Method for extracting and tracking under-crown navigation line based on semantic segmentation and feature point clustering

Abstract

The invention discloses a method for extracting and tracking a navigation line under a crown based on semantic segmentation and feature point clustering. The method comprises the following steps: S1, obtaining a color image between field ridges; s2, generating a navigation area mask graph through a semi-supervised semantic segmentation model UniMatch V2; s3, extracting a strip center point and fitting a navigation line through an RANSAC (Random Sample Consensus) algorithm; s4, defining a visual navigation working area to guide the power platform to advance; and S5, processing the current frame of mask image by combining with moving average filtering to realize continuous and stable tracking of the navigation line. According to the invention, through the semi-supervised semantic segmentation model UniMatch V2, under-crown inter-ridge navigation areas can be accurately identified in real time under limited manual annotation data volume, and navigation lines are extracted and tracked, so that stable and autonomous operation of the power platform under field crowns is ensured, mechanical damage is reduced, operation efficiency is improved, and the method is suitable for large-scale popularization and application. The device is especially suitable for field management tasks such as crop harvesting and transferring.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and image processing technology, and in particular to a method for extracting and tracking under-corona navigation lines based on semantic segmentation and feature point clustering. Background Technology

[0002] With the development of smart agriculture technology, crop harvesting is gradually moving towards mechanization and intelligence. Traditional manual harvesting methods, due to their high labor intensity and low efficiency, can no longer meet the needs of modern agriculture. To improve harvesting efficiency and operational quality, the application of automated navigation technology has become increasingly important. In crop harvesting operations, precise navigation is key to achieving automated operation. However, due to the complex environment under the canopy and between rows, and severe shading, traditional navigation area identification and navigation line extraction still face technical challenges, affecting the operational effectiveness and accuracy of the power platform.

[0003] Currently, traditional navigation technologies mainly rely on GNSS for positioning and navigation. However, during the crop harvest season, GNSS signals are frequently interfered with or fail due to severe leaf shading, and the related equipment is expensive, making it difficult to meet the requirements of high efficiency and low cost. At the same time, the terrain under the canopy and between rows is complex and dynamically changing. Traditional visual navigation methods are easily affected by extended leaves, changing lighting conditions, and ground weeds when identifying navigation areas, resulting in low accuracy in navigation line extraction and failing to meet the precision requirements of automated operations.

[0004] Therefore, there is an urgent need for an innovative navigation technology that can accurately identify navigation areas and effectively extract navigation lines in complex and ever-changing canopy harvesting environments, thereby automating crop harvesting operations. The implementation of this technology can not only significantly improve operational efficiency but also significantly reduce mechanical damage to crops, making it an important component of realizing smart agriculture. Summary of the Invention

[0005] One objective of this invention is to propose a canopy navigation line extraction and tracking method based on semantic segmentation and feature point clustering. This invention can effectively address environmental noise interference such as leaf shading and ground weeds, significantly improving the detection accuracy and tracking stability of the navigation lines. By accurately extracting the inter-row navigation lines, this invention can significantly improve the navigation performance of the power platform in complex canopy environments, thereby increasing the efficiency and accuracy of automated operations, reducing mechanical damage to crops, and promoting the development of crop harvesting operations towards a more intelligent and efficient direction.

[0006] The crown navigation line extraction and tracking method based on semantic segmentation and feature point clustering according to embodiments of the present invention includes the following steps:

[0007] S1: Acquire color images of the canopy between rows using an image acquisition device installed at the front of the power platform;

[0008] S2: Input the color image into the semi-supervised semantic segmentation model UniMatch V2, and output the mask map of the navigation region between ridges and under the crown;

[0009] S3: Perform horizontal strip segmentation and feature point clustering on the mask image, extract the central feature points corresponding to each strip, and use the random sampling consensus algorithm to fit the central feature points to obtain the navigation lines;

[0010] S4: Based on the navigation lines obtained from the fitting, determine the visual navigation working area of ​​the power platform;

[0011] S5: Based on the visual navigation working area corresponding to the previous frame image, perform region filtering processing on the mask image of the current frame image, and combine the filtering algorithm to smooth and optimize the navigation line fitted in the current frame, outputting a stable navigation line to realize continuous tracking of the navigation line between the canopy and the ridge in the field.

[0012] Furthermore, in step S1, the image acquisition device is an RGB camera, which is installed at the center of the front end of the power platform. The camera is 1.05m above the ground, and the angle between the camera's optical axis and the horizontal plane is 15°. The power platform travels at a constant speed of 0.5m / s. The RGB camera acquires SVO format video at 1080p resolution and 30 frames / second. The acquired color image covers the working area of ​​the power platform, and the acquisition time is during the crop harvest period.

[0013] Furthermore, in step S2, the improved features of the semi-supervised semantic segmentation model UniMatch V2 include:

[0014] A student-teacher network architecture is adopted, which combines a consistency regularization method from weak to strong enhancement and a pseudo-label-guided learning method to improve segmentation accuracy by utilizing the latent semantic information in unlabeled images.

[0015] By fusing image-level enhancement and feature-level enhancement into a unified learning stream, the redundant overhead of multi-stream training structures is reduced, computational resource consumption is decreased, and training efficiency is improved.

[0016] A complementary channel Dropout mechanism is introduced, which divides the feature map output by the encoder into two complementary subsets in the channel dimension. The two subsets are then processed by Dropout to obtain two feature perspectives that are semantically differentiated and informationally complementary, thereby enhancing the generalization ability of the model.

[0017] Furthermore, in step S3, the horizontal strip segmentation process of the mask image includes: first removing approximately 1 / 3 of the upper part of the mask image, and then dividing the remaining area into N equal-height strips along the vertical direction. The size of each equal-height strip is W×h, where W is the width of the mask image, h is the height of the strip, and N=10.

[0018] Furthermore, in step S3, the core of the feature point clustering is edge point extraction based on contour structure, and the specific process includes:

[0019] S31. First, extract the main contour of the segmented mask region, and generate a continuous strip region through contour filling operation to suppress boundary discontinuity problems caused by boundary jaggedness, contour splitting or small area false detection.

[0020] S32. Then, in the middle feature row of each band, the median value strategy is used to perform statistical analysis on several pixels on the leftmost and rightmost sides, replacing the point selection method based on extreme values, to determine the position of the edge points.

[0021] Furthermore, in step S3, the method for extracting the central feature points is as follows:

[0022] The arithmetic mean of the x-coordinates of the left and right edge points corresponding to each band is used to obtain the center feature point of the band.

[0023] Traverse all strips to generate a set of midpoint feature points, which serves as the base feature point set for fitting the navigation line.

[0024] Furthermore, in step S3, when fitting the midpoint feature point set using the linear regression method of the random sampling consensus algorithm, noise interference is suppressed by removing outliers that deviate from the main structure of the point set, while preserving the overall distribution trend of the midpoint feature point set, thus generating a smooth and robust navigation line.

[0025] Furthermore, in step S4, the visual navigation working area is the region of interest determined based on the outline of the field ridges on both sides of the driving area of ​​the power platform and the area between the ridges.

[0026] Furthermore, in step S5, the specific method for performing region filtering processing on the mask image of the current frame image is as follows:

[0027] Retain the pixels in the current frame mask that overlap with the visual navigation working area of ​​the previous frame, and set the pixels in the remaining areas of the current frame mask to zero to reduce computation and improve tracking accuracy.

[0028] Furthermore, in step S5, the filtering algorithm is an adaptive moving average filtering algorithm. The navigation line fitted in the current frame is smoothed by the adaptive moving average filtering algorithm to improve the stability and positioning accuracy of the navigation line.

[0029] The beneficial effects of this invention are:

[0030] 1. This invention proposes a canopy navigation line extraction and tracking method based on semantic segmentation and feature point clustering. This method integrates semi-supervised semantic segmentation and dynamic tracking techniques, enabling high-precision detection of navigation areas and real-time stable tracking of navigation lines in complex field environments during crop harvesting. Through semi-supervised semantic segmentation, this invention achieves pixel-level accurate segmentation of inter-row navigation areas even with limited manually labeled datasets, effectively reducing reliance on large-scale labeled samples and overcoming difficulties in navigation area identification caused by leaf occlusion. Even under complex interference conditions such as varying light levels and ground weeds, the robustness and accuracy of navigation area identification are still guaranteed.

[0031] 2. The canopy navigation line extraction and tracking method proposed in this invention, based on semantic segmentation and feature point clustering, combined with a dynamic tracking strategy, can accurately extract the canopy navigation lines between rows in the field and track their position and morphological changes in real time. Through continuous image frame processing and dynamic updates, the navigation route can be adaptively adjusted to ensure that the power platform travels precisely along the row navigation lines. Compared with traditional methods, this invention not only improves the accuracy of navigation line extraction but also significantly enhances the system's adaptability in complex field environments, ensuring the overall stability of operation.

[0032] 3. The canopy navigation line extraction and tracking method proposed in this invention, based on semantic segmentation and feature point clustering, can be directly applied to existing crop harvesters. It uses a semi-supervised semantic segmentation model to identify the inter-row navigation area in real time, and combines feature point extraction and tracking algorithms to dynamically output the navigation line. Based on the navigation results of this invention, the power platform can operate smoothly along the navigation line, avoiding mechanical damage to the crop, thereby effectively improving operational efficiency and quality. Attached Figure Description

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0034] Figure 1 This is a flowchart of the method for extracting and tracking the subcoronal navigation line based on semantic segmentation and strip feature point clustering according to the present invention.

[0035] Figure 2 These are color images of the field under the canopy and between rows, acquired using an RGB camera.

[0036] Figure 3 This is a model structure diagram of the semi-supervised semantic segmentation model UniMatch V2.

[0037] Figure 4Diagram of the unified enhancement and complementary channel structure for UniMatch V2.

[0038] Figure 5 This is a diagram illustrating the process of strip segmentation and feature point extraction.

[0039] Figure 6 Color images of the field under canopy and between rows under different lighting and shading conditions, along with the corresponding processing steps and their effects. Detailed Implementation

[0040] To make the technical means and objectives and effects of the present invention easier to understand, the technical solutions of the present invention will be described in detail below with reference to the embodiments and accompanying drawings, so as to ensure that those skilled in the art can accurately understand and reproduce the technical solutions of the present invention.

[0041] In this embodiment, the power platform is a wheeled field operation platform with a wheel track of 60cm, which is suitable for the common 40cm-80cm ridge width in the field. The computing unit on the operation platform is an NVIDIA Jetson AGX Xavier, which has strong parallel computing capabilities and can meet the real-time inference and navigation line processing requirements of the semi-supervised semantic segmentation model.

[0042] Appendix Figure 1 To be continued Figure 6 The accompanying diagrams correspond to the following: a flowchart of the method of this invention; an example of color image acquisition under the canopy and between rows in the field; a structural diagram of the semi-supervised semantic segmentation model UniMatch V2; a structural diagram of the unified enhancement and complementary channels; a diagram of the strip segmentation and feature point extraction process; and diagrams showing the effects of each processing step under different lighting and occlusion conditions. Each diagram can serve as an intuitive reference for the technical solution of this embodiment.

[0043] like Figure 1-6 As shown, this invention discloses a method for extracting and tracking under-corona navigation lines based on semantic segmentation and feature point clustering, including the following steps:

[0044] S1: Acquire color images of the canopy between rows using an image acquisition device installed at the front of the power platform;

[0045] like Figure 1 As shown, specifically, during the acquisition of color images under the canopy between rows, an RGB camera was selected as the image acquisition device. This camera features high frame rate and high resolution, and can stably output 1080p resolution color images. The camera is mounted on the front center of the power platform via a custom-made metal bracket. The bracket height is finely adjusted and fixed at 1.05m above the ground. This height is determined based on the crop plant height, which avoids the camera being directly blocked by leaves and ensures that the image covers the 2-3m area between rows in front of the platform, leaving sufficient reaction distance for navigation.

[0046] The angle between the camera's optical axis and the horizontal plane is set to 15°. This angle was determined through multiple field tests. It can balance the details between the ridges in the foreground and the extension trend between the ridges in the background in the image, and avoid the background being completely blocked by the upper leaves due to the optical axis being too flat, or the foreground ridge area being too small due to the optical axis being too steep.

[0047] The power platform travels at a constant speed of 0.5 m / s, which meets the actual needs of crop harvesting operations. Too fast a speed may cause the harvesting mechanism to miss or damage the crops, while too slow a speed will reduce the efficiency of the operation.

[0048] The RGB camera captures SVO format video at a frame rate of 30 frames per second. This frame rate matches the driving speed of the power platform. Each frame corresponds to a platform movement distance of approximately 1.67 cm. The overlap between adjacent frames exceeds 90%, providing coherent image information for subsequent continuous tracking.

[0049] Among them, the SVO format has a compression ratio of about 5:1, which can reduce the storage unit occupation of the power platform, while preserving image details and avoiding blurring of the edges between rows due to compression.

[0050] The image acquisition time was limited to the crop harvest period, when the field ridge structure was stable, the crop was mature and not withered in large areas, and the contrast between the inter-ridge area and the leaf area was high, which facilitated subsequent semantic segmentation.

[0051] The data collection scenarios covered various natural conditions, including strong light, weak light, weed cover, and leaf extension shading, to verify the adaptability of the invention in complex scenarios. The collected color images of the field under the canopy and between rows are shown below. Figure 2 As shown.

[0052] S2: Input the color image into the semi-supervised semantic segmentation model UniMatchV2 and output a mask image of the field ridge navigation area;

[0053] like Figure 3 and Figure 4 As shown, specifically, the training of the semi-supervised semantic segmentation model UniMatch V2 is based on the PyTorch deep learning framework. The training dataset contains 2000 color images of field rows, of which only 400 are manually labeled and 1600 are unlabeled. The labeling tool used is Labelme, with labeling accuracy down to the pixel level.

[0054] The core improvements and implementation details of this model are as follows:

[0055] First, the model adopts a student and teacher network architecture, and the encoders are all based on Vision Transformer DINOv2-B / 16. Compared with the traditional ResNet encoder, DINOv2 is pre-trained on large-scale unsupervised image corpora, and has stronger semantic feature representation ability and cross-domain transfer ability, improving the segmentation IoU by about 8% in complex field scenes;

[0056] The teacher network is updated using an exponential moving average, with an EMA decay coefficient set to 0.999. This balances the stability of the teacher network with its adaptability to new data, avoiding slow updates due to an excessively high coefficient or excessive fluctuations due to an excessively low coefficient.

[0057] Secondly, the model employs consistency regularization from weak to strong enhancement and pseudo-label-guided learning: the weak enhancement method is random horizontal flipping and brightness fine-tuning, and the strong enhancement method is random cropping, color jittering and Gaussian blur.

[0058] The pseudo-labels are generated based on the prediction results of the teacher network on the weakly enhanced image. Only the predicted pixels with a confidence score higher than 0.9 are retained as pseudo-labels. This threshold is determined by cross-validation. A score below 0.9 will increase the noise of the pseudo-labels, while a score above 0.9 will reduce the effective supervision information.

[0059] Third, the model integrates image-level enhancement and feature-level enhancement into a unified learning stream. Compared with the traditional multi-stream training structure of UniMatchV1, the training time is reduced by about 40%, the GPU memory usage is reduced by about 30%, and the segmentation accuracy is reduced by only about 1%, achieving a balance between efficiency and accuracy.

[0060] Fourth, the model introduces a complementary channel Dropout mechanism and employs a unified multi-level perturbation consistency mechanism during training to further improve the model's semi-supervised performance. The specific steps are as follows:

[0061] First, examine the original unlabeled image. Image obtained by applying weak enhancement Pseudo-labels are generated using an exponential moving average teacher model. First, it serves as a supervision signal in subsequent training; second, two strongly enhanced views are obtained by independently sampling from the same original image. and The feature maps are then extracted by inputting them into the shared encoder g. and Then, complementary channel dropout is applied to each pair of feature maps in the feature space, which is achieved by randomly sampling binary complementary masks along the channel dimension. Two non-overlapping feature subsets are generated respectively, as shown in equations (1) and (2).

[0062]

[0063]

[0064] In the formula, ⊙ represents element-wise multiplication, M is a binary mask in the channel dimension, and 1-M is a binary mask complementary to M.

[0065] Subsequently, these two perturbation features are fed into the shared decoder h to generate the model's predictions from different perspectives. and ;

[0066] Finally, the confidence level exceeds the threshold. In this embodiment =0.9, the corresponding pseudo-label is the supervision signal, and consistency constraint training is performed on the above prediction results. This mechanism can improve the segmentation IoU of the model by about 5% in low light scene and about 6% in weed-covered scene.

[0067] Fifth, the model's inference speed on the NVIDIA Jetson AGX Xavier can reach approximately 35 frames per second, which is higher than the camera's acquisition frame rate of 30 frames per second, meeting the requirements for real-time processing;

[0068] The output field ridge navigation area mask is a single-channel binary image with a foreground pixel value of 255 and a background pixel value of 0. The resolution is consistent with the input color image to ensure that the pixel positions correspond in subsequent processing.

[0069] S3: Perform horizontal strip segmentation and feature point clustering on the mask image, extract the central feature points corresponding to each strip, and use the random sampling consensus algorithm to fit the central feature points to obtain the navigation lines;

[0070] like Figure 5 As shown, specifically, during the navigation line fitting process, the mask image is first preprocessed: the upper 1 / 3 of the image is removed. This area suffers from severe structural compression due to the perspective projection principle and is prone to overlapping with the upper blade projection, leading to misjudgment of edge points. After removal, the edge point misjudgment rate is reduced from about 15% to about 5%.

[0071] The remaining 2 / 3 of the region is then divided into N=10 equal-height strips along the vertical direction. Each strip has a size of W×h, where W is the image width of 1920 pixels and h is the strip height. For example, in a 1080p image, the height of the remaining region is 720 pixels and h=72 pixels.

[0072] Choosing N=10 is the optimal value determined through comparative testing. When N=5, the strip height is too large, resulting in insufficient feature point density and local offset of the navigation line reaching 3-5cm. When N=15, the strip height is too small, resulting in a 30% increase in computational cost but less than a 5% improvement in accuracy. N=10 can balance feature point density and computational cost.

[0073] The core of feature point clustering is edge point extraction based on contour structure: the findContours function of OpenCV library is used to extract the main contour of the mask region, and the fillPoly function is used to fill the contour to generate continuous strip regions, suppressing boundary discontinuity problems caused by jagged edges, contour splits or small-area false detections. After filling, the continuity of the strip region boundary is improved by about 90%.

[0074] In the middle feature row of each band, the median value strategy is used to determine the edge points: the horizontal coordinates of the leftmost 10 consecutive non-zero pixels of the feature row are counted and the median value is taken as the horizontal coordinate of the left edge point. Similarly, the horizontal coordinates of the rightmost 10 consecutive non-zero pixels are counted and the median value is taken as the horizontal coordinate of the right edge point. Compared with the traditional extreme value selection method, this strategy can control the edge point offset caused by isolated noise from 5-8 pixels to within 1-2 pixels, which significantly improves stability.

[0075] The method for extracting the center feature points is as follows: calculate the arithmetic mean of the x-coordinates of the left and right edge points of each band, and the x-coordinate of the left edge point... x-coordinate of the right edge point x-coordinate of the central feature point The vertical axis is the vertical axis of the feature row, and the midpoint feature point set is generated by traversing 10 strips.

[0076] Finally, the random sampling consensus algorithm was used to perform linear fitting on the midpoint feature point set: the number of algorithm iterations was set to 100, the distance threshold was set to 2 pixels, and the confidence level was set to 0.95.

[0077] During iteration, two feature points are randomly selected each time to fit an initial straight line. The distances from the remaining eight points to the line are calculated, and the number of interior points with a distance less than 2 pixels is counted. The straight line with the most interior points is selected as the optimal fitting navigation line. This algorithm can effectively remove outliers and can still accurately fit the line when the proportion of outliers is ≤30%. The navigation line uses a straight line equation. Let x be the x-coordinate, y be the y-coordinate, k be the slope, and b be the intercept.

[0078] S4: Based on the navigation lines obtained from the fitting, determine the visual navigation working area of ​​the power platform;

[0079] Specifically, in the process of determining the working area for visual navigation, the ROI is determined based on the navigation line fitted in step S3 and the wheelbase of the power platform.

[0080] First, the camera internal parameter matrix and external parameter matrix are obtained through the Zhang-Zhengyou calibration method. The straight-line equation of the navigation line in the image coordinate system is converted into the world coordinate system to determine the position of the actual navigation line.

[0081] Taking the actual navigation line as the center, a rectangular area with a width of 1.0 m is formed by expanding 0.5 m on both sides. This area just covers the wheel track range of the power platform, ensuring that the platform travels along the navigation line without deviating from the inter-row area.

[0082] Then, the actual rectangular area is converted back to the image coordinate system to obtain the ROI area coordinates. , where is the starting vertical coordinate after removing the upper 1 / 3 area in step S3, is the vertical coordinate at the bottom of the image, = the horizontal coordinate of the navigation line at - 150, = the horizontal coordinate of the navigation line at y_min + 150.

[0083] This ROI area can reduce the number of pixels to be processed in the subsequent process from 2,073,600 pixels in the original mask image to about 345,600 pixels, reducing the computational amount by about 83%, and at the same time avoiding interference from irrelevant information outside the ROI to subsequent tracking.

[0084] S5: According to the visual navigation working area corresponding to the previous frame image, the mask image of the current frame image is subjected to region screening processing, and the navigation line obtained by fitting the current frame is smoothed and optimized in combination with the filtering algorithm, and a stable navigation line is output to achieve continuous tracking of the navigation line between the rows under the canopy in the field.

[0085] As Figure 6 shown, specifically, during the continuous tracking of the navigation line, the mask image of the current frame is processed first:

[0086] Read the ROI area coordinates determined in the previous frame:

[0087] ;

[0088] All pixels in the current frame mask image whose coordinates satisfy x < x_prev_min, x > x_prev_max, y < y_prev_min or y > y_prev_max are set to 0, and only the pixel values within the ROI area are retained.

[0089] This processing can use the navigation area information of the previous frame to constrain the processing range of the current frame. In a scenario where the local occlusion area reaches 40%, when not processed, the offset of the navigation line in the current frame reaches 8 - 10 cm, and after processing, the offset can be controlled within 2 - 3 cm, significantly improving the tracking accuracy.

[0090] Subsequently, an adaptive moving average filtering algorithm is used to smooth the navigation line fitted in the current frame: the size of the filtering window is dynamically adjusted according to the rate of change of the slope of the navigation line, and the rate of change of the slope is calculated as the absolute value of the difference between the current frame slope k_curr and the previous frame slope k_prev, |k_curr-k_prev|.

[0091] When |k_curr-k_prev|<0.05, the rows are straight and the navigation lines change gently. The window size is set to 5 frames. The slope and intercept of the current output navigation line are the average values ​​of the corresponding parameters in the last 5 frames, which enhances stability and avoids small fluctuations caused by single-frame noise.

[0092] When |k_curr-k_prev|≥0.05, the inter-row curvature and navigation line changes are obvious. The window size is set to 3 frames to ensure that the navigation line responds to the inter-row curvature trend in a timely manner and avoids navigation lag caused by an excessively large window. Compared with the moving average filtering with a fixed 5-frame window, this algorithm reduces the fluctuation range of the navigation line between straight rows from ±3cm to ±1.5cm, and the navigation lag distance between curved rows from about 10cm to about 5cm, balancing stability and response speed.

[0093] The final output of stable navigation line parameters (k, b) is transmitted to the motion control unit of the power platform in real time. The control unit adjusts the steering angle of the power platform according to the parameters to ensure that the power platform travels smoothly along the inter-row navigation line and the tracking error is controlled within ±3cm.

[0094] Field tests showed that using the navigation method of this invention reduced crop damage rate from about 8% with traditional navigation methods to about 2%, and improved operational efficiency by about 15%.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting and tracking crown navigation lines based on semantic segmentation and feature point clustering, characterized in that, Includes the following steps: S1: Acquire color images of the canopy between rows using an image acquisition device installed at the front of the power platform; S2: Input the color image into the semi-supervised semantic segmentation model UniMatch V2, and output the mask map of the navigation region between ridges and under the crown; S3: Perform horizontal strip segmentation and feature point clustering on the mask image, extract the central feature points corresponding to each strip, and use the random sampling consensus algorithm to fit the central feature points to obtain the navigation lines; S4: Based on the navigation lines obtained from the fitting, determine the visual navigation working area of ​​the power platform; S5: Based on the visual navigation working area corresponding to the previous frame image, perform region filtering processing on the mask image of the current frame image, and combine the filtering algorithm to smooth and optimize the navigation line fitted in the current frame, outputting a stable navigation line to realize continuous tracking of the navigation line between the canopy and the ridge in the field.

2. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S1, the image acquisition device is an RGB camera, which is installed at the center of the front end of the power platform. The camera is 1m-1.1m above the ground, and the angle between the camera's optical axis and the horizontal plane is 15°. The power platform travels at a constant speed of 0.4m / s-0.6m / s. The RGB camera acquires SVO format video at 1080p resolution and 30 frames / second. The acquired color image covers the working area of ​​the power platform, and the acquisition time is during the crop harvest period.

3. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S2, the improved features of the semi-supervised semantic segmentation model UniMatch V2 include: A student-teacher network architecture is adopted, which combines a consistency regularization method from weak to strong enhancement and a pseudo-label-guided learning method to improve segmentation accuracy by utilizing the latent semantic information in unlabeled images. By fusing image-level enhancement and feature-level enhancement into a unified learning stream, the redundant overhead of multi-stream training structures is reduced, computational resource consumption is decreased, and training efficiency is improved. A complementary channel Dropout mechanism is introduced, which divides the feature map output by the encoder into two complementary subsets in the channel dimension. The two subsets are then processed by Dropout to obtain two feature perspectives that are semantically differentiated and informationally complementary, thereby enhancing the generalization ability of the model.

4. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S3, the horizontal strip segmentation process of the mask image includes: first removing approximately 1 / 3 of the upper part of the mask image, and then dividing the remaining area into N equal-height strips along the vertical direction. The size of each equal-height strip is W×h, where W is the width of the mask image and h is the height of the strip.

5. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S3, the core of the feature point clustering is edge point extraction based on contour structure, and the specific process includes: S31. First, extract the main contour of the segmented mask region, and generate a continuous strip region through contour filling operation to suppress boundary discontinuity problems caused by boundary jaggedness, contour splitting or small area false detection. S32. Then, in the middle feature row of each band, the median value strategy is used to perform statistical analysis on several pixels on the leftmost and rightmost sides, replacing the point selection method based on extreme values, to determine the position of the edge points.

6. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S3, the method for extracting the center feature points is as follows: The arithmetic mean of the x-coordinates of the left and right edge points corresponding to each band is used to obtain the center feature point of the band. Traverse all strips to generate a set of midpoint feature points, which serves as the base feature point set for fitting the navigation line.

7. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S3, when fitting the midpoint feature point set using linear regression with the random sampling consensus algorithm, noise interference is suppressed by removing outliers that deviate from the main structure of the point set, while preserving the overall distribution trend of the midpoint feature point set, thus generating a smooth and robust navigation line.

8. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S4, the visual navigation working area is the region of interest determined by the outline of the field ridges and the area between the ridges on both sides of the driving area of ​​the power platform.

9. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S5, the specific method for performing region filtering on the mask image of the current frame is as follows: Retain the pixels in the current frame mask that overlap with the visual navigation working area of ​​the previous frame, and set the pixels in the remaining areas of the current frame mask to zero, in order to reduce the amount of computation and improve tracking accuracy.

10. The method for extracting and tracking undercoronal navigation lines based on semantic segmentation and feature point clustering according to claim 1, characterized in that, In step S5, the filtering algorithm is an adaptive moving average filtering algorithm. The adaptive moving average filtering algorithm is used to smooth the navigation line fitted in the current frame, thereby improving the stability and positioning accuracy of the navigation line.

Images