Dynamic path planning method and system for orchard weeding robot based on visual perception

Abstract

The application provides a kind of orchard weeding robot dynamic path planning method and system based on visual perception, comprising: constructing improved YOLO v8 network structure, OAK-D-PRO camera is carried on weeding robot to collect two side fruit tree image data and is passed to navigation control module, navigation control module utilizes target detection model to identify orchard inter-row trunk and obstacle, the midpoint of the lower edge of the output trunk recognition frame is used as the fitting reference point of tree row, the least square method is used to fit unilateral tree row line and tree row center navigation line, generate target navigation path, the execution mechanism controls weeding robot walking according to target navigation path, in the process of walking, navigation control module dynamically adjusts the moving direction and steering amplitude of weeding robot.The application effectively solves the problem that the navigation path in the prior art is greatly affected by the environment, and realizes efficient and accurate orchard robot path planning.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent agriculture and robot technology, and particularly relates to a dynamic path planning method and system for orchard weeding robots based on visual perception. BACKGROUND

[0002] Under the background of the rapid development of intelligent agriculture, the automatic navigation technology of orchard weeding robots has become the key to improving the management efficiency of orchards. However, the existing orchard weeding robots face significant challenges in complex environments: the target recognition model relied on by traditional navigation path planning mainly relies on the combination of hand-designed feature extractors and classifiers in orchard navigation path planning, extracts visual features with distinguishing degrees in images, and then uses machine learning methods for classification or positioning. When identifying tree trunks in orchards using traditional methods, image preprocessing is generally performed first, then feature extraction is performed, HOG descriptor is used to extract tree trunk edge direction features, then a classifier is trained, SVM is used to train positive samples (tree trunks) and negative samples (non-tree trunks), then the detection stage is entered, different size windows are slid on the image, HOG features are extracted for each window and SVM is used to judge whether it is a tree trunk, and finally non-maximum suppression (NMS) is performed to merge overlapping detection boxes to obtain the final tree trunk position.

[0003] Traditional methods have some limitations, they rely on hand-designed features, have limited generalization ability, are not suitable for complex environments such as light changes and occlusions, have insufficient recognition accuracy for different shapes and sizes of tree trunks and obstacles in orchards, and are easily disturbed by environmental factors such as leaf occlusion and light changes; at the same time, tree row positioning often uses the midpoint of the tree root, which is easily blocked by soil and weeds, as the base point, and when fitting the tree row line with the conventional least squares method, it is easily affected by noise data, resulting in a large deviation between the actual navigation path and the preset path, which seriously restricts the work efficiency and automation level of the robot.

[0004] Therefore, there is an urgent need for a path planning technology that can adapt to the complex environment of orchards, improve target recognition accuracy, and improve the robustness of path fitting. SUMMARY

[0005] In view of the deficiencies in the prior art, the present application provides a dynamic path planning method and system for orchard weeding robots based on visual perception, which effectively solves the problems of insufficient target recognition accuracy in traditional path planning methods, the recognition of tree trunks and obstacles by the model being easily disturbed by leaf occlusion and light changes, the fitting base point being easily blocked by soil or weeds, and the navigation path deviating from the preset trajectory due to the influence of noise data on the conventional least squares method.

[0006] The present application achieves the above technical purpose through the following technical means.

[0007] A visual perception-based orchard weeding robot dynamic path planning method, comprising the following processes:

[0008] Step 1: Collect photos under different weather conditions in different seasons to obtain an orchard road data set, and divide it into a training set, a validation set, and a test set according to the corresponding proportions;

[0009] Step 2: Construct an improved YOLO v8 network structure to obtain an improved YOLO v8 model: replace some C2f modules in the backbone network with ConvNeXt V2 modules, and add a multi-scale expansion attention mechanism, i.e., the MSDA mechanism, after the SPPF layer in the backbone;

[0010] Step 3: Train, validate, and test the improved YOLO v8 model using the data set established in step 1; load the best weight file obtained after training into the improved YOLO v8 model to obtain a target detection model, and load the target detection model into a navigation control module;

[0011] Step 4: The weeding robot carrying an OAK-D-PRO camera advances in the orchard, the OAK-D-PRO camera collects real-time image data of the trees on both sides of the orchard and transmits it to the navigation control module; the navigation control module uses the target detection model to identify the tree trunks and obstacles in the orchard in real time, ignores the trunk root area which is easily blocked, takes the midpoint of the lower edge of the trunk recognition box output by the target detection model as the tree row fitting reference point, and based on the fitting reference point, uses the least squares method to fit the single-side tree row line and the tree row center navigation line to generate a target navigation path;

[0012] Step 5: The executing mechanism controls the weeding robot to walk according to the target navigation path, in the walking process, the weeding robot collects visual data streams through the video acquisition device installed thereon and transmits them to the navigation control module, the navigation control module calculates the moving direction of the robot based on the visual data streams and the fitting algorithm, then compares and analyzes it with the target navigation path, calculates the deviation angle, and dynamically adjusts the moving direction and turning amplitude of the weeding robot.

[0013] Further, the specific process of step 4 is as follows:

[0014] First, after identifying the trees on both sides of the tree row through the target detection model, the coordinates of the opposite edges of the trunk recognition box are processed to obtain the midpoint coordinates of the lower edge of the trunk recognition box; then the coordinates of one side of the tree row are marked as , the fitting straight line equation is , the data set of the fitting straight line is , and the error sum of squares formula is:

[0015] (1)

[0016] wherein, , respectively represent the slope and intercept of the fitting straight line; , , represents the coordinate of the midpoint of the lower edge of each trunk recognition frame in the one-side tree row, represents the sum of squared errors, represents the y coordinate of the midpoint of the lower edge of the trunk recognition frame, represents the fitting straight line; represents the number of the trunk recognition frame in the one-side tree row; represents the number of the trunk recognition frame in the one-side tree row;

[0017] The partial derivatives of the parameters in the above formula (1) are respectively obtained as follows: ,

[0018] (2)

[0019] According to the above formula (1) and (2), the following is obtained:

[0020] (3)

[0021] The best estimation values of the parameters , in the formula (3) are obtained by solving the formula (3) as follows:

[0022] (4)

[0023] After the values of , are calculated, the expression of the one-side tree row line is obtained by combining the dataset of the midpoints of the lower edges of the trunk recognition frames as follows: , and the expression of the other-side tree row line is obtained as follows: ;

[0024] In the formula, represents the equation of the one-side fitting straight line, represents the slope of the equation of the one-side tree row line, represents the intercept of the equation of the one-side tree row line, represents the equation of the other-side fitting straight line, represents the slope of the equation of the other-side tree row line, represents the intercept of the equation of the other-side tree row line;

[0025] After the expressions of the two-side tree row lines are obtained, a plurality of points are selected as the reference datum for fitting at equal intervals on the one-side tree row line, and the coordinates of the one-side tree row reference points are P1 ​The longitudinal coordinates of the plurality of points are substituted into the expression of the tree row line on the other side respectively to obtain the corresponding coordinates P2 ;

[0026] Based on the plurality of reference points on the left and right sides, a plurality of reference points P of the navigation line located in the center of the tree row are calculated by coordinate mean value m wherein, , Subsequently, in combination with the central navigation line reference points, a straight line equation is fitted, and the formulas (1) to (4) are combined to solve the best straight line matching function of the navigation line reference points by minimizing the square error, to obtain the expression of the navigation line in the center of the orchard tree row wherein, represents the fitted navigation line equation, represents the slope of the fitted navigation line, represents the intercept of the fitted navigation line.

[0027] Further, in step 2, the ConvNeXt V2 module is a pure convolutional neural network architecture that introduces a sparse convolution mechanism to generate learning signals by randomly masking the original input. With the help of the full convolutional mask autoencoder FCMAE and the new global response normalization layer GRN, the inter-channel feature competition is enhanced.

[0028] Further, in step 2, the multi-scale expansion attention further enriches and improves the general representation ability of the YOLO v8 model by incorporating the self-attention mechanism. Its framework integrates three key links: multi-scale processing, attention calculation, and feature rearrangement, forming an optimized pattern recognition scheme. Specifically, the multi-scale expansion attention selectively focuses on the regions with higher attention scores in the input feature map for processing.

[0029] Further, in step 5, when the deviation angle exceeds the preset range, according to the deviation direction and angle value, the control parameters including the rotation direction and rotation angle are output to adjust the moving direction and turning amplitude of the weeding robot; if the deviation angle does not exceed the preset range or is completely coincident with the target navigation path, the weeding robot is controlled to maintain straight-line travel without adjustment.

[0030] A system for implementing the above-mentioned visual perception-based dynamic path planning method for orchard weeding robots, comprising

[0031] An OAK-D-PRO camera is deployed at the front end of the weeding robot to collect real-time image data of the trees on both sides of the orchard and transmit it to the navigation control module;

[0032] The navigation control module, mounted on the Jetson Xavier NX embedded platform, is used to run the target detection model and path planning algorithm, identify tree trunks and obstacles between rows in the orchard, extract the lower midpoint of the tree trunk recognition box as the tree row fitting positioning base point, and use the least squares method to fit the single-sided tree row line and the tree row center navigation line based on the positioning base point to generate the target navigation path, detect the path deviation in real time, and output control parameters to the actuator to adjust the movement direction of the weeding robot;

[0033] The actuator controls the servo motor via the CAN bus to drive the wheeled chassis of the weeding robot and execute path adjustment commands.

[0034] The present invention has the following beneficial effects:

[0035] By improving the YOLO v8 network architecture, replacing some c2f modules with ConvNeXt V2 modules and adding MSDA, the target recognition accuracy was improved. By combining the least squares method of fitting the tree line and navigation line with the midpoint of the bottom edge of the recognition box, the deviation between the actual walking path and the target navigation path was determined and adjusted in real time. This solved the problem of large deviation of the navigation path due to environmental influence in the existing technology, and realized efficient and accurate path planning for orchard robots. Attached Figure Description

[0036] Figure 1 To improve the network architecture diagram of the YOLO v8 model;

[0037] Figure 2 A schematic diagram showing the positional deviation between the actual walking path of the weeding robot and the target navigation path;

[0038] Figure 3 Here is a flowchart of the path planning method;

[0039] Figure 4 This is a schematic diagram of extracting the midpoint of the bottom edge of the tree trunk recognition box. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0041] Reference Figure 3 As shown, the dynamic path planning method for orchard weeding robots based on vision perception described in this invention includes the following process:

[0042] Step 1: Considering the variability of light intensity and the shape of the detected object, which may affect the detection speed and accuracy, photos were collected under different weather conditions in different seasons to obtain the orchard road dataset, which was then divided into training set, validation set and test set in a ratio of 7:2:1.

[0043] Step 2: Construct an improved YOLO v8 network structure based on the YOLO v8 algorithm to obtain the improved YOLO v8 model: Refer to... Figure 1 As shown, the ConvNeXt V2 module replaces part of the C2f module in the backbone network, and a multi-scale expanded attention (MSDA) mechanism is added after the SPPF layer in the backbone. The MSDA mechanism enhances the robustness of the tree trunk shape and obstacle recognition through multi-scale feature rearrangement and attention weight calculation. The above improvements not only enhance the model's recognition and pattern adaptation capabilities, but also improve feature extraction and further increase efficiency. Therefore, the modified model can accurately identify targets of different shapes and sizes while maintaining high efficiency, thus providing a significant performance improvement for segmentation tasks. Figure 1 In this module, ConvNX2 is also known as ConvNeXt V2. The 2D convolution is Conv2d, the bounding box loss is Bbox Loss, the classification loss is Cls Loss, the connection layer is Concat, and the upsampling is Upsample.

[0044] The ConvNeXt V2 module is a pure convolutional neural network architecture that introduces a sparse convolution mechanism. It generates learning signals by randomly masking the original input. This design forces the model to use the remaining context information to predict the missing parts, thereby significantly improving the model's ability to learn and understand visual data. The architecture effectively enhances feature competition between channels by using a fully convolutional masked autoencoder (FCMAE) and a novel global response normalization layer (GRN), making feature extraction more accurate and efficient.

[0045] Multi-Scale Expanded Attention (MSDA) further enriches the model's general representation capabilities by incorporating a self-attention mechanism. Its framework integrates three key components: multi-scale processing, attention computation, and feature rearrangement, forming an optimized pattern recognition scheme. Specifically, it selectively focuses on processing regions with high attention scores in the input feature map. This characteristic enables it to exhibit superior adaptability and processing accuracy when dealing with image data of different scales.

[0046] Step 3: Use the dataset established in Step 1 to train, validate, and test the improved YOLO v8 model; load the best weight file obtained after training into the improved YOLO v8 model to obtain the object detection model, and load the object detection model into the navigation control module.

[0047] Step 4: The weeding robot equipped with an OAK-D-PRO camera moves forward in the orchard. The OAK-D-PRO camera collects real-time image data of the fruit trees on both sides of the orchard and transmits it to the navigation control module; refer to Figure 4As shown, the navigation control module uses the target detection model to identify tree trunks and obstacles between rows in the orchard in real time, ignoring the easily obscured root area of ​​the tree trunks, and uses the midpoint of the bottom edge of the tree trunk recognition box output by the target detection model as the tree row fitting reference point; based on the fitting reference point, the least squares method is used to fit the single-sided tree row line and the tree row center navigation line to generate the target navigation path, that is, the driving route of the weeding robot during the operation.

[0048] The specific process is as follows:

[0049] First, after identifying the trees on both sides of the tree row using the object detection model, the coordinates of the opposite sides of the tree trunk recognition box are processed to obtain the coordinates of the midpoint of the bottom edge of the tree trunk recognition box. Preferably, the coordinates of the upper left corner of the tree trunk recognition box are set as (x1, y1), and the coordinates of the lower right corner are set as (x2, y2). Then, the coordinates of the midpoint of the bottom edge of the tree trunk recognition box are (x1, y1, x2, y2 ... 1+ (x2) / 2, y2);

[0050] After obtaining the coordinates of the lower midpoint of the tree trunk recognition box, mark the coordinates of one side of the tree row as follows: The fitted straight line equation is The dataset for fitting the straight line is The formula for the sum of squared errors is:

[0051] (1)

[0052] in, , These represent the slope and intercept of the fitted line, respectively. , , This represents the coordinates of the midpoint of the bottom edge of each tree trunk identification box in a row of trees on one side. Represents the sum of squared errors. This represents the y-coordinate of the midpoint of the bottom edge of the tree trunk recognition box. This represents the fitted straight line; Indicates the tree trunk identification box number in one side of the tree row; Indicates the number of tree trunk identification boxes in one side of the tree row;

[0053] For the parameters in equation (1) above , Taking the partial derivatives separately, we can obtain

[0054] (2)

[0055] According to equations (1) and (2) above, we can obtain:

[0056] (3)

[0057] Solving equation (3) yields the parameters. , Best estimate:

[0058] (4)

[0059] Calculated , After obtaining the values, and combining them with the dataset of the midpoints of the tree trunk recognition boxes, the expression for one side of the tree line can be obtained. Similarly, the expression for the treeline on the other side can be obtained. ;

[0060] In the above formula, This represents the fitted linear equation for one side (the left side in this example). This represents the slope of the treeline equation on one side. This represents the intercept of the treeline equation on one side. This indicates the fitted linear equation for the other side (the right side in this example). This represents the slope of the tree line equation on the other side. This represents the intercept of the treeline equation on the other side.

[0061] After obtaining the expressions for the tree rows on both sides, eight points are selected at equal intervals on one side of the tree row as reference benchmarks for center fitting. The coordinates of the reference points on one side of the tree row are P1. ,in =1, 2, ..., 8. Substitute the ordinates of these 8 points into the expression for the tree line on the other side to obtain the corresponding coordinates P2. ,in =1, 2, ..., 8.

[0062] Based on 16 reference points on the left and right sides, 8 reference points P for the navigation line located in the center of the tree row were calculated using the coordinate mean. m ,in, , , =1, 2, ..., 8. Then, combining these 8 central navigation line reference points, the straight line equation is fitted. By combining equations (1) to (4), the optimal straight line matching function with the navigation line reference points is solved by minimizing the squared error, and the expression of the navigation line in the center of the orchard tree row is obtained. ,in, This indicates the fitted navigation line equation. This represents the slope of the fitted navigation line. This represents the intercept of the fitted navigation line.

[0063] Step 5: The actuator controls the weeding robot to move according to the target navigation path. During movement, the weeding robot collects visual data streams through its onboard video acquisition equipment and transmits them to the navigation control module. The navigation control module calculates the robot's movement direction based on the visual data streams and a fitting algorithm, then compares and analyzes it with the target navigation path to calculate the deviation angle. The closed-loop control system then dynamically adjusts the weeding robot's movement direction and turning angle to ensure the orchard robot moves along the target navigation path. The specific process is as follows:

[0064] Reference Figure 2 As shown, the initial direction of movement of the weeding robot is its initial actual walking path. There are two types of deviations between the actual walking path of the weeding robot and the target navigation path: the deviation to one side of the target navigation path exceeds a predetermined angle, or the deviation angle overlaps with the target navigation path or is within a preset range.

[0065] In response to the two situations described above, this embodiment determines the deviation in the following specific way:

[0066] The location and direction of movement of the weeding robot are continuously acquired to determine the deviation between the actual walking path and the target navigation path. This deviation includes both the magnitude of the deviation angle and the direction of deviation.

[0067] The robot's movement direction is corrected in real time based on the calculated deviation, ensuring that it travels along the target navigation path. Specifically:

[0068] If the actual walking path is as follows Figure 2 As shown, if the angle of deviation to the left or right exceeds the preset range, control parameters including rotation direction and rotation angle will be output according to the deviation direction and angle value to adjust the movement direction and steering amplitude of the weeding robot; if the deviation angle does not exceed the preset range, or completely coincides with the target navigation path, the weeding robot will be controlled to keep moving in a straight line without adjustment.

[0069] The visual perception-based orchard weeding robot dynamic path planning system of the present invention includes:

[0070] The OAK-D-PRO camera, deployed at the front of the weeding robot, is installed at a height of 1.2m above the ground with a pitch angle of 15°. It is used to collect real-time image data of fruit trees on both sides of the orchard and transmit it to the navigation control module.

[0071] The navigation control module, mounted on the Jetson Xavier NX embedded platform, is used to run the target detection model and path planning algorithm, identify tree trunks and obstacles between rows in the orchard, extract the lower midpoint of the tree trunk recognition box as the tree row fitting positioning base point, and use the least squares method to fit the single-sided tree row line and the tree row center navigation line based on the positioning base point to generate the target navigation path, and detect the path deviation in real time, output control parameters to the actuator to adjust the movement direction of the weeding robot so that the weeding robot walks along the target navigation path;

[0072] The actuator controls the servo motor via CAN bus to drive the wheeled chassis of the weeding robot, with a steering accuracy of ±0.5°, and executes path adjustment commands.

[0073] In calculating the target navigation path, this invention uses the fruit trees on both sides in front of the robot as references, takes the midpoint of each pair of opposite fruit trees on both sides, and then connects these midpoints sequentially to form a centerline. This centerline is the robot's movement path. Based on the above method for determining the target navigation path, the reasonable selection of features is key to achieving accurate navigation when the weeding robot travels in an orchard environment. Orchard environments are often complex, and the roots of fruit trees may be obscured by problems such as branching and weeds, making it difficult to identify root features. However, fruit trees themselves have significant usable characteristics, namely, the trunk and crown are clearly separated, and the trunk is relatively tall and not easily obscured by weeds. In addition, the positional deviation between the midpoint of the trunk base and the center of the root is small. Combining these advantages, this invention selects the midpoint of the bottom of the fruit tree trunk as the reference point for fitting the navigation path.

[0074] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A dynamic path planning method for an orchard weeding robot based on visual perception, characterized in that, The process includes the following: Step 1: Collect photos under different weather conditions in different seasons to obtain orchard road dataset, and divide it into training set, validation set and test set according to the corresponding proportions; Step 2: Construct an improved YOLO v8 network structure to obtain an improved YOLO v8 model: replace some c2f modules with ConvNeXt V2 modules in the backbone network, and add a multi-scale expanded attention mechanism, i.e., MSDA mechanism, after the SPPF layer in the backbone. Step 3: Use the dataset established in Step 1 to train, validate, and test the improved YOLO v8 model; load the best weight file obtained after training into the improved YOLO v8 model to obtain the object detection model, and load the object detection model into the navigation control module; Step 4: The weeding robot equipped with an OAK-D-PRO camera moves forward in the orchard. The OAK-D-PRO camera collects real-time image data of the fruit trees on both sides of the orchard and transmits it to the navigation control module. The navigation control module uses a target detection model to identify tree trunks and obstacles between rows in real time, ignoring the easily obscured root areas of the tree trunks. The midpoint of the bottom edge of the tree trunk recognition box output by the target detection model is used as the tree row fitting reference point. Based on the fitting reference point, the least squares method is used to fit the single-sided tree row line and the central navigation line of the tree row to generate the target navigation path. Step 5: The actuator controls the weeding robot to walk according to the target navigation path. During the walking process, the weeding robot collects visual data streams through the video acquisition device installed on it and transmits them to the navigation control module. The navigation control module calculates the robot's movement direction based on the visual data streams and fitting algorithms, and then compares and analyzes it with the target navigation path to calculate the deviation angle. It then dynamically adjusts the movement direction and turning amplitude of the weeding robot. The specific process of step 4 is as follows: First, the trees on both sides of the tree row are identified using an object detection model. Then, the coordinates of the opposite sides of the tree trunk recognition box are processed to obtain the coordinates of the midpoint of the bottom edge of the tree trunk recognition box. Next, the coordinates of one side of the tree row are designated as (x1, y1), and the fitted straight line equation is y = ax + b. The data set for the fitted straight line is A = [(x1, y1), (x2, y2), ..., (x...]. n ,y n The formula for the sum of squared errors is: Where a and b represent the slope and intercept of the fitted line, respectively; (x1,y1), (x2,y2), (x n ,y n ) represents the coordinates of the midpoint of the bottom edge of each tree trunk recognition box in one side of the tree row, δ represents the sum of squared errors, and y i ax represents the y-coordinate of the midpoint of the bottom edge of the tree trunk recognition box. i +b represents the fitted straight line; i represents the tree trunk identification box number in one side of the tree row; n represents the number of tree trunk identification boxes in one side of the tree row; Taking partial derivatives of parameters a and b in equation (1) above, we get: Based on equations (1) and (2) above, we obtain: Solving equation (3) yields the best estimates for parameters a and b: After calculating the values ​​of a and b, and combining them with the dataset of the midpoints of the lower edge of the tree trunk recognition box, the expression L for one side of the tree line is obtained. l =a l x+b l Similarly, the expression for the treeline on the other side, L, is obtained. r =a r x+b r ; In the formula, L l Let a represent the equation of the fitted line on one side. l b represents the slope of the treeline equation on one side. l L represents the intercept of the treeline equation on one side. r This represents the fitted straight line equation on the other side, a r b represents the slope of the tree line equation on the other side. r This represents the intercept of the treeline equation on the other side; After obtaining the expressions for the tree rows on both sides, multiple points are selected at equal intervals on one side of the tree row as reference benchmarks for center fitting. The coordinates of the reference points on one side of the tree row are P1(x i ,y i Substitute the ordinates of these multiple points into the tree line expression on the other side to obtain the corresponding coordinates P2(x). j ,y j ); Based on multiple reference points on the left and right sides, multiple reference points P for the navigation lines located in the center of the tree row are calculated using the coordinate mean. m (x k ,y k ),in, Subsequently, by combining these central navigation line reference points, a straight line equation is fitted. By combining equations (1) to (4) and minimizing the squared error, the optimal straight line matching function with the navigation line reference points is obtained, thus yielding the expression L for the central navigation line of the orchard tree row. m =a m x+b m , where L m This represents the fitted navigation line equation, a m b represents the slope of the fitted navigation line. m This represents the intercept of the fitted navigation line.

2. The dynamic path planning method for orchard weeding robots based on visual perception according to claim 1, characterized in that, In step 2, the ConvNeXt V2 module is a pure convolutional neural network architecture that introduces a sparse convolution mechanism. It generates learning signals by randomly masking the original input. This architecture enhances feature competition between channels by using a fully convolutional masked autoencoder (FCMAE) and a novel global response normalization layer (GRN).

3. The dynamic path planning method for orchard weeding robots based on visual perception according to claim 1, characterized in that, In step 2, multi-scale expanded attention further enriches and improves the general representation capability of the YOLO v8 model by incorporating a self-attention mechanism. Its framework integrates three key links: multi-scale processing, attention calculation, and feature rearrangement, forming an optimized pattern recognition scheme. Specifically, multi-scale expanded attention selectively focuses on processing regions with high attention scores in the input feature map.

4. The dynamic path planning method for orchard weeding robots based on visual perception according to claim 1, characterized in that, In step 5, if the deviation angle exceeds the preset range, control parameters including the rotation direction and rotation angle are output according to the deviation direction and angle value to adjust the movement direction and steering amplitude of the weeding robot; if the deviation angle does not exceed the preset range, or completely coincides with the target navigation path, the weeding robot is controlled to maintain straight-line travel without adjustment.

5. A system for implementing the dynamic path planning method for an orchard weeding robot based on vision perception as described in claim 1, characterized in that, include The OAK-D-PRO camera, deployed at the front end of the weeding robot, is used to collect real-time image data of fruit trees on both sides of the orchard and transmit it to the navigation control module. The navigation control module, mounted on the Jetson Xavier NX embedded platform, is used to run the target detection model and path planning algorithm, identify tree trunks and obstacles between rows in the orchard, extract the lower midpoint of the tree trunk recognition box as the tree row fitting positioning base point, and use the least squares method to fit the single-sided tree row line and the tree row center navigation line based on the positioning base point to generate the target navigation path, detect the path deviation in real time, and output control parameters to the actuator to adjust the movement direction of the weeding robot; The actuator controls the servo motor via the CAN bus to drive the wheeled chassis of the weeding robot and execute path adjustment commands.

Images