Agricultural robot navigation path detection method for inter-ridge planting

Abstract

The invention discloses an agricultural robot navigation path detection method for inter-ridge planting, and relates to the technical field of unmanned working vehicles. The method comprises the following steps: S1, acquiring an image between ridges in real time through an image acquisition device; s2, a VPDXYOLOv8n model is adopted for image processing, prediction frames containing transverse crop rows are generated, and each prediction frame corresponds to crop areas on the two sides and a middle driving area; s3, extracting geometric midpoints of the prediction frames as a navigation positioning point set; s4, performing curve fitting on the navigation positioning point set by applying a recursive least square method to generate an optimal navigation path, reducing the parameter quantity and the calculation quantity of a model by adopting a VPDXYOLOv8n model and using convolution such as PDWConv and DSConv, and increasing the performance of multi-scale feature detection; a navigation positioning point is directly obtained by directly detecting a transverse crop row, so that a navigation path extraction step is simplified; and the recursive least square method is adopted to perform curve fitting on the navigation positioning point set, so that interference points can be effectively eliminated, and the robustness is high.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned operation vehicles, and particularly to a navigation path detection method for an agricultural robot for inter-ridge planting. Background Art

[0002] Intelligent robots are gradually applied to various industries. In agriculture, agricultural robots can efficiently replace manual labor to achieve various tasks such as precise fertilization and watering, weed identification and control, and pest and disease identification and detection. Among the various technologies that make up agricultural robots, navigation technology is extremely important. Currently, the mainstream navigation methods include the Global Positioning System (GPS) and lidar. However, GPS navigation may have inaccurate positioning in a certain range of farmland areas, and lidar is costly and difficult to identify some low-growing crops. Visual navigation is a technology that collects surrounding environmental information through a camera or RGB camera and extracts features in the environment to achieve positioning and movement. Among them, the extraction of the navigation path line is an important prerequisite.

[0003] One type of method for extracting the center line in crop navigation is to utilize the difference in color features, adopt the super-green feature factor method and threshold segmentation to segment the crops, thereby determining the driving area of the agricultural robot, and then further determining the feature points and fitting the navigation line. This traditional image processing method not only has many steps but is also extremely sensitive to light changes, making it difficult to extract the navigation line in a complex environment. Another type is to introduce deep learning methods, using object detection or semantic segmentation models to detect the navigation line. However, some deep learning network models have a large amount of computation and poor real-time performance.

[0004] For example, Chinese Patent CN110243372B discloses an intelligent agricultural machinery navigation system and method based on machine vision, including: an image sensing module that acquires a farmland image on the navigation path; an image processing module that preprocesses the farmland image, extracts the seedling line and the navigation line from the farmland image, calculates the lateral deviation value and the heading angle deviation value between the agricultural machinery and the navigation line, and judges the reliability of the calculation results; an agricultural machinery path planning module that plans the field operation path according to the navigation line; and an agricultural machinery steering system that calculates the correct steering wheel angle of the agricultural machinery according to the field operation path planning and controls the steering of the steering wheel of the agricultural machinery. However, in image processing, it is necessary to perform binary segmentation on farmland crops and the background. When it comes to complex environments with large crop density and many weeds, segmenting each crop will have inadaptability. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a navigation path detection method for an agricultural robot for inter-ridge planting.

[0006] To achieve the above object, the present invention provides the following technical solutions: An agricultural robot navigation path detection method for inter-ridge planting, comprising the following steps: S1: Real-time acquisition of images between ridges by an image acquisition device; S2: Using the VPDX_YOLOv8n model for image processing to generate prediction boxes containing horizontal crop rows, and each prediction box corresponds to the crop regions on both sides and the driving region in the middle; S3: Extracting the geometric midpoints of each prediction box as a navigation positioning point set; S4: Applying the recursive least squares method to perform curve fitting on the navigation positioning point set to generate an optimal navigation path.

[0007] The VPDX_YOLOv8n model includes: An input module for receiving feature maps; A backbone network composed of Conv modules and multiple cascaded VanillaBlock modules, where through the stacking of several VanillaBlock modules, the feature map size is gradually reduced, the number of channels is increased, and multi-scale features are obtained; A neck network using Conv modules and PC2f modules for image downsampling and feature extraction. On the basis of the FPN top-down path, a bottom-up path is introduced for multi-scale feature fusion; A head network using DS_Detect.

[0008] The PC2f module includes: A PDWConv module for performing convolution operations on the input; A split function for segmentation: n cascaded bottleneck modules, and the output of each bottleneck module is used as the input of the next bottleneck module for performing per-channel convolution on partial features of the input; A Concat function for splicing all branches; And a Conv module for compressing the spliced feature map.

[0009] Each bottleneck module contains a PConv module and a DWConv module.

[0010] The DS_Detect includes two branches, and the front section of each branch is a cascaded Conv module and DSConv module.

[0011] The DSConv module consists of a DWConv module for performing independent convolution operations on each channel of the input features and a PWConv module for converting the number of output feature channels of the DWConv module to the target number of feature channels using a 1×1 convolution kernel.

[0012] Introduce the XIoU loss function to solve the aspect ratio penalty term υ for. .

[0013] In S4, S41 Randomly select two navigation positioning points to determine a straight line; S42 Calculate the distance from each navigation positioning point to the straight line x ; S43 By comparing all x with the set threshold w , if there is one x < w , then the vote count T is incremented by 1, and the maximum value of T is the number of navigation positioning points; S44 Perform iterative processing for a set number of times. If T≥8, stop the iteration and select all x <w navigation positioning points and use the least squares method to fit the selected navigation positioning points; if T<8, when the vote count is the largest, select all x < w navigation positioning points and use the least squares method for fitting.

[0014] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned navigation path detection method for an agricultural robot for inter-row planting.

[0015] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned navigation path detection method for an agricultural robot for inter-row planting.

[0016] Advantages of the present invention: 1. Disclosed is a lightweight VPDX_YOLOv8n model with good multi-scale feature extraction. Using convolutions such as PDWConv and DSConv reduces the number of model parameters and computational complexity, while improving the performance of multi-scale feature detection; 2. Directly obtain navigation positioning points by detecting the form of horizontal crop rows, simplifying the navigation path extraction steps; 3. Using the recursive least squares method to perform curve fitting on the set of navigation positioning points can effectively exclude interference points and has strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of the navigation path detection method disclosed by the present invention.

[0018] Figure 2 It is a prediction map of the lateral crop rows of cantaloupe seedlings.

[0019] Figure 3 It is a schematic diagram of the network structure of the VPDX_YOLOv8n model disclosed by the present invention.

[0020] Figure 4 It is a schematic diagram of the structure of the PDC2f module disclosed by the present invention.

[0021] Figure 5 It is a structural diagram of the PDWConv disclosed by the present invention.

[0022] Figure 6 It is the structure of the DS_Detect detection head disclosed by the present invention.

[0023] Figure 7 It is a prediction map of the lateral crop rows.

[0024] Figure 8 It is the RLS fitting algorithm process.

[0025] Figure 9 It is the navigation center line after RLS fitting, where the dashed line is the least squares method, the dotted line is the random sample consensus algorithm, and the solid line is the RLC fitting algorithm.

[0026] Figure 10 It is the input image of the VPDX_YOLOv8n model. Specific implementation manners

[0027] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts belong to the scope of protection of the present invention.

[0028] It should be noted that all the directional indications (such as up, down, left, right, front, back...) in the embodiments of the present invention are only used to explain the relative position relationship and movement conditions between components in a specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications will also change accordingly.

[0029] As Figure 1 shown, the present invention discloses an agricultural robot navigation path detection method for inter-ridge planting, which includes the following steps: S1 obtains the images between ridges in real time through an image acquisition device; S2 uses the VPDX_YOLOv8n model for image processing to generate prediction boxes containing horizontal crop rows, and each prediction box corresponds to the crop areas on both sides and the driving area in the middle; S3 extracts the geometric midpoints of each prediction box as the navigation positioning point set; S4 applies the recursive least squares (RLS) method to perform curve fitting on the navigation positioning point set to generate the optimal navigation path.

[0030] The image acquisition device is placed on the tracked robot, and the distance from the ground is 1.2 m. The image acquisition device can be a camera or a video camera, and the images collected between ridges are transmitted to the host computer in real time for processing. The main contents include: obtaining the prediction boxes of the images through the VPDX_YOLOv8n model, and each prediction box includes a horizontal crop row (the crops on both sides and the driving area in the middle), as Figure 2 shown, then taking the midpoint of each prediction box as the navigation positioning point, using the RLS algorithm to fit the navigation path, and finally transmitting the detected navigation path information to the lower computer.

[0031] This application mainly includes two major parts, namely the VPDX_YOLOv8n detection model and the RLS navigation path fitting algorithm.

[0032] The existing YOLOv8n model has 3.01M parameters and 8.2G computational complexity. To meet the real-time requirements, the VPDX_YOLOv8n model is proposed based on the YOLOv8n model. The parameters of this model are 1.17, and the computational complexity is 3.7G. It is more lightweight than YOLOv8n, and at the same time, it increases the performance of multi-scale feature extraction. The average precision mAP50 in the detection of cantaloupe seedling crop rows has increased by 3.19%. In addition, the proposed method for detecting horizontal crop rows between ridges is more concise and efficient than the common method for detecting single plants.

[0033] Combining the Random Sample Consensus (RANSAC) algorithm with the Least Square method (LS), the RLS algorithm is proposed, which can effectively eliminate the interference points generated in actual situations, improve the path recognition accuracy, and has better robustness.

[0034] The network structure of the VPDX_YOLOv8n model is as Figure 3 shown. To meet the real-time and multi-scale feature detection requirements during the operation of agricultural robots, improvements are made in four aspects: the backbone network, C2f of the neck network, the head network, and the loss function.

[0035] The VPDX_YOLOv8n model includes: an input module, a backbone network, a neck network, and a head network.

[0036] The input module is used to receive the feature map. In this embodiment, the image size is 640×640. The backbone network is composed of a Conv module and multiple cascaded VanillaBlock modules. By stacking several VanillaBlock modules, the size of the feature map is gradually reduced, the number of channels is increased, and multi-scale features are obtained. Using the VanillaBlock module instead of the backbone network, that is, the backbone network of the VPDX_YOLOv8n model is composed of a Conv module and the lightweight VanillaBlock module which is good at extracting multi-scale features, as Figure 3 shown. The convolution kernel of the Conv module is 4×4, while the convolution kernels of the 3 VanillaBlock modules are 3×3. Due to its simple and efficient design and the further aggregation of features by MaxPool2d in the VanillaBlock module, in this paper, by stacking the VanillaBlock modules, the size of the input feature map is continuously reduced to obtain a larger receptive field, and the number of channels is continuously increased to obtain richer feature information, improving the detection ability for different scale features and reducing the model parameters and computational complexity.

[0037] The neck network uses a Conv module and a PC2f module for image downsampling and feature extraction. Based on the top-down path of the FPN, a bottom-up path is introduced for multi-scale feature fusion. The head network adopts DS_Detect.

[0038] The PC2f module includes: A PDWConv module for performing convolution operations on the input; A split function for segmentation; n cascaded bottleneck modules, and the output of each bottleneck module is used as the input of the next bottleneck module for performing per-channel convolution on part of the features of the input; A Concat function for splicing all branches; And a Conv module for compressing the spliced feature map.

[0039] Each bottleneck module contains a PConv module and a DWConv module.

[0040] That is, the PC2f module is designed to replace the original four C2f (CSP Bottleneck with 2 Convolutions) structures in the neck, and its structure is as follows Figure 4 shown. The input features first pass through a Partial Depthwise Convolution (PDWConv), and then enter n bottleneck modules (PD_Bottleneck). Each PD_Bottleneck contains a Partial convolution (PConv) as follows Figure 5 shown, and a Depthwise Convolution (DWConv). The output features of multiple PD_Bottleneck modules are concatenated to fuse multi-scale information, and finally a standard convolution (Conv) is used to compress the concatenated feature map to output a feature map with the target number of channels.

[0041] The structure of PDWConv is as follows Figure 5 shown, which is a combination of PConv and DWConv, and performs per-channel convolution on the input partial features. Since the feature maps in the neck have been extracted and fused multiple times, the features of some channels may have become redundant, while partial depth convolution can better focus on important feature regions, reduce the interference of redundant information, and inherit the channel independence of depth convolution, enabling the extraction of richer channel features. In this paper, the number of parameters and computational amount of PDWConv are , where is the number of output channels.

[0042] The DS_Detect described above includes two branches, and the front section of each branch is a cascaded Conv module and DSConv module.

[0043] The DSConv module consists of a DWConv module for performing independent convolution operations on each channel of the input features and a PWConv module for converting the number of output feature channels of the DWConv module to the target number of feature channels using a 1×1 convolution kernel.

[0044] For further weight reduction, in this paper, the second Conv of the detection head is replaced with a depthwise separable convolution (DSConv), which consists of a DWConv and a pointwise convolution (PWConv). The DWConv can perform independent convolution operations on each channel of the input features, and the number of convolution kernels is the same as the number of channels of the input features. The PWConv uses a 1×1 convolution kernel to convert the number of output feature channels of the DWConv into the number of target feature channels. Since the DSConv can efficiently perform inter-channel fusion on the basis of existing rich feature information, it is placed after the standard convolution. Such a design not only effectively reduces the computational complexity and the number of parameters, but also maintains good detection performance. The structure of the improved DS_Detect detection head is as Figure 6 shown.

[0045] To further improve the detection accuracy of the model without increasing the model complexity, the XIoU loss function is used to replace the loss function of the Yolov8n model. The CIoU loss function is used by default in the Yolov8n model. This loss function takes into account three set parameters: the overlapping area, the distance between the center points, and the aspect ratio. The calculation formula of the CIoU bounding box loss function is as follows: (1) (2) (3) In the formula IoU represents the intersection over union between the predicted bounding box and the ground truth bounding box, is the distance between the center point of the predicted bounding box and the center point of the ground truth bounding box, is the coordinate of the center point of the predicted bounding box, is the coordinate of the center point of the ground truth bounding box, is the aspect ratio penalty coefficient, υ is the aspect ratio penalty term.

[0046] However, the aspect ratio penalty term υ in the CIoU loss function formula is obtained using the arctangent function, which results in υ having weak robustness, being sensitive to outliers, and being greatly affected by outliers, leading to large fluctuations in the value of the loss function, affecting the performance of the loss function. Secondly, due to the limitations of the arctangent function itself, the value range cannot meet the normalization requirements of the loss function. Therefore, the XIoU loss function is introduced. Compared with the CIoU, the XIoU loss function performs a new solution for the aspect ratio penalty term υ , and υ The calculation formula of is as follows: (4) Most crops are planted in the inter-row manner, that is, crop seeds are sown in rows at a certain spacing. The common method is to detect single plants on both sides of the driving area of the agricultural robot, and then find the navigation midpoint of the middle area based on the positioning points on both sides. However, this patent adopts the form of directly detecting the horizontal crop rows, that is, including the crops on both sides and the driving area in the middle. The number of prediction boxes in this way is half of the number of prediction boxes for detecting crops separately. At the same time, since the horizontal crop rows determine the scope of the driving area, directly taking the center point of the prediction box as the navigation midpoint saves the steps of first determining both sides and then determining the navigation midpoint, and can better adapt to the complex situation with many weeds, improving the detection efficiency.

[0047] Taking the seedlings of Hami melons as an example in this patent, as Figure 7 shown, use VPDX_YOLOv8n to predict the horizontal crop rows.

[0048] In actual situations, misdetection may occur, resulting in the navigation positioning point appearing outside the navigation area, that is, interference points. For the least squares method, since all points need to participate in the fitting, the fitted navigation line is verified to deviate from the navigation area ( Figure 9 dotted line). For the random sample consensus algorithm, it can exclude interference points, but due to the limited range collected by the camera, only 5 - 10 horizontal crop rows can be predicted. There is still a certain error in selecting two points to determine the straight line after excluding interference points ( Figure 9 dashed-dotted line). Therefore, in this method, the two are combined to propose the RLS navigation midline fitting algorithm ( Figure 9 solid line), and the algorithm flow is as Figure 8 shown. The specific method is as follows: S41 Randomly select two navigation positioning points to determine a straight line; S42 Calculate the distance from each navigation positioning point to the straight line x ; S43 By comparing all x with the set threshold w , if there is one x < w , then the vote count T is incremented by 1, and the maximum value of T is the number of navigation positioning points; S44 Perform iterative processing for a set number of times. If T≥8, stop the iteration, select all x <w navigation positioning points, and use the least squares method to fit the selected navigation positioning points; if T<8, when the vote count is the largest, select all the navigation positioning points that meet x < w , and use the least squares method for fitting.

[0049] Assume 100 iterations. If during the iteration process, T ≥ 8, directly jump out of the iteration process and select all x navigation positioning points less than w, and use the least squares method to fit the selected navigation positioning points. If during the iteration process, T < 8, select all x navigation positioning points less than w when T is at its maximum, and fit them. If T has the same maximum value multiple times, select all x navigation positioning points less than w when T first reaches its maximum value, and fit them.

[0050] The loss function in the VPDX_YOLOv8n model serves to measure the prediction results and the true labels, mainly acting during training, including bounding box regression loss, class classification loss, and confidence loss. Among them, the bounding box regression loss is used to optimize the positional relationship between the predicted box and the true box, and its position directly affects the subsequent extraction of the navigation path line. In this patent, XIoU is used to replace the original CIoU. Specific embodiments The input is an inter-row image with a resolution of 640×480. Taking the inter-row image of cantaloupe seedlings as an example in this patent, as Figure 10 shown. The output is an inter-row image with a resolution of 640×480 and with prediction box information, as Figure 2 shown.

[0052] The specific prediction process of the VPDX_YOLOv8n model is similar to that of the original YOLOv8n model. However, due to some structural improvements in the VPDX_YOLOv8n model, the model is more lightweight, has a faster prediction speed when predicting horizontal crop rows, and maintains the prediction accuracy. The specific prediction process is as Figure 3 shown.

[0053] The overall network structure is divided into three parts. The backbone network is used for feature extraction, accounting for the largest proportion in the model. The neck network is responsible for multi-scale feature fusion, and the head network is used to generate the output of object detection.

[0054] First, the inter-row image collected by the visual sensor is transmitted to the host computer in real time. The image input to the VPDX_YOLOv8n model is 640×640×3, that is, with a resolution of 640×640 and 3 channels of RGB.

[0055] Layers 0 to 3 are the backbone network. Layer 0 is a normal convolution (Conv) with a kernel size of 4×4. After convolution in Layer 0, the image becomes 160×160×32, that is, the image resolution is 160×160 and the number of channels is 32. Layers 1 - 3 are VanillaBlock modules with a kernel size of 3×3. After Layer 1, the image sizes become 80×80×64, 40×40×128, and 20×20×256 in sequence. In this part of the backbone network, the size of the input feature map continuously decreases to obtain a larger receptive field, and the number of channels continuously increases to obtain richer feature information, improving the detection ability for different-scale features. Compared with the original YOLOv8n backbone network, the backbone network of the improved model uses only 4 layers to replace the original 10 layers, greatly reducing the computational amount and the number of parameters of the model. At the same time, due to the stacking of VanillaBlock modules, the detection ability for horizontal crop rows of different scales is improved.

[0056] Layers 4 to 15 are the neck of the model, responsible for feature fusion, combining multi-scale features extracted by the backbone to generate richer feature representations. Layer 4 is an upsampling layer that doubles the resolution of the feature map, that is, it becomes 40×40×256. Then Layer 5 is a concatenation layer. By fusing feature maps at different levels, the detection effect for multi-scale targets is improved. The output of Layer 2 with a feature map of 40×40 and the output of Layer 4 are concatenated together, and the feature map becomes 40×40×384. Layer 6 is the PC2f layer designed in the present invention. Compared with the original C2f layer, convolutions such as PDWConv are introduced. Since the feature maps in the neck are extracted and fused multiple times, the features of some channels may have become redundant, while some depth convolutions can better focus on important feature regions, reducing the interference of redundant information. At the same time, it inherits the channel independence of depth convolutions and can extract richer channel features. The output feature map is 40×40×128. Layers 7 - 9 repeat the operations of Layers 4 - 6. After convolution in Layer 7, the feature map is 80×80×128. After concatenation in Layer 8, the feature map is 80×80×192. After convolution in Layer 9, the feature map is 80×80×64. Layer 10 is a normal convolution layer with a kernel size of 3×3. After convolution, the feature map is 40×40×64. Then Layer 11 performs concatenation, and the output feature map is 40×40×192. Layer 12 is a PC2f layer, and the output feature map is 40×40×128. Layers 13 - 15 repeat the operations of Layers 10 - 12, and the sequentially output feature maps are 20×20×128, 20×20×384, and 20×20×256.

[0057] The last part is the head. The VPDX_YOLOv8n model inherits the 3 detection heads of the YOLOv8n model. However, in the present invention, a lightweight DS_Detect detection head is designed to replace the original Detect detection head. The role of the detection head is to be responsible for generating the final detection effect, such as predicting information such as bounding boxes, confidence levels, and detection categories. The 3 DS_Detect detection heads respectively input the image information after convolution of the 9th, 12th, and 15th layers. Through these feature maps of three different scales, different-sized targets can be better predicted.

[0058] The present invention also provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the above-mentioned navigation path detection method for an agricultural robot for inter-ridge planting is implemented.

[0059] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned navigation path detection method for an agricultural robot for inter-ridge planting is implemented.

[0060] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, Digital Subscriber Line (DSL)) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a Digital Video Disc (DVD)), or a semiconductor medium (such as a Solid State Disk (SSD)), etc.

[0061] The embodiments should not be regarded as limitations to the present invention, but any improvements made based on the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. An agricultural robot navigation path detection method for inter-ridge planting, characterized in that: It includes the following steps: S1: Real-time obtain images between ridges by an image acquisition device; S2: Use the VPDX_YOLOv8n model for image processing to generate prediction boxes containing horizontal crop rows, and each prediction box corresponds to the crop areas on both sides and the driving area in the middle; S3: Extract the geometric midpoints of each prediction box as a set of navigation positioning points; S4: Apply the recursive least squares method to perform curve fitting on the set of navigation positioning points to generate an optimal navigation path.

2. The agricultural robot navigation path detection method for inter-ridge planting according to claim 1, characterized in that: The VPDX_YOLOv8n model includes: An input module for receiving a feature map; A backbone network composed of Conv modules and multiple cascaded VanillaBlock modules, where by stacking several VanillaBlock modules, the size of the feature map is gradually reduced, the number of channels is increased, and multi-scale features are obtained; A neck network that uses Conv modules and PC2f modules for image downsampling and feature extraction. Based on the top-down path of FPN, a bottom-up path is introduced for multi-scale feature fusion; A head network that uses DS_Detect.

3. A method for detecting the navigation path of an agricultural robot for inter-ridge planting according to claim 2, characterized in that: The PC2f module includes: A PDWConv module for performing convolution operations on the input; A split function for segmentation: n cascaded bottleneck modules, and the output of each bottleneck module is used as the input of the next bottleneck module for performing per-channel convolution on some features of the input; A Concat function for splicing all branches; And a Conv module for compressing the spliced feature map.

4. A method for detecting the navigation path of an agricultural robot for inter-ridge planting according to claim 3, characterized in that: Each bottleneck module contains a PConv module and a DWConv module.

5. The agricultural robot navigation path detection method for inter-ridge planting according to claim 2, wherein: The DS_Detect includes two branches, and the front section of each branch is a cascaded Conv module and DSConv module.

6. A method for detecting the navigation path of an agricultural robot for inter-ridge planting according to claim 5, characterized in that: The DSConv module consists of a DWConv module for performing independent convolution operations on each channel of the input features and a PWConv module for using a 1×1 convolution kernel to convert the output feature channels of the DWConv module into the target number of feature channels.

7. A method for detecting the navigation path of an agricultural robot for inter-ridge planting according to claim 2, characterized in that: Introduce the XIoU loss function to solve the penalty term for the aspect ratio υ . .

8. A method for detecting the navigation path of an agricultural robot for inter-ridge planting according to claim 1, characterized in that: In S4, S41: Randomly select two navigation positioning points to determine a straight line; S42 Calculate the distance from each navigation positioning point to the straight line x ; S43 compares all x with the set threshold w . If there is a x < w , the vote count T is incremented by 1, with T having a maximum value equal to the number of navigation positioning points. S44 performs iterative processing for a set number of times. If T≥8, stop the iteration, select all x navigation positioning points < w, and use the least squares method to fit the selected navigation positioning points; if T < 8, when the number of votes is the largest, all x < w navigation positioning points that meet the requirements, and use the least squares method for fitting.

9. A computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for detecting the navigation path of an agricultural robot for inter-ridge planting as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the method for detecting the navigation path of an agricultural robot for inter-ridge planting as described in any one of claims 1 to 8.

Images