and organize the prediction results according to spatial coordinates to obtain the intersection prediction dataset.

[0187] Calculate the intersection coordinates (X_inter, Y_inter) between the adjusted obstacle avoidance section and the speculated path of the moving obstacle. If the intersection exists, then calculate the time t_veh for the device to reach the intersection t_veh = s_veh / V_veh (s_veh is the distance from the device to the intersection) and the time t_obs for the obstacle to reach the intersection t_obs = s_obs / V_obs (s_obs is the distance from the obstacle to the intersection). When |t_veh - t_obs| < t_overlap (time overlap threshold), it is determined as a possible intersection, and mark the intersection area coordinates, time node, and risk level (high / medium / low) to form the intersection prediction dataset.

[0188] Step S1495: Based on the intersection prediction dataset, further adjust the direction of the obstacle avoidance section for areas where intersection is possible, expand the avoidance space range of the corresponding area or change the path direction of the obstacle avoidance section, so that the adjusted obstacle avoidance section completely avoids the predicted movement range of the moving obstacle, and obtains the intersection avoidance dataset.

[0189] For high-risk intersection areas in the intersection prediction dataset, one of the following strategies is used to adjust the obstacle avoidance sections: 1) Increase the avoidance space range, increasing the offset distance to Δd' = Δd + 0.5 meters; 2) Change the path direction, with the detour direction opposite to the obstacle's movement direction. After adjustment, the intersection probability is recalculated to ensure no high-risk intersections occur. The final section coordinate sequence and adjustment strategy are recorded to form the intersection avoidance dataset.

[0190] Step S1496: Extract relevant data on driving speed from the cooperative adaptation parameters, combine them with the features of the intersection avoidance dataset, adjust the adaptation value of driving speed in obstacle avoidance sections close to the predicted range of moving obstacles, reduce the driving speed adaptation value of the corresponding road sections, label the speed adaptation value for each road section, and obtain the speed optimization adjustment dataset.

[0191] The minimum driving speed V_min is extracted from the cooperative adaptation parameters. For obstacle avoidance sections in the intersection avoidance dataset, the speed adaptation value is reduced to V_adj=max(V_min, V_current×0.7) to allow reaction time for changes in obstacle movement. The adjusted speed value, adjustment ratio, and effective range are labeled for each road segment to form a speed optimization adjustment dataset.

[0192] Step S1497: Import the path alignment adjustment dataset, intersection prediction dataset, intersection avoidance dataset, and speed optimization adjustment dataset into the data integration process, and connect the adjusted parameters with the core harvested road segments of the road segment marking trajectory dataset to obtain the preliminary corrected trajectory dataset.

[0193] The adjustment parameters (path coordinates, speed, and direction) of the above four datasets are integrated to ensure a smooth transition at the connection points between obstacle avoidance sections and core harvesting sections (direction angle change ≤ θ_max). The data is then reorganized according to the trajectory coordinate sequence to form a preliminary corrected trajectory dataset containing the adjusted complete trajectory parameters.

[0194] Step S1498: Extract the spatial coordinates of all adjustment points in the trajectory from the preliminary correction trajectory dataset, record the steering angle, driving speed value, and harvesting component operation position parameters of each adjustment point, and obtain the adjustment point marking dataset according to the motion characteristics specification of the unmanned driving equipment.

[0195] Extract the points of change in direction angle, speed, and altitude from the preliminary corrected trajectory dataset as adjustment points. Record the coordinates (Xa, Ya), steering angle θ_turn, driving speed V_adj, and harvester position L_adj for each adjustment point. Ensure that θ_turn ≤ θ_max (maximum steering angle), V_adj is within the range [V_min, V_max], and L_adj is within the range [L_min, L_max] (component lifting / lowering range). Standardize the parameter format to floating-point numbers, retaining two decimal places, to form an adjustment point marker dataset.

[0196] Step S1499: Extract the ground bearing data set from the coupled model data, match the ground bearing data set with the adjustment point marker data set, and fine-tune the driving speed and harvesting component operation position parameters of the adjustment points in areas corresponding to soft soil and areas corresponding to pothole distribution to obtain the ground adaptation correction data set.

[0197] The coordinates of the adjustment point marker dataset are matched with the ground bearing capacity dataset. If the adjustment point is located in a soft soil area (BCI < 0.4), the speed V_adj is reduced by 5%, and the harvesting part position L_adj is increased by ΔL_ground = 0.1 meters; if it is located in a pitted area (D_pit > D_pit_th), the speed is reduced by 8%, and L_adj is increased by ΔL_ground = 0.2 meters. The fine-tuned parameters are recorded to form the ground adaptation correction dataset.

[0198] Step S14910: Merge the preliminary corrected trajectory dataset, the adjustment point marker dataset, and the ground adaptation correction dataset, integrate the adjusted trajectory parameters, and supplement the adjustment point information and parameters of speed and harvesting component operation position to the corresponding positions of the trajectory to obtain the dynamic corrected trajectory dataset.

[0199] The trajectory parameters (coordinates, velocity, altitude, and adjustment points) of the three datasets are fused in a coordinate sequence. Each trajectory point contains complete motion and operation parameters to form a dynamically corrected trajectory dataset, which can dynamically respond to moving obstacles and changes in ground conditions.

[0200] Step S1410: Integrate all information from the dynamic correction trajectory dataset, including the path coordinate sequence, speed adaptation value, harvesting component operation position adjustment value, connecting node coordinates, and steering angle, and form an optimized obstacle avoidance trajectory according to the format and dimensions of all parameters in accordance with the motion parameter constraint specifications of the unmanned driving equipment.

[0201] Extract the path coordinate sequence (X0, Y0)...(Xn, Yn), the velocity V_i of each point, the position L_i of the harvesting component, the coordinates of the connecting node, and the turning angle θ_i from the dynamically corrected trajectory dataset. Check and correct the parameters according to the motion parameter constraints of the unmanned vehicle (e.g., maximum turning angular velocity ω_max = θ_max / Δt, maximum acceleration a_max): if θ_i / Δt > ω_max, increase the turning time Δt; if |V_i - V_{i-1}| / Δt > a_max, adjust the acceleration. Standardize the format of all parameters (e.g., coordinates are rounded to three decimal places, velocity is in meters per second), and finally form the optimized obstacle avoidance trajectory, which is the final path scheme for the equipment to perform the harvesting operation.

[0202] Step S150: Based on the optimized obstacle avoidance trajectory, convert the unmanned driving equipment into execution instructions, adjust the driving direction and speed of the equipment according to the optimized obstacle avoidance trajectory, adjust the working position and harvesting frequency of the harvesting components according to the working position adaptation value of the harvesting components, and integrate and output the execution instructions of the unmanned driving equipment.

[0203] The optimized obstacle avoidance trajectory obtained in step S1410 is converted into execution commands for the unmanned driving equipment. For driving control, the trajectory coordinate sequence is converted into steering angle commands: for each sampling period Δt_samp, the direction angle θ_cmd = arctan[(Y_{i+1}-Y_i) / (X_{i+1}-X_i)] between the current position and the next trajectory point is calculated and output to the steering system; the speed command V_cmd is output to the drive system according to the speed adaptation value V_i of the trajectory. For harvesting component control, L_i is converted into displacement commands for the lifting cylinder (through the conversion relationship between L_i and cylinder displacement), and the harvesting frequency f_cut = (V_i × W × k1 × ρ) / N_cut (ρ is the forage density, N_cut is the number of cuts per square meter) is calculated based on the speed V_i and the lateral coverage size W × k1 and output to the harvesting component drive system. All control commands (steering angle, speed, cylinder displacement, harvesting frequency) are integrated, packaged into an execution command set according to time sequence, and sent to each execution unit of the equipment via bus to complete the unmanned obstacle avoidance operation of forage harvesting.

[0204] In one exemplary embodiment, an unmanned obstacle avoidance system for hay harvesting is provided. This unmanned obstacle avoidance system for hay harvesting can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 2As shown, the unmanned obstacle avoidance system for forage harvesting includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements an unmanned obstacle avoidance method for forage harvesting. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the shell of an unmanned obstacle avoidance system used for hay harvesting, or an external keyboard, touchpad, or mouse, etc.

[0205] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. An unmanned obstacle avoidance method for hay harvesting, characterized in that, The method includes: By combining the terrain features of the forage harvesting area with the operating parameters of unmanned driving operations, the cooperative adaptation relationship between the obstacle avoidance perception component and the harvesting component is adjusted and integrated to form cooperative adaptation parameters. Based on the collaborative adaptation parameters, the synchronous acquisition process of multiple sensing components is initiated to capture obstacle feature information, pasture growth information, and ground condition information in the pasture harvesting area, and integrate them to form a three-dimensional sensing dataset. A target coupling model was built based on the 3D perception dataset. In the target coupling model, the avoidance space corresponding to different types of obstacles was divided. The spatial coordinates of the core harvesting area and the avoidable area of ​​pasture harvesting were marked and integrated to form the coupling model data. By combining coupled model data and collaborative adaptation parameters, a dynamic obstacle avoidance trajectory is planned. The overall direction of the dynamic obstacle avoidance trajectory is adjusted to match the adaptation value of the harvesting component's operating position, so that the direction of the dynamic obstacle avoidance trajectory fits the obstacle avoidance space and the lateral coverage size of the equipment's single harvest. This integration yields an optimized obstacle avoidance trajectory. Based on the optimized obstacle avoidance trajectory, the unmanned driving equipment executes commands. The driving direction and speed of the equipment are adjusted according to the optimized obstacle avoidance trajectory. The working position and harvesting frequency of the harvesting components are adjusted according to the adaptation value of the harvesting component's working position. The unmanned driving equipment executes commands in an integrated manner.

2. The unmanned obstacle avoidance method for forage harvesting according to claim 1, characterized in that, The synchronous acquisition process of multiple sensing components is initiated based on the collaborative adaptation parameters to capture obstacle feature information, pasture growth information, and ground condition information within the pasture harvesting area, forming a three-dimensional sensing dataset, including: Based on the collaborative adaptation parameters, activate the visual perception component, terrain perception component, material perception component, and soil perception component. Set the collection range, collection frequency, and data accuracy of each perception component according to the requirements of forage harvesting operations. Start the synchronous collection action of each perception component to obtain visual collection data, terrain collection data, material collection data, and soil collection data respectively. Visual data, terrain data, material data, and soil data are imported into a multi-source data processing workflow. The data content is compared point by point according to the spatial coordinates of the harvesting area. Conflicting data fragments from different data sources are eliminated, and core information such as obstacle outlines, ground undulations, material hardness, soil compaction, and pasture distribution is retained and integrated into a preliminary fusion dataset. Obstacle contour information is extracted from the preliminary fusion dataset. Contour completion is performed on the blurred edges of the visual acquisition data caused by grass obstruction. The planar projection offset value of the obstacle position in the terrain acquisition data is calculated by combining the terrain slope and sensor view parameters. The offset value is then added to the corresponding spatial coordinate position of the obstacle contour information to obtain the contour correction dataset. Extract pasture growth-related information from the contour correction dataset, perform spatial labeling operation on pasture growth areas, divide pasture growth areas according to the coordinate range of plant distribution, record the spatial coordinates of pasture plant height and coverage area in each area, and integrate all labeling results according to the spatial block of harvesting area to form pasture partition dataset. Based on the contour correction dataset, mobile obstacles in the pasture harvesting area are captured, and the movement trajectory segments and real-time movement status of the mobile obstacles are continuously recorded. The material acquisition data is matched with the spatial coordinates of the mobile obstacles, and the corresponding material hardness attribute is labeled for each mobile obstacle. All the information of the mobile obstacles is integrated according to the time series to form a mobile obstacle trajectory dataset. The pasture partition dataset is correlated with the acquisition parameters of each sensing component. The acquisition frequency of the visual sensing component, terrain sensing component, material sensing component, and soil sensing component is adjusted according to the plant distribution coordinate range in the pasture partition dataset. The acquisition frequency is increased in the core area of ​​pasture plant distribution and the area around moving obstacles, and the acquisition frequency is decreased in the edge area of ​​pasture plant distribution, and integrated into a frequency-optimized dataset. Soil compaction distribution information is extracted from soil data, and the spatial range in which equipment is prone to sinking is marked according to the distribution characteristics of soil compaction. Spatial coordinates of pothole distribution are extracted from topographic data and matched with the sinking area. Topographic feature information of the sinking area is supplemented, and the ground risk dataset is integrated. The contour correction dataset, pasture zoning dataset, moving obstacle trajectory dataset, frequency optimization dataset, and ground risk dataset are imported into the data integration process. The data format, data dimensions, and spatial coordinate correspondence of all datasets are unified. All data content is integrated according to the spatial block division of pasture harvesting area to form an integrated perception dataset. Obstacles hidden by grass and located in potholes are captured from the integrated perception dataset. Visual acquisition data, terrain acquisition data, and material acquisition data of the corresponding area of ​​the obstacle are extracted. The feature information of various types of data is combined to infer the outline size and material hardness of the corresponding obstacle. All information of the corresponding obstacle is integrated according to spatial coordinates to form a supplementary dataset of hidden obstacles. The integrated perception dataset and the supplementary dataset for hidden obstacles are imported into the spatial coordinate matching process. The obstacle feature information, pasture growth information, and ground condition information at the same spatial location are matched point by point. The information of hidden obstacles is supplemented to the corresponding position of the integrated perception dataset, and the three-dimensional perception dataset is obtained.

3. The unmanned obstacle avoidance method for forage harvesting according to claim 1, characterized in that, The target coupling model is constructed based on a 3D perception dataset. Within this model, avoidance spaces corresponding to different types of obstacles are divided. The spatial coordinates of the core harvesting area and the avoidable area for forage harvesting are marked, and these coordinates are integrated to form the coupling model data, including: Obstacle feature information is extracted from the 3D perception dataset. Obstacles are classified into rigid obstacles and flexible obstacles according to material hardness and contour size. The contour size and edge distribution of rigid obstacles are recorded for each obstacle, and the contour boundary and deformation buffer space requirements of flexible obstacles are recorded for each obstacle. All information is integrated according to obstacle type to form an obstacle classification dataset. Extract pasture growth information from the 3D perception dataset, filter the core pasture growth area within the pasture harvesting area according to plant distribution and coverage, mark the spatial coordinates of the core pasture growth area and draw complete boundary lines for each area, determine the overall spatial range of the core harvesting area, and integrate the information of the core harvesting area according to spatial blocks to form a core area dataset. Ground condition information is extracted from the 3D perception dataset. The distribution information of soil compaction is matched with the information of ground undulation. The bearing capacity information of the ground in the pasture harvesting area is labeled by spatial coordinate. The distribution boundary of ground bearing capacity is drawn and the corresponding spatial coordinates are marked. The data are then integrated into a ground bearing capacity dataset. Based on the obstacle classification dataset, the avoidance space corresponding to rigid obstacles and flexible obstacles is divided. The avoidance distance is fixed and extended outward according to the outline size of the rigid obstacle, and the boundary coordinates of the avoidance space are drawn. The minimum avoidance distance is set according to the deformation buffer space requirement of the flexible obstacle, and the boundary coordinates of the avoidance space are drawn. The information of the avoidance space is integrated for each obstacle to form an obstacle avoidance dataset. The spatial coordinates of the core harvesting area are extracted from the core area dataset. The actual location of all obstacles is marked within the spatial range of the core harvesting area. The ground bearing capacity dataset is matched with the location of the obstacles. The bearing capacity information of the surrounding ground is marked for each obstacle. The obstacle information in the core harvesting area is integrated according to the spatial coordinates to form a regional obstacle dataset. The obstacle classification dataset, core area dataset, ground bearing capacity dataset, obstacle avoidance dataset, and regional obstacle dataset are imported into the model building process. The various datasets are mapped to the corresponding spatial locations of the target coupled model according to the actual spatial coordinates of the pasture harvesting area. The various datasets are then integrated to obtain the basic data of the model. Import the basic model data into the target coupling model for computation. Calculate the impact range of obstacles on forage harvesting operations within the core harvesting area for each obstacle. Mark the spatial coordinates of the overlapping parts of the obstacle avoidance space and the core harvesting area. Integrate relevant information on the impact range according to obstacle type to form an impact range dataset. Based on the impact range dataset, adjust the boundary coordinates of the core harvesting area and the boundary coordinates of the obstacle avoidance space, maintain the safe range of the obstacle avoidance space, reduce the area occupied by the obstacle avoidance space in the core harvesting area, redraw the adjusted area boundary and mark the coordinates, and integrate them into a boundary adjustment dataset. In the target coupling model, dynamic adjustment nodes are set for the local differences between the motion state of the moving obstacle and the ground state. The spatial coordinates of the dynamic adjustment nodes are marked for each node, the triggering conditions and data update range of each node are recorded, and all information of the dynamic adjustment nodes are integrated according to their spatial location to form a model node dataset. The influence range dataset, boundary adjustment dataset, and model node dataset are supplemented to the corresponding spatial location of the target coupled model, and the information of the basic model data and various adjustment datasets are integrated to form coupled model data.

4. The unmanned obstacle avoidance method for forage harvesting according to claim 1, characterized in that, The process involves combining coupled model data and collaborative adaptation parameters to plan a dynamic obstacle avoidance trajectory. The overall direction of the dynamic obstacle avoidance trajectory is adjusted to match the adaptation value of the harvesting component's operating position, ensuring the trajectory conforms to the obstacle avoidance space and the lateral coverage size of a single harvest. This integration yields an optimized obstacle avoidance trajectory, including: The coupled model data and collaborative adaptation parameters are imported into the trajectory planning basic processing flow. Information on obstacle avoidance space and core harvesting area boundary is extracted from the coupled model data. Information on the lateral coverage size, travel speed and harvesting component operation position of a single harvest is extracted from the collaborative adaptation parameters. The data are then integrated into trajectory planning basic data according to the operation requirements to form the trajectory planning basic dataset. Based on the trajectory planning dataset, the initial obstacle avoidance trajectory within the work area is drawn. The overall direction of the initial obstacle avoidance trajectory is planned according to the coordinates of the obstacle avoidance space, so that the entire path of the initial obstacle avoidance trajectory avoids all obstacle avoidance spaces, while covering the entire spatial range of the core harvesting area. The information of the initial obstacle avoidance trajectory is integrated according to the coordinate sequence to form the initial trajectory dataset. The spatial coordinates of the turning points of the initial obstacle avoidance trajectory are extracted from the initial trajectory dataset. Smoothing operations are performed on all the turning points of the initial obstacle avoidance trajectory. The coordinate positions of the turning points are adjusted to reduce the sudden angle of the device's driving direction so that the adjusted trajectory conforms to the motion parameter constraints of the autonomous driving device. The results are then integrated into a smooth trajectory dataset. Dynamic adjustment nodes are extracted from the model node dataset from the coupled model data. Data update nodes are set at the corresponding spatial locations in the smooth trajectory dataset. A time window for trajectory correction is reserved for each data update node according to the motion state of the moving obstacle. The coordinates and triggering conditions of the data update nodes are marked to form a node-adapted trajectory dataset. Extract driving speed related data from the collaborative adaptation parameters, adjust the driving speed adaptation value of road segments close to the obstacle avoidance space according to the road segment division of the node adaptation trajectory dataset, set the uniform speed driving adaptation value for road segments in the core harvesting area, and label the speed adaptation value for each road segment to form a speed adaptation trajectory dataset. The location coordinates of the flexible obstacle are extracted from the coupled model data. The trajectory around the flexible obstacle is fine-tuned in the velocity adaptation trajectory dataset to reduce the area occupied by the flexible obstacle avoidance space, retain the working space of the core harvesting area, and integrate them into a flexible adaptation trajectory dataset. The ground bearing capacity dataset is extracted from the coupled model data. The coordinates of soil compaction and pothole distribution are matched with the coordinate sequence of the flexible adaptation trajectory. The adaptation values ​​of the harvester operation position for soft soil road sections and pothole distribution road sections are adjusted to improve the adaptation values ​​of the harvester operation position for the corresponding road sections, thus forming a highly adapted trajectory dataset. Extract all coordinate sequences of the trajectory from the highly adapted trajectory dataset, divide the trajectory into core harvesting segments and obstacle avoidance segments, mark the spatial coordinates of the connection points between the core harvesting segments and obstacle avoidance segments segment by segment, set the driving speed switching value for each connection point, and integrate all the information of the segment division and connection points to form a segment-marked trajectory dataset. The motion state information of moving obstacles is extracted from the coupled model data. The direction of the obstacle avoidance road segments in the road segment marker trajectory dataset is adjusted according to the motion trend of the moving obstacles to reserve buffer space for the motion of the moving obstacles. The coordinate sequence of the obstacle avoidance road segments is adjusted to fit the motion characteristics of the moving obstacles, forming a dynamically corrected trajectory dataset. The system integrates all information from the dynamic correction trajectory dataset, including the path coordinate sequence, speed adaptation value, harvesting component operation position adjustment value, connecting node coordinates, and steering angle. It then forms an optimized obstacle avoidance trajectory by conforming to the motion parameter constraint specifications of unmanned driving equipment in terms of the format and dimensions of all parameters.

5. The unmanned obstacle avoidance method for forage harvesting according to claim 2, characterized in that, The process involves extracting obstacle contour information from the initial fusion dataset, performing contour completion on blurred areas in the visually acquired data caused by overgrazing, calculating the planar projection offset of obstacle locations in the terrain acquisition data based on terrain slope and sensor viewpoint parameters, and then supplementing the corresponding spatial coordinates of the obstacle contour information with the offset values. This results in a contour-corrected dataset, including: Obstacle contour information and visual acquisition data are extracted from the preliminary fusion dataset. The blurred edge areas caused by grass occlusion in the visual acquisition data are marked by spatial coordinates. The spatial range and corresponding coordinate information of the blurred areas are recorded. The marking results of the blurred areas are integrated according to the obstacle type to form a blurred marking dataset. The fuzzy labeled dataset is matched with the material acquisition data. The contour extension method is used to complete the obstacle contour in the fuzzy area. The contour shape is extended according to the different shapes of rigid and flexible obstacles. The completed contour fits the actual material characteristics of the obstacle and is integrated into the contour extension dataset. Extract terrain data and ground undulation information from the preliminary fusion dataset. Combine terrain slope and sensor view parameters to calculate the planar projection offset of obstacle position in the perception data due to ground undulation. Record the direction and specific amount of offset. Integrate the offset calculation results according to obstacle coordinates to form an offset calculation dataset. The contour extension dataset and offset calculation dataset are imported into the obstacle information correction process. The original obstacle information in the visual acquisition data is replaced with the completed obstacle contour and the corrected position coordinates. All the information of the obstacle contour is re-integrated according to the spatial coordinates to form a preliminary correction dataset. Hardness feature information is extracted from the material acquisition data. The hardness feature information is matched with the preliminary correction dataset to verify the fit between the corrected obstacle contour shape and the material hardness. For areas that do not fit, the detailed coordinates of the contour are fine-tuned so that the contour shape fits the material hardness, thus forming a verification correction dataset. The contour information of all obstacles is extracted from the verification and correction dataset. The corner positions and dimensions of rigid obstacles are marked for each obstacle, and the edge curvature of flexible obstacles is marked for each obstacle. Various detailed feature information of obstacle contours is supplemented. The marking results of detailed features are integrated according to obstacle coordinates to form a detailed supplement dataset. The detailed supplementary dataset and the preliminary correction dataset are integrated, and the detailed features are added to the corrected obstacle contour information to improve the morphological description of the obstacle contour. All the information of the obstacle contour is then reorganized according to spatial coordinates to form the contour optimization dataset. Color feature information is extracted from visual acquisition data, and the color feature information is matched with the contour optimization dataset to correct the color deviation in the contour optimization dataset, enhance the color contrast between obstacles and pasture, so that the visual presentation of obstacle contours is clearer, and integrate them into a color correction dataset. The color correction dataset and the preliminary fusion dataset are imported into the data integration process. The pasture growth information and ground condition information in the preliminary fusion dataset are retained, while redundant and duplicate data generated during the contour correction process are removed. All information is integrated according to spatial coordinates to form the contour correction base dataset. By integrating all the corrected obstacle contour information, position coordinates, and detailed features from the contour correction base dataset, and unifying the data format and spatial coordinate correspondence, a contour correction dataset is formed.

6. The unmanned obstacle avoidance method for forage harvesting according to claim 3, characterized in that, The obstacle avoidance dataset is divided into rigid and flexible obstacle avoidance spaces based on the obstacle classification dataset. A fixed avoidance distance is extended outwards according to the outline dimensions of the rigid obstacle, and the boundary coordinates of the avoidance space are plotted. A minimum avoidance distance is set according to the deformation buffer space requirements of the flexible obstacle, and the boundary coordinates of the avoidance space are plotted. The avoidance space information is integrated for each obstacle to form an obstacle avoidance dataset, including: Extract all relevant information about rigid obstacles from the obstacle classification dataset, record the length, width, and height of each rigid obstacle, accurately mark the spatial coordinates of the outermost contour point of the rigid obstacle, and integrate the feature information of the rigid obstacles according to the obstacle number to form a rigid contour dataset. Based on the rigid contour dataset, a fixed avoidance distance is extended outward for each rigid obstacle according to the operational requirements of the unmanned driving equipment. The boundary coordinates of the avoidance space are drawn for each obstacle, and the range, area and shape of the avoidance space for each rigid obstacle are recorded. The information is integrated according to the obstacle number to form a rigid avoidance dataset. Extract all relevant information about flexible barriers from the obstacle classification dataset, record the contour boundary and spatial requirements of deformation buffer for each flexible barrier, mark the specific spatial range of deformation buffer of flexible barriers in combination with material hardness, and integrate the feature information of flexible barriers by barrier number to form a flexible characteristic dataset. Based on the flexible characteristic dataset, a minimum avoidance distance is set for each flexible obstacle. The set minimum avoidance distance covers the deformation buffer space of the flexible obstacle. The boundary coordinates of the avoidance space are drawn for each obstacle. The range area of ​​the avoidance space of each flexible obstacle is recorded. The information is integrated according to the obstacle number to form a flexible avoidance dataset. The rigid obstacle avoidance dataset and the flexible obstacle avoidance dataset are imported into the spatial comparison process. The coordinates of the overlapping area of ​​the rigid obstacle avoidance space and the flexible obstacle avoidance space within the operation area are marked. The original range of the rigid obstacle avoidance space is preserved. The boundary coordinates of the flexible obstacle avoidance space are adjusted according to the overlapping area to form the overlapping adjustment dataset. Extract all information on ground bearing capacity from the ground bearing capacity dataset, match the ground bearing capacity information with the rigid obstacle avoidance dataset, expand the range of rigid obstacle avoidance space in soft soil sections and pothole-prone sections, and maintain the original range of rigid obstacle avoidance space in sections with good ground bearing capacity, thus forming a ground-adaptive avoidance dataset. Extract all information on material hardness from the obstacle classification dataset, match the material hardness information with the ground adaptation avoidance dataset, verify the fit between avoidance distance and material hardness, adjust the boundary coordinates of the avoidance space according to the material hardness so that the avoidance distance fits the material hardness, and form a hardness adaptation dataset. Extract the boundary coordinates of the core harvesting area from the core area dataset, match the boundary coordinates of the core harvesting area with the hardness adaptation dataset, mark the intersection coordinates of the obstacle avoidance space boundary and the core harvesting area boundary, record the number and distribution of the intersection points, integrate the intersection point information according to the spatial coordinates, and form an intersection point marking dataset. The rigid avoidance dataset, flexible avoidance dataset, overlap adjustment dataset, ground-adaptive avoidance dataset, hardness-adaptive dataset, and intersection mark dataset are imported into the data integration process. All relevant information about obstacle avoidance space is integrated according to obstacle type, and the data format and spatial coordinate correspondence are unified to form an integrated avoidance space dataset. The boundary coordinates, area, and adaptation conditions of all obstacle avoidance spaces in the standard obstacle avoidance space integration dataset are used to form an obstacle avoidance dataset.

7. The unmanned obstacle avoidance method for forage harvesting according to claim 4, characterized in that, The process involves extracting ground bearing capacity datasets from coupled model data, matching the coordinates of soil compaction and pothole distribution with the coordinate sequences of flexible adaptation trajectories, adjusting the adaptation values ​​of harvester operation positions for soft soil sections and pothole-distributed sections, improving the adaptation values ​​of harvester operation positions for corresponding sections, and forming a highly adapted trajectory dataset, including: The ground bearing capacity dataset was extracted from the coupled model data. The forage harvesting operation area was divided into soft soil sections, pothole sections, and flat ground sections according to the distribution of soil compaction and the coordinates of pothole distribution. The boundary coordinates of each section were marked, and the soil compaction value and pothole depth parameter were recorded to form a ground zoning dataset. The complete coordinate sequence of the trajectory is extracted from the flexible adaptation trajectory dataset. The trajectory coordinate sequence is accurately matched with the spatial coordinates of the ground partition dataset. The coordinates of each road segment through which the trajectory passes, such as soft soil road segment, pothole-distributed road segment, and flat ground road segment, are marked. The matching results are integrated by road segment to form a trajectory ground association dataset. Extract the portion of the trajectory that passes through the soft soil section from the trajectory ground association dataset, match the soil compaction value with the coordinates of the corresponding section, adjust the adaptation value of the harvesting component operation position according to the soil compaction, increase the adaptation value of the harvesting component operation position of the corresponding section according to the soil softness, and label the adaptation value for each coordinate to form a soft area height dataset. Extract the portion of the track that passes through the pothole distribution section from the trajectory ground association dataset, match the pothole depth and width parameters with the coordinates of the corresponding section, adjust the adaptation value of the harvesting component operation position according to the pothole parameters, increase the adaptation value of the harvesting component operation position of the corresponding section according to the pothole depth, reserve buffer space for the pothole edge, and label the adaptation value for each coordinate to form a pothole area height dataset. Extract the portion of the trajectory that passes through the flat road section from the trajectory ground association dataset, extract the original adaptation value of the harvesting component operation position from the cooperative adaptation parameters, use the original adaptation value in the corresponding part of the road section, and label the adaptation value of the harvesting component operation position coordinate by coordinate to form a flat area height dataset. The height datasets of soft areas, potholes, and flat areas are matched point by point with the coordinate sequence of the flexible adaptation trajectory. The corresponding harvesting component operation position adaptation value is labeled for each trajectory coordinate point. The information of all adaptation values ​​is integrated according to the coordinate sequence to form the height adaptation basic dataset. Extract the coordinates of the switching nodes for the adaptation values ​​of the harvesting component operation position from the highly adapted base dataset, adjust the spacing between the switching nodes, label the gradient of the adaptation value change, and form a highly smooth dataset. Obstacle location information is extracted from the coupled model data. The obstacle location information is matched with the height smoothing dataset. The adaptation value of the harvesting component operation position of the trajectory coordinate points around the obstacle is fine-tuned in the height smoothing dataset. The adaptation value of the corresponding coordinate points is increased to avoid the obstacle, thus forming an obstacle adaptation height dataset. Extract the spatial coordinates of all harvesting component operation position adjustment nodes from the obstacle adaptation height dataset, record the harvesting component operation position adaptation value and switching speed of each adjustment node, and form a height node labeling dataset according to the component motion parameter constraint specification of unmanned driving equipment. The height node marker dataset and the flexible adaptation trajectory dataset are integrated. The adaptation values ​​of the harvesting component operation position and the information of the height adjustment nodes are added to the corresponding positions of the flexible adaptation trajectory, replacing the harvesting component operation position parameters in the original trajectory, thus forming the height adaptation trajectory dataset.

8. The unmanned obstacle avoidance method for forage harvesting according to claim 2, characterized in that, The process involves capturing obstacles hidden by pasture and located in potholes from the integrated perception dataset, extracting visual acquisition data, terrain acquisition data, and material acquisition data for the corresponding areas of the obstacles, combining the feature information of various data to infer the outline size and material hardness attributes of the corresponding obstacles, and integrating all information of the corresponding obstacles according to spatial coordinates to form a supplementary dataset for hidden obstacles, including: Extract pasture zoning dataset and ground risk dataset from integrated sensing dataset, mark the spatial boundary coordinates of the core area of ​​pasture plant distribution and the area of ​​pothole distribution, record the relevant data of pasture plant height and pothole depth in each area, and integrate the marking results according to spatial blocks to form hidden area marking dataset. By matching the integrated perception dataset and the hidden area labeling dataset, visual acquisition data, terrain acquisition data, and material acquisition data corresponding to the core distribution area of ​​pasture plants and the distribution area of ​​potholes are selected. The content of various acquisition data is integrated according to the region to form the hidden area acquisition dataset. Topographic data was extracted from the dataset collected from hidden areas. Combined with the numerical characteristics of ground elevation changes, the locations of completely obscured obstacles in the core area of ​​pasture distribution were inferred. The elevation changes caused by the height of the pasture itself were removed, and the spatial coordinates of the inferred obstacles were marked to form a dataset of obscured obstacle locations. The dataset of occlusion location is matched with the dataset of hidden area collection. The hardness characteristics of material collection data and the color residue information of visual collection data are combined to infer the outline size and material hardness attributes of the obstacles occluded by pasture at each location. The inference results are integrated by coordinate to form the occlusion feature dataset. Topographic data is extracted from the hidden area dataset. Combined with the spatial coordinate features of pothole morphology, the actual location of obstacles within the pothole distribution area is inferred. The differences in topographic features between obstacles and pothole edges are distinguished. The spatial coordinates of the inferred obstacles are marked to form a pothole obstacle location dataset. The dataset of pothole obstacle locations is matched with the dataset of hidden areas, and the hardness attribute of the material data and the compaction difference information of the soil data are supplemented. The material hardness attribute of the obstacle in the pothole distribution area is inferred at each location. The inference results are integrated by coordinate to form a pothole obstacle feature dataset. The occlusion obstacle feature dataset and the pothole obstacle feature dataset are imported into the data integration process. The location coordinates, outline size and material hardness attributes of all hidden obstacles are integrated. All information is integrated in blocks according to the spatial coordinates of the pasture harvesting area to form the basic dataset of hidden obstacles. Extract other obstacle information within the working area from the integrated perception dataset, match the other obstacle information with the hidden obstacle base dataset, and adjust the outline size and position coordinates of the obstacles in the hidden obstacle base dataset according to the spatial distribution of other obstacles to avoid conflicts with the spatial distribution of other obstacles, thus forming a hidden obstacle correction dataset. Extract all inferred obstacle information from the hidden obstacle correction dataset, label the data source for the inferred obstacle outline size and material hardness properties, mark the spatial coordinates of the uncertain areas of the data inference, and integrate the labeling results by coordinate to form a hidden obstacle labeling dataset; By merging the hidden obstacle correction dataset and the hidden obstacle labeling dataset, and integrating information such as the location coordinates, outline dimensions, material hardness attributes, and data source of all hidden obstacles, the data format and the correspondence between spatial coordinates are standardized to obtain the hidden obstacle supplementary dataset.

9. An unmanned obstacle avoidance system for forage harvesting, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the unmanned obstacle avoidance method for hay harvesting as described in any one of claims 1 to 8 by executing the machine-executable instructions.

10. A computer program product, characterized in that, The computer program product includes machine-executable instructions stored in a computer-readable storage medium. A processor of the unmanned obstacle avoidance system for hay harvesting reads the machine-executable instructions from the computer-readable storage medium and executes the machine-executable instructions, causing the unmanned obstacle avoidance system for hay harvesting to perform the unmanned obstacle avoidance method for hay harvesting as described in any one of claims 1 to 8.

Images