Non-crop field ridge navigation line extraction method, control method and system

Abstract

The invention relates to the technical field of agricultural machinery visual navigation, and discloses a crop-free field ridge navigation line extraction method, control method and system, and the method employs instance segmentation, dynamic middle axis extraction, segmented least square, and application of an inter-segment continuity constraint link to achieve the extraction of a global navigation curve. The problems that in the prior art, a crop-free field ridge navigation line extraction method is insufficient in end-side real-time performance and prone to instability under complex interference can be solved.

Description

Technical Field

[0001] This invention relates to the field of agricultural machinery visual navigation technology, and discloses a method, control method and system for extracting navigation lines in fields without crops. Background Technology

[0002] Existing visual navigation and ridge recognition solutions mostly follow two paths: "detection-fitting" or "traditional image processing-geometric regression". One approach, such as the Chinese patent application CN120274747A, discloses a navigation path detection method for inter-row operations. This method uses a lightweight YOLO derived network (VPDX_YOLOv8n) to generate crop row prediction boxes and performs recursive least-squares curve fitting using the set of geometric midpoints of the boxes. While this method has the advantages of simple link and low inference overhead, the approach of "using the box center instead of the row axis" is easily affected by box size / occlusion, exhibiting systematic biases in scenarios with no crops, jagged boundaries, or broken rows. Furthermore, the global fitting lacks inter-segment continuity constraints, leading to abrupt curve changes during ridge bending, which is detrimental to smooth output of control quantities. Secondly, as proposed in Chinese patent application CN120526303A, a cascaded encoder-decoder approach is used to segment seedlings / weeds and extract field ridge boundary lines. Semantic segmentation and connected component analysis are then performed to support operation planning. This approach emphasizes crop-weed semantics and boundary derivation, making it suitable for weed management and field identification. However, the network scale and processing chain are relatively complex, putting pressure on the real-time performance and power consumption of the end-side. Furthermore, the target products are mostly category masks and boundaries, without directly outputting navigation quantities such as bottom edge intersections and bottom tangent angles that can be called by the vehicle controller. Secondary geometric modeling is still required for real-time ridge walking. Thirdly, as proposed in Chinese patent application CN120628084A, a binary map is obtained based on HSV-V grayscale enhancement, dynamic thresholding, and morphology. Then, bidirectional boundary scanning and linear regression are used to construct the left and right boundaries and calculate the heading deviation. This approach has a low engineering implementation threshold and is usable on the end-side. However, this route is sensitive to conditions such as strong light / shade / straw cover. The linear regression model is difficult to track stably when the boundary breaks or the ridges are nonlinearly curved, and its robustness and curvature adaptation capabilities are limited.

[0003] It is evident that existing methods for extracting navigation lines in uncropped fields suffer from insufficient real-time performance at the end and instability under complex interference. Summary of the Invention

[0004] This invention provides a method, control method, and system for extracting navigation lines in uncropped fields, to solve the problems of insufficient real-time performance at the end side and instability under complex interference in existing methods for extracting navigation lines in uncropped fields.

[0005] Firstly, this application provides a method for extracting navigation lines in crop-free fields, including: S1: Obtain the original image of the field ridge area in the forward direction, and determine the input frame based on the original image; S2: An improved single-stage instance segmentation network model is used to perform pixel-level segmentation of the ridge region in the input frame to obtain an instance mask; S3: Perform morphological opening operation on the instance mask and extract the main central axis point sequence in a row-by-row sliding window manner. Remove outliers by interquartile range method and perform one-dimensional smoothing on the main central axis point sequence to obtain the central axis point sequence. S4: Divide the central axis point sequence into several parts along the longitudinal direction of the original image, perform least squares fitting for each part of the central axis points, and apply continuity constraints between the central axis points of each part to form a global navigation curve.

[0006] Secondly, this application provides a control method for vehicles operating in crop-free field ridges, including: The global navigation curve is extracted using the cropless field ridge navigation line extraction method described in the first aspect; Based on the global navigation curve, calculate its intersection point with the bottom edge of the image and the tangent direction at that intersection point to obtain the yaw angle. Calculate its distance from the midpoint of the bottom edge of the image to obtain the pixel error. A validity flag is generated based on the navigation curve, pixel error, and yaw angle, and vehicle control is performed when the validity flag is valid.

[0007] Thirdly, this application provides a cropless field ridge navigation line extraction system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0008] Fourthly, this application provides a control system for a crop-free field ridge vehicle, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the second aspect above.

[0009] The present invention has the following beneficial effects: The method for extracting navigation lines in uncropped fields proposed in this application uses instance segmentation, dynamic central axis extraction, and a link structure with piecewise least squares and inter-segment continuity constraints to extract global navigation curves. This method can solve the problems of insufficient real-time performance at the end side and instability under complex interference in existing methods for extracting navigation lines in uncropped fields.

[0010] In a further technical solution, when extracting navigation curves, noise and breaks are suppressed through sliding window centroid, IQR rejection, and sequence smoothing. Constraints can prevent abrupt changes in the curve.

[0011] In a further technical solution, the unified output can be directly used for the controller's navigation parameters, thereby achieving synergistic optimization in terms of end-side latency, geometric robustness, and control availability.

[0012] In a further technical solution, introducing longitudinal receptive field enhancement and orientation recalibration into the network model can improve the expression of strip-shaped ridges.

[0013] In addition to the objectives, features and advantages described above, the present invention has other objectives, features and advantages.

[0014] The present invention will now be described in further detail with reference to the figures. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a preferred embodiment of the present invention for extracting navigation lines in a field without crops; Figure 2 This is a schematic diagram of the Ghost-DW module according to a preferred embodiment of the present invention; Figure 3 This is a schematic diagram of the Dir-SE module according to a preferred embodiment of the present invention; Figure 4 This is a flowchart of the navigation line extraction process according to a preferred embodiment of the present invention, wherein (a) is a schematic diagram of dynamic extraction of the central axis point, and (b) is a schematic diagram of segmented fitting of the navigation line; Figure 5 This is a schematic diagram of the system structure of a preferred embodiment of the present invention; Figure 6 This is a schematic diagram of the automatic control system for the differential steering clutch of a transplanter according to a preferred embodiment of the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a," and similar terms, do not indicate a quantity limitation, but rather indicate the presence of at least one.

[0018] It should be understood that currently, the key technical challenges faced by fully automated transplanters in ridge-walking scenarios without crop ridges include: insufficient stability in ridge target extraction under complex conditions such as changes in moisture levels, stubble obstruction, mud and water reflection, field edges, and broken ridges; limited sensitivity of conventional convolutional structures to the directional geometry of longitudinal strip-shaped ridges, leading to segmentation boundary jitter and missed detections; difficulty in achieving low-latency inference under the constraints of computing power and power consumption on edge computing platforms; discontinuities in the central axis point during mask skeletonization due to noise and local gaps; and insufficient error diffusion, curve continuity, and smoothness in single-segment global fitting under irregular boundary conditions. Therefore, a visual perception method that can run in real-time at the edge and is robust to complex disturbances is needed to stably generate the aforementioned navigation quantities and meet the closed-loop control requirements for ridge walking.

[0019] Please see Figure 1 This application provides a method for extracting navigation lines in cropless fields, comprising: S1: Obtain the original image of the field ridge area in the forward direction, and determine the input frame based on the original image.

[0020] In this step, the forward-looking image of the operation can be acquired based on the onboard camera of the edge platform. The original image is then subjected to size unification and normalization processing, and camera calibration parameters are loaded to obtain the input frame. Here, the forward-looking image of the operation refers to the close-up view of the field ridge.

[0021] S2: An improved single-stage instance segmentation network model is used to perform pixel-level segmentation of the ridge region in the input frame to obtain an instance mask.

[0022] In this step, the improved single-stage instance segmentation network model is named R-YOLO-Prune, which includes a backbone network, a neck network, and a head network connected in sequence. A Ghost deep convolution with a large vertical kernel is set at the bottleneck position of the backbone network and the neck network, and a direction-aware channel recalibration module is introduced before the head network.

[0023] The operational steps when using the improved instance segmentation network model include: replacing the original depthwise convolutions with Ghost depthwise convolutions (Ghost-DW, kernel selection of 7×1) at the bottleneck positions of the backbone and neck networks to expand the vertical receptive field and enhance the orientation sensitivity to strip-shaped ridge geometry; introducing an orientation-aware channel recalibration module (Dir-SE) before fusing multi-scale features to the segmentation / detection head, splitting global pooling into vertical statistical and horizontal preservation paths, applying 1D convolution and two fully connected layers to the 1×W orientation histogram, and obtaining channel weights through Sigmoid to recalibrate feature channels, suppressing horizontal noise and highlighting vertical structure; using BN-γ based sparsification training and performing structured pruning, pruning low-contribution channels according to channel importance thresholds, and fine-tuning the recovery to obtain a lightweight instance segmentation model; exporting the model to ONNX for INT8 quantization and mapping it to the inference engine of the edge platform for operation. The network model outputs bounding boxes, class confidence scores, mask coefficients, and prototype masks. The linear combination of these and thresholding yields the field instance mask.

[0024] S3: Perform morphological opening operation on the instance mask and extract the main central axis point sequence in a row-by-row sliding window manner. Remove outliers by interquartile range method and perform one-dimensional smoothing on the main central axis point sequence to obtain the central axis point sequence.

[0025] In this step, an opening operation is performed on the mask to remove spurs and isolated points, and the shape and scale of the structuring element are matched with the image resolution. Then, a sliding window is used from the bottom of the image upwards with a step size s, selecting the largest connected component in each row and calculating the horizontal centroid as the central axis point. The Q1 to Q3 quartile method is used on the horizontal coordinate sequence to remove outliers with a threshold of 1.5×IQR, and the retained sequence is smoothed by median or local mean to obtain a continuous, low-noise central axis point sequence.

[0026] S4: Divide the central axis point sequence into several parts along the longitudinal direction of the original image, perform least squares fitting for each part of the central axis points, and apply continuity constraints between the central axis points of each part to form a global navigation curve.

[0027] In this step, the central point sequence is divided into K segments (K = 4 to 6) at equal intervals along the longitudinal direction, and the local curve is obtained by least-squares fitting for each segment. and at the splicing point of adjacent segments Apply at least Continuity constraints At the same time, encourage with soft constraints Continuity, added to the objective function ( (as weights), thereby suppressing abrupt changes in seams and error propagation, forming a global navigation curve.

[0028] In one example, the above parameters can be configured as follows: vertical depth convolution kernel length 7×1, Ghost artifact channel ratio 0.5–1.0; sparse regularization coefficient The pruning threshold is set to the 50th percentile of the BN-γ distribution, and fine-tuning is performed for 300 rounds after pruning; the row-by-row step size s is 10 pixels, and the IQR coefficient is 1.5; the number of segments K is 5, and the fitting order is 11; the continuity constraint is... Required Soft constraints Weight Choose option 1. The input dimensions on the end side are fixed at 640×640. The above-mentioned method for extracting navigation lines in fertile fields without crops uses instance segmentation, dynamic central axis extraction, and segmented least squares combined with link-based continuous constraints between segments to extract global navigation curves. This method can solve the problems of insufficient real-time performance at the end side and instability under complex interference in existing methods for extracting navigation lines in fertile fields without crops.

[0029] The steps of the above-mentioned method for extracting navigation lines in fields without crops are described in detail below, using a complete example: First, the onboard camera on the edge platform acquires the forward image of the operation, performs size unification and normalization processing on the original image, and loads the camera calibration parameters to obtain the input frame.

[0030] In this application, the improved single-stage instance segmentation network model includes a backbone network, a neck network, and a head network connected in sequence. The backbone takes a normalized 640×640×3 image as input. After initial feature extraction via Input and two Conv layers, it enters multiple cascades of C3k2 and Conv, downsampling to form three semantically progressive features: P3 with a resolution of approximately 80×80 and a stride of 8 is obtained from C3k2 (Stage-P3); P4 with a resolution of approximately 40×40 and a stride of 16 is obtained by downsampling via Conv and then by C3k2 (Stage-P4); after further downsampling, it enters SPPF, and two structural modifications are inserted between this point and C2PSA: a Ghost depthwise separable convolution with a large vertical kernel (Ghost-DW, kernel size 7×1) and a direction-aware channel recalibration module (Dir-SE).

[0031] In this application, the processing flow of Ghost-DW is as follows: Figure 2 As shown, DWConv(7×1) is first executed, which aggregates only along the vertical direction. The output of the c-th channel is:

[0032] in Output to the main branch. Learnable vertical convolutional kernel weights, no horizontal convolution, output resolution preserved. ; The radius of the convolution kernel is 1. , These represent the vertical and horizontal coordinates of the feature map. This represents the vertical sampling offset of the convolution kernel. The results are then sequentially activated using BN and SiLU:

[0033] in The batch mean and variance of this channel. Scaling and biasing of BN. The output characteristics after BN and SiLU activation are shown. It is a constant.

[0034] Next, a Ghost linear transformation is performed to generate a cheap channel, which is then concatenated with the main channel. The output maintains the H×W size. The main channel is first obtained through a simple linear mapping. , and then Applying a cheap linear transformation Generate additional channels, and finally concatenate them according to the channel dimensions to obtain the Ghost output.

[0035] The processing flow of Dir-SE is as follows: Figure 3 As shown, the directional statistics of 1×W are obtained by first averaging the longitudinal direction of each channel. The longitudinal average of the feature X of the c-th channel is: In the formula, The height of the feature map, The width of the feature map. Indicates the first The pixel values ​​of each channel. For vertical pixel coordinate index, Horizontal pixel coordinate index; Then perform a one-dimensional convolution in the horizontal direction: The channel weights α are obtained by performing GAP_w, FC1 and ReLU, and FC2 and Sigmoid, and the original features are recalibrated to suppress lateral texture noise and highlight the strip structure.

[0036] In this application, the horizontal global average processing satisfies the following relationship: Let the input features be ,in For the number of channels, For feature map height, The width of the feature map; the first The passage is in the location The pixel value is denoted as , , ; for the first The channel is obtained by vertical averaging and horizontal convolution:

[0037] Again Dimensions are used for horizontal global averaging After processing, we obtain:

[0038] In the formula, It is a horizontal global average. As a dimension, No. transverse response sequence of the channel The channel statistics vector obtained by horizontal global averaging Number of channels; It is the symbol for the set of real numbers. Representing vectors The dimension is ; Then, the weight scalar for each channel is obtained through two fully connected layers. Let the compression ratio be... ,but

[0039] The channel weights obtained after Sigmoid normalization are as follows:

[0040] In the formula, The compressed intermediate channel vector The weights and biases of the two fully connected layers; For unnormalized channel scores; FC1 performs dimensionality reduction on the channel dimension ( ReLU is activated by linear rectification, and FC2 restores the channel dimension ( Finally, it is normalized to Sigmoid. :

[0041] Will Broadcast to the spatial dimension, and perform channel recalibration on the original features:

[0042] In the formula, For the first The passage is The recalibrated pixel value at the location; For the first Channel weighting coefficient For the first The passage is The original pixel value at that location.

[0043] After completing the above two modifications, the image enters C2PSA, forming a P5 with a resolution of approximately 20×20 and a step size of 32. Both modifications are located after SPPF and before C2PSA.

[0044] Neck employs an FPN / PAN fusion structure: First, P5 is upsampled and concatenated with P4 at the Cat node, then passed through a C3k2 layer to obtain mid-scale fused features. Subsequently, it is upsampled again, concatenated with P3 at the Cat node, and passed through another C3k2 layer to obtain high-resolution fused features. Simultaneously, a side branch from the mid-scale C3k2 is set up, connected in parallel with the main branch through a Conv layer to supplement semantic and detail information. Thus, Neck outputs three feature paths, corresponding to the upper and middle layer C3k2 nodes and the bottom layer Conv node, respectively, corresponding one-to-one with the three inputs of the Head. The Head consists of multi-scale Detect sub-headers: top-level features pass through C3k2 and enter Detect, outputting Predict1; middle-level features pass through C3k2 and enter Detect, outputting Predict2; bottom-level features pass through Conv and enter Detect, also outputting Predict2. The decoupling implementation of Detect is consistent with CV2 and CV3: the classification branch consists of two Conv layers and one Conv2D layer, and calculates ClsLoss; the regression branch consists of Conv, DWConv, another Conv layer, a second DWConv, and Conv2D, and calculates BoxLoss. During the training phase, the classification and regression losses are calculated separately for the CV2 and CV3 branches, and BN-γ sparsity and lightweight strategies can be combined; during the deployment phase, the ONNX model is exported and mapped to the edge inference engine.

[0045] It should be noted that Ghost-DW (7×1) and Dir-SE are deployed only after Backbone's SPPF and before C2PSA.

[0046] It is worth explaining that channel recalibration refers to the process of dynamically enhancing useful features and suppressing redundant information by adaptively calculating the importance weights of each channel after convolutional feature extraction and then reapplying these weights to the feature map channels.

[0047] In other words, in this application, the improved single-stage instance segmentation network model includes a backbone network, a neck network, and a head network connected in sequence. A Ghost depthwise convolution with a large vertical kernel is placed at the bottleneck positions of the backbone and neck networks. A direction-aware channel recalibration module is placed before the head network. The composition and function of each module are as follows: (1) Backbone network: It consists of Input, two Conv layers, multi-level C3k2, SPPF, vertical large kernel Ghost depthwise separable convolution (Ghost-DW, 7×1), orientation-aware channel recalibration module Dir-SE and C2PSA concatenated in sequence.

[0048] Input / Conv: Completes pixel normalization and initial feature extraction; C3k2: This is a residual-type cross-stage convolution module that uses a bottleneck structure of 1×1 and 3×3 convolutions and some cross-layer connections to handle channel compression, nonlinear representation, and feature reuse. SPPF (Spatial Pyramid Pooling-Fast): Enhances the perception of long strip textures and curved ridges by pooling large receptive field contexts through multi-scale pooling. Ghost-DW (7×1): Depth-separable convolutions aggregated only along the vertical axis of the image, followed by BN and SiLU, and generating inexpensive channels through Ghost linear branches. Its function is to expand the effective receptive field of the monstrosity without significantly increasing the computational cost. Dir-SE: First, the directional statistics of each channel are obtained by averaging the vertical direction. Then, 1D convolution is performed in the horizontal direction and the channel weights are calculated by GAP_w, FC1 / ReLU, and FC2 / Sigmoid. The features are recalibrated with orientation awareness to suppress horizontal noise and highlight the directional structure. C2PSA: As the bottleneck unit of the backbone output, it integrates channel / spatial attention and cross-stage connectivity, integrates the aforementioned multi-scale semantics, and provides P3 / P4 / P5 three-scale features to the neck network (resolutions of approximately 80×80, 40×40, and 20×20, corresponding to strides of 8 / 16 / 32).

[0049] In this process, the features from Conv / C3k2 / SPPF / Ghost-DW / Dir-SE are reweighted in both channels and spatially and fused with cross-stage branches to form multi-scale semantics. In other words, C2PSA, as the bottleneck unit of the backbone output, reweights the upstream features in both channels and spatially and fuses them with cross-stage branches, outputting three feature maps P3, P4, and P5 (corresponding to strides of 8 / 16 / 32) for subsequent neck network multi-scale fusion.

[0050] (2) Neck network: Employs an FPN / PAN fused topology, consisting of uplink and downlink paths composed of Upsample, Cat, and C3k2 / Conv. Its functions are: High-semantic low-resolution features are upsampled using Upsample and concatenated with adjacent scales at Cat. C3k2 fusion convolution is used to achieve top-down semantic downpropagation and bottom-up detail uppropagation. Parallel Conv side branches are used to supplement local details and stabilize gradients. Finally, the F3 / F4 / F5 three-way fused features corresponding to P3 / P4 / P5 are output for the head network to decode.

[0051] (3) Head network: a multi-scale decoupled detection / segmentation head (Detect).

[0052] Decoupled structure: Each scale includes a classification branch (CV2: Conv→Conv→Conv2D→ClsLoss) and a regression branch (CV3: Conv→DWConv→Conv→DWConv→Conv2D→BoxLoss), decoupling class discrimination from boundary / geometric regression to improve convergence and accuracy; Output format: In the detection configuration, the class confidence and bounding box parameters are output, and scaled predictions such as Predict1 / Predict2 are generated. In the instance segmentation configuration, the prototype mask and mask coefficients (or equivalent mask branches) are output simultaneously. The two are linearly combined and thresholded to obtain the instance mask, which is used in the downstream "mask → central point → segmented fitting" link. It is worth explaining that the above three parts work together in the manner of backbone semantic extraction—neck multi-scale fusion—head decoupling decoding. Among them, Ghost-DW (7×1) and Dir-SE are only set between SPPF and C2PSA in the backbone to enhance the monstrous geometric representation and suppress lateral interference. The remaining sub-modules are connected in the order shown in the figure and keep the naming consistent.

[0053] Furthermore, such as Figure 4 As shown, the geometry processing unit performs statistically robust extraction and piecewise fitting of the central axis points within the mask space. First, the binary mask... Using structural elements Performing morphological opening operations to remove glitch and isolated points is defined as:

[0054] In the formula, This indicates a corrosion operation. Indicates an expansion operation; Then, a sliding window scan is performed line by line from the bottom of the image upwards at a predetermined step size, preferably 10 pixels, with a window width of 3% to 8% of the image width. Select the largest connected component pixel set within the window Its transverse centroid is taken as the principal midline point of the row, and the calculation formula is:

[0055] In the formula, Let x be the horizontal coordinate value of the pixel within the connected component. Let x be the horizontal centroid coordinates of the i-th row. Let be the number of pixels in the connected region of the i-th row.

[0056] To suppress the influence of occasional outliers on the fitting, the abscissa sequence was... Outliers were removed using the interquartile range method, retaining only... Samples within the interval are processed, and one-dimensional median or local mean smoothing is applied to the retained sequence to obtain a continuous, low-noise central point sequence.

[0057] In this application, fitting and continuity processing are performed on the central point sequence. The point sequence is divided longitudinally along the image. part, Take 4 to 6; the first The segment uses a first-order or second-order polynomial. Least squares fitting is performed to obtain the local curve. To avoid abrupt changes and error propagation at the joints between adjacent segments, [the following is done at the joint points]. Apply a continuity constraint: at least the function value must be continuous. Furthermore, to encourage first-order derivative continuity using soft constraints, the following weights are added to the objective function:

[0058] in, The value is the weighting coefficient for the smoothing term. In this application, it is set to 1.0. This is only an example and is not intended to be limiting. Let be the fitting function for the k-th segment. The splicing position of adjacent fitted segments, Let be the first derivative of the fitted function for the k-th segment at the splicing point. The constrained piecewise curves are then spliced ​​together to form the global navigation curve. .

[0059] Furthermore, in some implementations, this application also provides a method for controlling a crop-free field ridge vehicle, including: The global navigation curve was extracted using the cropless field ridge navigation line extraction method described above. Based on the global navigation curve, calculate its intersection point with the bottom edge of the image and the tangent direction at that intersection point to obtain the yaw angle. Calculate its distance from the midpoint of the bottom edge of the image to obtain the pixel error. A validity flag is generated based on the navigation curve, pixel error, and yaw angle, and vehicle control is performed when the validity flag is valid.

[0060] In one embodiment, the system publishes data at a fixed update frequency when running on an edge platform. Preferably, parameters such as input resolution, row step size, IQR coefficient, number of segments, and fitting order are loaded in the form of a configuration file to facilitate rapid tuning under different plots and lighting conditions. In this embodiment, the edge computing platform calculates the deviation of the vehicle relative to the field ridge navigation line in real time based on the visual detection results, and generates a validity flag accordingly to achieve differential steering correction.

[0061] In one embodiment, the edge computing platform calculates the deviation of the vehicle relative to the field ridge navigation line in real time based on the visual detection results, and generates an validity flag accordingly to achieve differential steering correction.

[0062] Wherein, the horizontal coordinate of the center pixel at the bottom edge of the vehicle Determined by the center pixel position of the bottom edge of the image, it serves as a reference position in the image coordinate system, representing the vehicle's forward direction. When there is a camera installation offset, the offset can be obtained through calibration. At this point, the x-coordinate of the vehicle's center pixel is defined as:

[0063] in, This represents the image width.

[0064] The equation of the navigation curve is denoted as Its vertical coordinate at the bottom edge The x-coordinate of the point is the intersection of the bottom edges of the navigation lines. The lateral deviation of a vehicle is defined as:

[0065] This deviation reflects the lateral offset of the vehicle relative to the center line of the field in the image coordinate system.

[0066] The vehicle's current heading angle is defined as the vehicle's direction of travel being consistent with the vertical direction of the image, i.e. At this point, the heading deviation angle The tilt angle of the navigation line relative to the vertical direction from the camera's perspective can be calculated from the slope of the tangent line at the bottom edge of the navigation curve:

[0067] When the tilt angle of the navigation line increases, the absolute value of the yaw angle increases accordingly, reflecting the degree to which the vehicle's attitude deviates from the direction of the ridge line.

[0068] To achieve vision-based automatic steering control, the system employs a dynamic fusion control method based on lateral deviation and heading deviation. The edge computing platform determines the steering direction based on the lateral deviation. , heading angle deviation and vehicle speed A nonlinear combination model is constructed to generate the comprehensive deviation output. The calculation relationship is as follows:

[0069] in, This is the lateral deviation gain coefficient. The vehicle's speed. To prevent small constants with a denominator of zero, this output quantity... It characterizes the overall deviation of the vehicle and is used to drive the actuators to make steering adjustments.

[0070] To prevent the system from frequently switching between left and right drive within a small deviation range, which could lead to vehicle heading instability, this embodiment sets a deviation tolerance threshold in the control logic. When the absolute value of the overall deviation output is less than this threshold, the condition is met:

[0071] When the system determines that the vehicle is in normal navigation mode ( No direction correction is needed. When When the tolerance range is exceeded, the control module writes coil on / off commands to the relay module via the Modbus RTU protocol to drive the electromagnet to engage / disengage, thereby pulling the left and right clutches to achieve steering. Specifically: when When the vehicle is determined to be veering to the right, a power-on command is written to the address of the left electromagnet coil corresponding to the relay module, while a power-off command is written to the address of the right electromagnet coil; when When the vehicle is determined to be veering to the left, a power-on command is written to the right coil address, and a power-off command is written to the left coil address; when At that time, power-off commands are written to the addresses of the left and right coils respectively, and the vehicle continues to travel straight.

[0072] In this application, the relay performs on / off control to drive the electromagnet to engage / disengage, thereby pulling the left and right clutches to achieve steering.

[0073] In one instance, such as Figure 6As shown, the electromagnets are push-pull type, with two electromagnets arranged on the left and right sides of the transplanter, respectively. They are connected to the clutch levers of the drive wheels on the left and right sides of the transplanter via traction steel wires. When it is necessary to control the transplanter to turn left or right, the electromagnet on that side is engaged, pulling back the electromagnet core and actuating the clutch lever on that side via the traction steel wire. This disengages the clutch on that side, cutting off power to the drive wheel on that side, thus achieving steering. The steering angle is controlled by the duration of the electromagnet engagement, which is also the duration of the clutch disengagement. The relationship between the steering angle and the electromagnet engagement time is determined through pre-calibration tests.

[0074] In summary, this application achieves the following beneficial effects: by introducing a large longitudinal kernel and orientation-aware structural enhancement into the instance segmentation network, it significantly reduces the jitter and missed detection of strip-shaped ridge boundaries; by combining sparsification and structured pruning with quantized deployment, it enables low-latency inference on low-computing-power, low-power edge platforms; by using a row-by-row sliding window centroid and a sequence of central points constructed through statistical elimination / smoothing, as well as a fitting strategy of piecewise least squares with continuity constraints, it effectively suppresses error diffusion and angle jumps caused by broken rows and irregular boundaries, resulting in stable and controllable output navigation quantities, facilitating direct interface with the controller and improving ridge walking performance and engineering feasibility under complex field conditions.

[0075] This application also provides a system for extracting field ridge navigation lines without crops, such as Figure 5 As shown, it includes: A camera is used to acquire raw images of the field area in the forward direction of operations; an image acquisition and processing module is used to determine the input frame based on the raw images. The instance segmentation module is used to perform pixel-level segmentation of the ridge region in the input frame to obtain an instance mask using an improved single-stage instance segmentation network model; The navigation line extraction module is used to perform morphological opening operations on the instance mask and extract the main central axis point sequence in a row-by-row sliding window manner. Outliers are removed by interquartile range method and the main central axis point sequence is smoothed in one dimension to obtain the central axis point sequence. It is also used to divide the central axis point sequence into several parts along the longitudinal direction of the original image, perform least squares fitting for each part of the central axis points, and apply continuity constraints between the central axis points of each part to form a global navigation curve.

[0076] The publishing module is used to calculate the intersection point of the global navigation curve with the bottom edge of the image and the tangent direction at the intersection point, thereby obtaining the yaw angle, and calculate the distance between the global navigation curve and the midpoint of the bottom edge of the image to obtain the pixel error; generate a validity flag based on the navigation curve, pixel error and yaw angle, and perform vehicle control when the validity flag is valid.

[0077] The image acquisition and preprocessing unit is integrated into the edge platform and calls its onboard camera; the instance segmentation and perception unit realizes network transformation, pruning and quantization inference from S2-1 to S2-4; the navigation line extraction unit realizes morphological preprocessing, row-by-row sliding window centroid, outlier removal and sequence smoothing; the fitting and navigation interface unit realizes piecewise fitting and continuity constraints, navigation quantity calculation and external release; the release unit is responsible for outputting the variables required by the controller.

[0078] In summary, the present invention has the following effects: Lightweight multi-scale feature extraction backbone network design: An improved Ghost module with vertical large kernel convolution is combined with a C2PSA module to reduce computational complexity and enhance orientation awareness. While ensuring feature representation ability, the network inference speed is significantly improved, thereby improving the real-time performance on the terminal side. Introducing the Direction-Aware Channel Recalibration (Dir-SE) module: By performing longitudinal statistics and lateral weighting on channel features, the network's response to the ridge strip structure is enhanced, effectively suppressing instability caused by interference features such as complex backgrounds, shadows, and stubble. Model compression strategy based on BN-γ sparsification and structured pruning: By sparse training and channel importance assessment, low-contribution channels are pruned and fine-tuned to restore them, reducing computational load and enabling lightweight model deployment, thus enabling the system to have real-time detection capabilities on edge platforms.

[0079] Through the above-mentioned technical improvements, the present invention significantly enhances the real-time performance and anti-interference stability of the extraction of navigation lines in fields without crops.

[0080] This application also provides a system for extracting navigation lines in ridges without crops, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method for extracting navigation lines in ridges without crops. This system can implement various embodiments of the method for extracting navigation lines without crops and achieve the same beneficial effects, which will not be elaborated upon here.

[0081] This application also provides a control system for a cropless field ridge vehicle, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned control method for the cropless field ridge vehicle. This control system for the cropless field ridge vehicle can implement various embodiments of the control method for the cropless field ridge vehicle and achieve the same beneficial effects, which will not be elaborated upon here.

[0082] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for extracting navigation lines in cropless fields, characterized in that, include: S1: Obtain the original image of the field ridge area in the forward direction, and determine the input frame based on the original image; S2: An improved single-stage instance segmentation network model is used to perform pixel-level segmentation of the ridge region in the input frame to obtain an instance mask; S3: Perform morphological opening operation on the instance mask and extract the main central axis point sequence in a row-by-row sliding window manner. Remove outliers by interquartile range method and perform one-dimensional smoothing on the main central axis point sequence to obtain the central axis point sequence. S4: Divide the central axis point sequence into several parts along the longitudinal direction of the original image, perform least squares fitting for each part of the central axis points, and apply continuity constraints between the central axis points of each part to form a global navigation curve.

2. The method for extracting navigation lines in cropless fields according to claim 1, characterized in that, The improved single-stage instance segmentation network model includes a backbone network, a neck network, and a head network connected in sequence. At the bottleneck positions of the backbone network and the neck network, there is a Ghost deep convolution with a large vertical kernel. A direction-aware channel recalibration module is provided before the head network.

3. The method for extracting navigation lines in cropless fields according to claim 2, characterized in that, S2 includes: The input frame is fed into the backbone network. After initial feature extraction through the input layer and two Conv layers, it enters multiple C3k2 layers for channel compression, nonlinear representation, and feature reuse. After further downsampling, it enters the SPPF layer. Between the SPPF layer and the C2PSA layer, a vertically large kernel Ghost depth-separable convolutional layer and a direction-aware channel recalibration module are inserted. The vertically large kernel Ghost depth-separable convolutional layer generates inexpensive channels through Ghost linear branches, expanding the effective receptive field of the monstrosity. The direction-aware channel recalibration module first calculates the 1×W direction statistics for each channel in the vertical direction, and then performs 1D convolution in the horizontal direction and calculates the channel weights to perform direction-aware channel recalibration on the features. The C2PSA layer integrates channel / spatial attention and cross-stage connections, integrates multi-scale semantics, and provides three-scale features to the neck network. The three-scale features are upsampled by the Upsample layer of the neck network to improve the high semantic low-resolution features, and then concatenated with the adjacent scales at the Cat level. The C3k2 layer is used for fusion convolution to achieve top-down semantic downpropagation and bottom-up detail uppropagation. The Conv side branch connected in parallel with the C3k2 layer is used to supplement local details and stabilize gradients. Finally, the three-way fused features corresponding to the three-scale features are output for the head network to decode. The head network outputs bounding boxes, class confidence, mask coefficients and prototype masks under the detection configuration based on the three-way fusion features. The mask coefficients and prototype masks are linearly combined and thresholded to obtain the field instance mask.

4. The method for extracting navigation lines in cropless fields according to claim 3, characterized in that, The Ghost processing flow is as follows: First, execute DWConv(7×1) which aggregates only along the vertical direction. The output of the c-th channel is: in, Output to the main branch. Learnable vertical convolutional kernel weights, no horizontal convolution, output resolution preserved. ; The radius of the convolution kernel is 1. , These represent the vertical and horizontal coordinates of the feature map. This represents the vertical sampling offset of the convolution kernel; The output of the c-th channel is activated sequentially using BN and SiLU: in, The batch mean and variance of this channel. Scaling and biasing for BN. The output characteristics after BN and SiLU activation are shown. It is a constant; Next, a Ghost linear transformation is performed to generate a cheap channel, which is then concatenated with the main channel. The output maintains the H×W size. The main channel is first obtained through linear mapping. , and then Applying a cheap linear transformation Generate additional channels, and finally concatenate them according to the channel dimensions to obtain the Ghost output; The processing flow of the orientation-aware channel recalibration module is as follows: First, we average the longitudinal direction of each channel to obtain 1×W directional statistics. The longitudinal average of feature X of the c-th channel is: In the formula, The height of the feature map, The width of the feature map. Indicates the first The pixel values ​​of each channel. For vertical pixel coordinate index, Horizontal pixel coordinate index; Performing one-dimensional convolution laterally: The system performs horizontal global averaging, calculates the scalar weight of each channel to obtain the channel weight, recalibrates the original features by channels, reduces the channel dimension, restores the channel dimension, and finally normalizes it using Sigmoid.

5. The method for extracting navigation lines in cropless fields according to claim 4, characterized in that, The horizontal global averaging process satisfies the following relationship: Let the input features be ,in For the number of channels, For feature map height, The width of the feature map; the first The passage is in the location The pixel value is denoted as , , ; for the first The channel is obtained by vertical averaging and horizontal convolution: Again Dimensions are used for horizontal global averaging After processing, we obtain: In the formula, It is a horizontal global average. As a dimension, For the first The transverse response sequence of the channel; This is the channel statistics vector obtained by horizontal global averaging; Number of channels; It is the symbol for the set of real numbers. Representing vectors The dimension is ; The calculation of the weight scalar for each channel satisfies the following: Let the compression ratio be ,but: The channel weights obtained after Sigmoid normalization are as follows: In the formula, This is the compressed intermediate channel vector; The weights and biases of the two fully connected layers; For unnormalized channel scores; The original features are recalibrated via channels, satisfying the following relationship: In the formula, For the first The passage is The recalibrated pixel value at the location; For the first Channel weighting coefficients; For the first The passage is The original pixel value at that location.

6. The method for extracting navigation lines in cropless fields according to claim 1, characterized in that, S3 includes: For binary masks Using structural elements Perform morphological opening operations to remove burrs and isolated points, satisfying the following relation: In the formula, This indicates a corrosion operation. Indicates an expansion operation; A sliding window scan is performed line by line from the bottom of the original image upwards at predetermined steps, in the lines... Select the largest connected component pixel set within the window Its transverse centroid is taken as the main midline point of the row, and the calculation formula is as follows: ； In the formula, Let x be the horizontal coordinate value of the pixel within the connected component. Let i be the horizontal centroid coordinates corresponding to the i-th row. This represents the number of pixels within the connected component in the i-th row. For the x-axis sequence The Q1-Q3 quartile method was used to remove outliers by a 1.5×IQR threshold, and the retained sequence was smoothed by median or local mean to obtain a continuous, low-noise central point sequence.

7. The method for extracting navigation lines in cropless fields according to claim 1, characterized in that, S4 includes: Divide the central point sequence along the vertical axis of the image into Paragraph; in the The segment uses a first-order or second-order polynomial. Least squares fitting is performed to obtain the local curve; at the splicing point of adjacent segments Apply a continuity constraint: at least ensure that the function values ​​are continuous. Furthermore, to encourage first-order derivative continuity through soft constraints, the following weights are added to the objective function: in, These are the weighting coefficients for the smoothing term. Let be the fitting function for the k-th segment. This represents the splicing position between adjacent fitted segments. Let be the first derivative of the fitted function for the k-th segment at the splice point; The constrained piecewise curves are then stitched together to form a global navigation curve. .

8. A control method for a vehicle operating in a field without crop ridges, characterized in that, include: The global navigation curve is extracted using the method for extracting navigation lines in uncropped fields as described in any one of claims 1-7; Based on the global navigation curve, calculate its intersection point with the bottom edge of the image and the direction of the tangent at that intersection point to obtain the yaw angle. Calculate its distance from the midpoint of the bottom edge of the image to obtain the pixel error. A validity flag is generated based on the navigation curve, pixel error, and yaw angle, and vehicle control is performed when the validity flag is valid.

9. A cropless field ridge navigation line extraction system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of any of the methods described in claims 1-7.

10. A control system for a crop-free field ridge vehicle, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method described in claim 8.

Images