Light intelligent orchard edge navigation method

Abstract

The invention relates to a light intelligent orchard edge navigation method which comprises the following steps: aerial photography is performed on the terrain of an orchard through an unmanned aerial vehicle to obtain an orchard topographic map; carrying out image preprocessing on the obtained orchard topographic map; establishing a coordinate system in the orchard topographic map after image preprocessing, and obtaining coordinates of each fruit tree in the orchard topographic map; historical orchard edge path information data and steering motion information data of the robot are called to form a historical motion data set; based on the historical motion data set of the robot and the coordinates of each fruit tree, constructing a three-dimensional orchard model by adopting a digital twin technology; training the historical motion data set of the robot to obtain a historical optimal steering path state recognition model of the robot; and determining a real-time starting point and a real-time ending point of the robot, and acquiring a real-time optimal steering inflection point path of the robot by utilizing a historical optimal path state recognition model. The image definition can be improved, and the fruit tree recognition effect is improved.

Description

Technical Field

[0001] The present invention relates to the field of robot automatic navigation, and specifically relates to a light-intelligent orchard field head navigation method. Background Art

[0002] The high-quality orchards where the main fruits are produced in China are mainly distributed in hilly areas. With the continuous progress of technology, some automated equipment has begun to replace traditional manual operations and is gradually applied to the production, management, and harvesting of hilly orchards, reducing labor intensity and improving production efficiency.

[0003] Inter-row navigation and field head navigation are necessary links for orchard robots to achieve full-process autonomous operation. However, currently, the research on orchard robot navigation mainly focuses on the inter-row environment, with less research on turning at the field head, and most of the research on field head navigation is carried out around the field environment. The difficulties are as follows: First, the orchard is closed, resulting in the loss of GNSS positioning information and the inability to achieve real-time absolute coordinate positioning; second, restricted by factors such as terrain conditions and planting patterns, the field head situation varies greatly, it is difficult to identify the field head, and the difficulty of automatically changing rows is relatively high. Therefore, generally, these robots switch to the manual driving mode when turning. Therefore, solving the difficulties of orchard field head navigation has practical significance for the autonomous operation of orchard robots. Summary of the Invention

[0004] Aiming at the defects of the prior art, the purpose of the present invention is to provide a light-intelligent orchard field head navigation method to solve the problem of difficult determination of the best turning strategy of the robot at the field head in the above-mentioned prior art.

[0005] In order to achieve the above purpose, the technical solution adopted by the present invention is: A light-intelligent orchard field head navigation method, characterized by including the following steps: S1. Aerial photograph the terrain of the orchard by using an unmanned aerial vehicle to obtain an orchard topographic map; S2. Perform image preprocessing on the obtained orchard topographic map, and the image preprocessing includes image denoising, image enhancement, and image segmentation; S3. Establish a coordinate system in the orchard topographic map after image preprocessing to obtain the coordinates of each fruit tree in the orchard topographic map; S4. Retrieve the historical orchard field head path information data and turning motion information data of the robot, and form a historical motion data set; S5. Based on the historical motion data set of the robot and the coordinates of each fruit tree, construct a three-dimensional orchard model by using digital twin technology; S6. Based on the neural network model, train the historical motion data set of the robot to obtain a historical best turning path state recognition model of the robot; S7. Determine the real-time starting point and the real-time ending point of the robot, retrieve the pre-set starting and ending point similarity matching model and the inflection point matching model from the historical motion data set of the robot, and use the historical optimal path state recognition model to obtain the real-time optimal turning inflection point path of the robot.

[0006] As an optimal solution, in step S2, image denoising includes the following steps: S21. Perform adaptive wavelet threshold denoising on the orchard topographic map; S22. Determine the optimal wavelet threshold by constructing a probability model of wavelet coefficients and using Bayesian estimation; S23. Perform adaptive threshold processing on the wavelet coefficients and perform wavelet reconstruction to obtain the denoised image.

[0007] As an optimal solution, in step S2, image enhancement includes the following steps: S24. Perform adaptive histogram equalization processing on the denoised image; S25. Determine the cumulative distribution function of the local area by calculating the gray-level histogram of the denoised image; S26. Based on the cumulative distribution function, adaptively adjust the gray-level mapping relationship of the local area to obtain the enhanced image.

[0008] As an optimal solution, in step S2, image segmentation includes the following steps: S27. Segment the enhanced image into multiple superpixels by the superpixel segmentation algorithm; S28. Determine the similarity matrix between superpixels as the edge weights of the undirected graph, and construct the superpixel graph cut undirected graph; S29. Combine the pre-set probability distribution of the text area, and solve the global optimal solution of the superpixel graph cut undirected graph by the maximum flow minimum cut algorithm; S210. Obtain the segmentation result of the fruit tree area to obtain the orchard topographic map after image preprocessing.

[0009] As an optimal solution, step S3 includes the following sub-steps: S31. Connect the diagonal points in the orchard topographic map after image preprocessing to obtain two diagonals; S32. Based on the intersection part of the two diagonals, determine the midpoint of the topographic map; S33. Take the midpoint of the topographic map as the origin of the coordinate system, and establish a quadrant coordinate system in the orchard topographic map after image preprocessing; S34. Analyze the position of each fruit tree in the orchard topographic map to obtain the horizontal distance and the vertical distance of each fruit tree from the origin; S35. Based on the horizontal distance and the vertical distance, determine the coordinate information of each fruit tree.

[0010] As a preferred solution, step S4 includes the following sub-steps: S41. Retrieve the historical driving path information of the robot from the database, and assign English and numerical labels A1, A2,.......An; S42. Analyze the historical driving path information each time to determine the inflection point information of the robot; S43. Assign numerical labels A1-1, A1-2,.......A1-n to each inflection point information in the historical driving path information each time; S44. Based on the path information labels and the inflection point information labels in each path, form a set to constitute the historical motion data set of the robot.

[0011] As a preferred solution, step S5 includes the following sub-steps: S51. Construct a three-dimensional map based on the pre-processed orchard topographic map of the image; S52. According to the coordinate information of each fruit tree, map the position of each fruit tree in the three-dimensional map; S53. Incorporate the historical motion data set of the robot into the three-dimensional map of the orchard in the form of dynamic simulation to construct a three-dimensional orchard model.

[0012] As a preferred solution, step S6 includes the following sub-steps: S61. Based on the feedforward neural network, construct an optimal steering path state recognition model; S62. Based on machine learning, label and divide the multiple sample historical motion data sets to obtain a training set, a validation set, and a test set; S63. Use the training set, the validation set, and the test set to perform supervised training, validation, and testing on the optimal steering path state recognition model to obtain the optimal steering path state recognition model with an accuracy rate meeting the preset accuracy rate requirements.

[0013] As a preferred solution, in step S7, the start-end point similarity matching model is: In the formula, S i is the similarity between the real-time start-end point data and the i-th feature of the historical optimal steering inflection point path of the robot, w ij is the j-th feature index value of the i-th feature of the real-time start-end point data, v ij is the j-th feature index value of the i-th feature of the historical optimal steering inflection point path, and n is the total number of the i-th feature index values.

[0014] As a preferred solution, in step S7, the inflection point matching model is: In the formula, C is the fitness degree between the real start and end point data and the historical best turning inflection point path, and α i is the weight value of the i-th characteristic parameter, and m is the total number of characteristic parameters.

[0015] The beneficial effects of the present application are as follows: The present invention obtains the orchard topographic map through a drone, removes the noise in the orchard topographic map to improve the accuracy of subsequent fruit tree recognition. In order to enhance the contrast of the image and improve the image clarity to enhance the effect of fruit tree recognition, image enhancement processing is performed on the denoised image, and the image is segmented into different regions to better identify and locate the fruit tree regions. Secondly, a quadrant coordinate system is constructed with the intersection point of the two diagonals of the orchard topographic map as the origin. Then, according to the horizontal distance and vertical distance of each fruit tree from the origin, the coordinates of each fruit tree are obtained. A three-dimensional orchard model is constructed based on the digital twin technology, and the motion trajectory data of the historical robot in the orchard is repeatedly trained based on the neural network model to determine the optimal turning path state of the robot. Based on this, according to the real-time start and end points of the robot, the similarity matching model and the inflection point matching model are used, and the historical best path state recognition model is utilized to obtain the real-time best turning inflection point path of the robot. Brief Description of the Drawings

[0017] Figure 1 is a schematic diagram of the present invention;

[0018] Figure 2 is a schematic diagram of image denoising in the present invention;

[0019] Figure 3 is a schematic diagram of image enhancement in the present invention;

[0020] Figure 4 is a schematic diagram of image segmentation in the present invention;

[0021] Figure 5 is a schematic diagram of obtaining the coordinates of each fruit tree in the orchard topographic map in the present invention. Detailed Embodiments

[0022] The following further describes the present invention in detail with reference to the drawings and specific embodiments.

[0023] Combined with Figures 1-5 shown, the present application provides a light intelligent orchard headland navigation method, including the following steps: S1. Aerial photograph the terrain of the orchard through a drone to obtain the orchard topographic map; S2. Perform image preprocessing on the obtained orchard topographic map, and the image preprocessing includes image denoising, image enhancement and image segmentation; S3. Establish a coordinate system in the pre-processed orchard topographic map, and obtain the coordinates of each fruit tree in the orchard topographic map; S4. Retrieve the historical orchard field path information data and turning motion information data of the robot, and form a historical motion data set; S5. Based on the historical motion data set of the robot and the coordinates of each fruit tree, construct a 3D orchard model using digital twin technology; S6. Based on the neural network model, train the historical motion data set of the robot to obtain a historical optimal turning path state recognition model of the robot; S7. Determine the real-time starting point and real-time ending point of the robot, retrieve the pre-set starting and ending point similarity matching model and inflection point matching model from the historical motion data set of the robot, and use the historical optimal path state recognition model to obtain the real-time optimal turning inflection point path of the robot.

[0024] Among them, as shown in Figure 2, in step S2, image denoising includes the following steps: S21. Perform adaptive wavelet threshold denoising on the orchard topographic map; S22. By constructing a probability model of wavelet coefficients, use Bayesian estimation to determine the optimal wavelet threshold; S23. Perform adaptive threshold processing on the wavelet coefficients and perform wavelet reconstruction to obtain a denoised image.

[0025] Combined with Figure 3 As shown in, in step S2, image enhancement includes the following steps: S24. Perform adaptive histogram equalization processing on the denoised image; S25. By calculating the gray histogram of the denoised image, determine the cumulative distribution function of the local area; S26. Based on the cumulative distribution function, adaptively adjust the gray mapping relationship of the local area to obtain an enhanced image.

[0026] Combined with Figure 4 As shown in, in step S2, image segmentation includes the following steps: S27. Segment the enhanced image into multiple superpixels by the superpixel segmentation algorithm; S28. Determine the similarity matrix between superpixels as the edge weights of the undirected graph, and construct a superpixel graph cut undirected graph; S29. Combine the pre-set probability distribution of the text area, and solve the global optimal solution of the superpixel graph cut undirected graph by the maximum flow minimum cut algorithm; S210. Obtain the segmentation result of the fruit tree area to obtain the pre-processed orchard topographic map of the image.

[0027] Combined with Figure 5As shown in the figure, step S3 includes the following sub-steps: S31. Connect the diagonal points in the pre-processed orchard topographic map to obtain two diagonals; S32. Determine the midpoint of the topographic map based on the intersection of the two diagonals; S33. Take the midpoint of the topographic map as the origin of the coordinate system, and establish a quadrant coordinate system in the pre-processed orchard topographic map; S34. Analyze the position of each fruit tree in the orchard topographic map to obtain the horizontal distance and vertical distance of each fruit tree from the origin; S35. Determine the coordinate information of each fruit tree based on the horizontal distance and vertical distance.

[0028] Specifically, step S4 includes the following sub-steps: S41. Retrieve each historical driving path information of the robot from the database, and label it with English and numbers A1, A2,.......An; S42. Analyze each historical driving path information to determine the inflection point information of the robot; S43. Label each inflection point information in each historical driving path information with numbers A1-1, A1-2,.......A1-n; S44. Based on the path information label and the inflection point information label in each path, form a set to constitute the historical motion data set of the robot.

[0029] Step S5 includes the following sub-steps: S51. Construct a three-dimensional map according to the pre-processed orchard topographic map; S52. Map the position of each fruit tree in the three-dimensional map according to the coordinate information of each fruit tree; S53. Incorporate the historical motion data set of the robot into the three-dimensional map of the orchard in the form of dynamic simulation to construct a three-dimensional orchard model.

[0030] Specifically, step S6 includes the following sub-steps: S61. Based on the feedforward neural network, construct an optimal steering path state recognition model; S62. Based on machine learning, label and divide the multiple sample historical motion data sets to obtain a training set, a validation set, and a test set; S63. Use the training set, the validation set, and the test set to perform supervised training, validation, and testing on the optimal steering path state recognition model to obtain the optimal steering path state recognition model with an accuracy rate meeting the preset accuracy rate requirements.

[0031] Specifically, in step S7, the start-end point similarity matching model is: Wherein, S i is the similarity between the real-time start and end point data and the i-th feature of the historical best turning inflection point path of the robot, w ij is the j-th feature index value of the i-th feature of the real-time start and end point data, v ij is the j-th feature index value of the i-th feature of the historical best turning inflection point path, and n is the total number of the i-th feature index values.

[0032] More specifically, in step S7, the inflection point matching model is as follows: Wherein, C is the fitness between the real-time start and end point data and the historical best turning inflection point path, α i is the weight value of the i-th feature parameter, and m is the total number of feature parameters. It should be noted that the parts not detailed in this application are all prior arts.

[0033] Secondly, in step S7, the present invention first determines the path, start point and end point. For example, there are two fruit trees, the start point is denoted as K, and the end point is denoted as L. There are multiple paths to choose from when moving from the start point K to the end point L. According to the start and end point similarity matching model, the most suitable path is selected. Since when the robot turns, different inflection point forms are different, that is, similar to the "radius" of the robot's turning, which will also affect the arrival efficiency of the robot. Therefore, based on the historical best turning path state recognition model, the robot can find the best state every time it turns, so as to reach quickly and meet the requirements.

[0034] The present application also provides a lightweight intelligent orchard field navigation system, including a shooting module, an image processing module, a coordinate acquisition module, a retrieval module, a first modeling module, a second modeling module, and an optimal turning inflection point path determination module. The shooting module is used to shoot the terrain of the orchard by using UAV aerial photography technology to obtain an orchard topographic map. The image processing module is used to perform image preprocessing on the obtained orchard topographic map, and the preprocessing includes image denoising, image enhancement, and image segmentation. The coordinate acquisition module is used to establish a coordinate system in the preprocessed orchard topographic map to obtain the coordinates of each fruit tree in the orchard topographic map. The retrieval module is used to retrieve the historical orchard field path information data and turning motion information data of the robot and form a historical motion data set. The modeling module is used to construct a three-dimensional orchard model by using digital twin technology based on the coordinates of each fruit tree and the historical motion data set of the robot. The second modeling module is used to train the historical motion data set of the robot based on a neural network model to obtain the historical optimal turning path state of the robot. The optimal turning inflection point path determination module is used to determine the real-time starting point and the real-time ending point of the robot, retrieve the preset starting and ending point similarity matching model and inflection point matching model from the historical motion data set of the robot, and use the historical optimal path state recognition model to obtain the real-time optimal turning inflection point path of the robot.

[0035] The present invention obtains the orchard topographic map through UAV technology, removes the noise in the orchard topographic map to improve the accuracy of subsequent fruit tree recognition. In order to enhance the contrast of the image and improve the image clarity to enhance the effect of fruit tree recognition, image enhancement processing is performed on the denoised image, and the image is segmented into different regions to better identify and locate the fruit tree region. Secondly, a quadrant coordinate system is constructed with the intersection of the two diagonals of the orchard topographic map as the origin, so as to obtain the coordinates of each fruit tree according to the horizontal distance and vertical distance of each fruit tree from the origin. A three-dimensional orchard model is constructed based on digital twin technology, and the historical motion trajectory data of the robot in the orchard is repeatedly trained based on a neural network model to determine the optimal turning path state of the robot. Based on this, according to the real-time starting point and ending point of the robot, a similarity matching model and an inflection point matching model are used, and the historical optimal path state recognition model is used to obtain the real-time optimal turning inflection point path of the robot.

[0036] The above are only the preferred embodiments of the present invention and do not impose any formal restrictions on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments with equivalent changes within the scope of the technical solution of the present invention by using the above-disclosed technical content. However, as long as it does not depart from the content of the technical solution of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A light intelligent orchard field navigation method, characterized in that, It includes the following steps: S1. Use a drone to conduct aerial photography of the orchard terrain to obtain an orchard topographic map; S2. Perform image preprocessing on the obtained orchard topographic map. The image preprocessing includes image denoising, image enhancement, and image segmentation; S3. Establish a coordinate system in the orchard topographic map after image preprocessing, and obtain the coordinates of each fruit tree in the orchard topographic map; S4. Retrieve the historical orchard head path information data and turning motion information data of the robot, and form a historical motion data set; S5. Based on the historical motion data set of the robot and the coordinates of each fruit tree, use digital twin technology to construct a three-dimensional orchard model; S6. Based on the neural network model, train the historical motion data set of the robot to obtain a historical optimal turning path state recognition model of the robot; S7. Determine the real-time starting point and real-time ending point of the robot, retrieve the preset starting and ending point similarity matching model and inflection point matching model from the historical motion data set of the robot, and use the historical optimal path state recognition model to obtain the real-time optimal turning inflection point path of the robot.

2. The light intelligent orchard field navigation method according to claim 1, wherein In step S2, Image denoising includes the following steps: S21. Perform adaptive wavelet threshold denoising on the orchard topographic map; S22. By constructing a probability model of wavelet coefficients, use Bayesian estimation to determine the optimal wavelet threshold; S23. Perform adaptive threshold processing on the wavelet coefficients and perform wavelet reconstruction to obtain a denoised image.

3. The method for light intelligent orchard field navigation according to claim 2, characterized in that, In step S2, image enhancement includes the following steps: S24. Perform adaptive histogram equalization processing on the denoised image; S25. By calculating the gray histogram of the denoised image, determine the cumulative distribution function of the local area; S26. Based on the cumulative distribution function, adaptively adjust the gray mapping relationship of the local area to obtain an enhanced image.

4. A light-intelligent orchard field navigation method according to claim 3, characterized in that, In step S2, image segmentation includes the following steps: S27. Segment the enhanced image into multiple superpixels by the superpixel segmentation algorithm; S28. Determine the similarity matrix between superpixels as the edge weights of an undirected graph, and construct a superpixel graph cut undirected graph; S29. Combine the preset probability distribution of the text area, and solve the global optimal solution of the superpixel graph cut undirected graph by the maximum flow minimum cut algorithm; S210. Obtain the segmentation result of the fruit tree area to obtain the orchard topographic map after image preprocessing.

5. The method for light intelligent orchard field navigation according to claim 4, wherein Step S3 includes the following sub-steps: S31. Connect the diagonal points in the orchard topographic map after image preprocessing to obtain two diagonals; S32. Based on the intersection part of the two diagonals, determine the midpoint of the topographic map; S33. Take the midpoint of the topographic map as the origin of the coordinate system, and establish a quadrant coordinate system in the orchard topographic map after image preprocessing; S34. Analyze the position of each fruit tree in the orchard topographic map to obtain the horizontal distance and vertical distance of each fruit tree from the origin; S35. Based on the horizontal distance and vertical distance, determine the coordinate information of each fruit tree.

6. The light intelligent orchard field navigation method according to claim 5, wherein Step S4 includes the following sub-steps: S41. Retrieve each historical driving path information of the robot from the database, and perform English and digital labeling A1, A2,.......An; S42. Analyze each piece of historical driving path information to determine the inflection point information of the robot; S43. Numerically label each inflection point information in each piece of historical driving path information as A1-1, A1-2,.......A1-n; S44. Based on the path information label and the inflection point information label in each path, form a set to constitute the historical motion data set of the robot.

7. A lightweight intelligent orchard field navigation method according to claim 6, characterized in that, Step S5 includes the following sub-steps: S51. Construct a 3D map according to the pre-processed orchard topographic map; S52. Map the position of each fruit tree in the 3D map according to the coordinate information of each fruit tree; S53. Incorporate the historical motion data set of the robot into the 3D map of the orchard in the form of dynamic simulation to construct a 3D orchard model.

8. A lightweight intelligent orchard field navigation method according to claim 4, characterized in that Step S6 includes the following sub-steps: S61. Based on the feedforward neural network, construct an optimal steering path state recognition model; S62. Based on machine learning, label and divide the multiple sample historical motion data sets to obtain a training set, a validation set, and a test set; S63. Use the training set, the validation set, and the test set to supervise the training, validation, and testing of the optimal steering path state recognition model to obtain the optimal steering path state recognition model with an accuracy rate meeting the preset accuracy rate requirements.

9. A light intelligent orchard field navigation method according to claim 8, characterized in that, In step S7, the start-end point similarity matching model is: Where S i is the similarity between the real-time start and end point data and the i-th feature of the historical best turning inflection point path of the robot, w ij is the j-th feature index value of the i-th feature of the real-time start and end point data, v ij is the j-th feature index value of the i-th feature of the historical best turning inflection point path, and n is the total number of the i-th feature index values.

10. A light-intelligent orchard field navigation method according to claim 9, characterized in that, In step S7, the inflection point matching model is: Where C is the fitness between the real-time origin-destination data and the historical best turning inflection point path, and α i is the weight value of the i-th characteristic parameter, and m is the total number of characteristic parameters.

Images