An autonomous navigation inspection method for a mobile robot in a tea garden

Abstract

The present invention relates to an autonomous navigation inspection method for a mobile robot in a tea garden, comprising: obtaining tea garden road image information in two-dimensional space; constructing a target navigation path in three-dimensional space based on the tea garden road image information; the mobile robot in the tea garden moves based on the target navigation path, and monitors tea leaves in real time through disease visual sensing during the movement. The present invention can truly present complex operation scenes in tea gardens, more effectively improve the navigation accuracy and autonomous following stability of mobile robots, and at the same time can replace traditional manual operations, reduce labor costs, improve the detection rate and efficiency of tea diseases, ensure the quality of tea production, provide scientific and accurate protection measures for tea leaves, and further provide theoretical and practical support for improving the level of intelligence in rural areas and building rural smart agriculture.

Description

Technical Field

[0001] The invention relates to the technical field of tea disease detection, and in particular to an autonomous navigation inspection method of a mobile robot in a tea garden. Background Art

[0002] China is a producer and exporter of many teas in the world. It not only has a large planting area, but also has a high tea yield. With the annual growth of tea plantation area, tea plantation diseases are becoming more and more serious, which has a serious impact on tea yield and quality. At present, the traditional tea disease detection method mainly relies on manual work, which mainly has the following problems:

[0003] (1) Due to the large planting area and complex geographical environment of tea gardens, and the fact that tea leaves grow under the cover of weeds, branches and other obstructions, the manual labor method for detecting tea diseases is inefficient, has low detection accuracy, high labor costs, and takes a lot of time;

[0004] (2) On-site identification and diagnosis are based on the tea farmers’ planting experience, and the test results are highly subjective. Tea farmers take corresponding protective measures against tea diseases based on their personal experience, and tea cannot receive scientific and accurate treatment, resulting in poor protection effects;

[0005] (3) Hiring experienced plant protection experts to conduct field inspections is costly and has poor real-time performance, making it impossible to diagnose and prevent tea diseases in a timely manner.

[0006] The existing traditional tea disease detection method relies on a large amount of manual work and has a low degree of automation. For the production and development of the entire tea garden, there are problems such as high production costs, low accuracy of manual detection, and difficulty in stably ensuring tea quality. It is unable to provide scientific and accurate tea protection measures.

[0007] Since there are few patents on autonomous navigation and inspection of mobile machines in tea gardens, we studied the patents of application scenarios related to autonomous navigation and inspection and found the following:

[0008] 1. An autonomous navigation inspection robot based on a distributed framework. This patent mainly uses GPS / Beidou units to measure speed and position. When the signal is weak, it uses the visual odometer to detect obstacles before detection. When the light is strong or insufficient, it automatically switches to the laser sensor to measure and collect map data.

[0009] 2. A tea garden inspection system based on visual technology and inspection robots. This patent includes an acquisition device, a control device and a visual system. A supervised lightweight MobileNetV2 model is built, and the inspection system is built by combining the hardware STM32 and the software system.

[0010] 3. Visual navigation mobile robot based on deep learning. This patent includes a mobile vehicle body and a terminal device. Through the image processing module, the route is replanned at the location of the new obstacle to control the mobile vehicle body, so that the mobile robot can plan the route more scientifically and reasonably.

[0011] 4. A visual navigation method for an indoor mobile robot. This patent mainly uses a laser radar to scan the environment to obtain point cloud information, segment the point cloud, use the RANSAC algorithm for three-dimensional fitting, use YOLOv5 to enhance semantic information, and finally perform path planning.

[0012] 5. A field autonomous navigation inspection and intelligent mist spraying device. This patent provides an autonomous navigation driving system that uses a ranging algorithm to provide three-dimensional space point cloud data and the reflectivity of surrounding objects and automatically plans the optimal path. It is used as an inspection and monitoring system for patrolling field environments and crop growth conditions.

[0013] Combined with the above-mentioned related patent research and analysis, the following problems are found:

[0014] (1) Currently, there are few autonomous navigation inspection methods for tea garden operation scenarios, and the research methods of related application scenarios have a narrow scope of application and are difficult to adapt to autonomous navigation in complex tea garden operation scenarios;

[0015] (2) Most of the related research methods are based on laser radar and GPS navigation. Although the laser radar method can scan the entire operation scene, it is expensive and has low universal applicability. In complex operation scenes, the GPS navigation method cannot stably receive signals due to crop occlusion and environmental interference. Summary of the invention

[0016] The purpose of the present invention is to provide an autonomous navigation inspection method for a tea garden mobile robot to solve the problems existing in the above-mentioned prior art.

[0017] To achieve the above object, the present invention provides the following solutions:

[0018] An autonomous navigation inspection method for a tea garden mobile robot, comprising:

[0019] Obtain tea garden road image information in two-dimensional space;

[0020] Based on the tea garden road image information, construct a target navigation path in three-dimensional space;

[0021] The tea garden mobile robot moves based on the target navigation path, and monitors the tea leaves in real time through disease visual sensing during the movement.

[0022] Optionally, acquiring the tea garden road image information in two-dimensional space includes:

[0023] Collect tea bush image data in tea gardens;

[0024] Based on the tea bush image of the tea garden, construct a tea garden navigation path VOC dataset;

[0025] Based on the tea garden navigation path VOC dataset, the improved Deeplabv3Plus model is trained and tested to obtain a road segmentation model;

[0026] Based on the road segmentation model, the tea garden road image information in two-dimensional space is acquired.

[0027] Optionally, collecting tea bush image data in a tea garden includes:

[0028] The tea bushes in the tea garden are selected as the preset navigation path of the mobile robot in the tea garden. The visual sensor is installed on the mobile robot in the tea garden to collect video data of the tea bushes in the tea garden in different time periods and weather environments.

[0029] The tea garden tea bush image data is extracted from the collected tea garden tea bush video data.

[0030] Optionally, improve the Deeplabv3Plus model by:

[0031] The lightweight Mobilenetv2 network is used to replace the original Xecption network as the backbone network of the network model; a separate feature processing branch is added to the ASPP module, and the ECA attention mechanism is introduced in the branch to extract features. At the same time, the CBAM attention mechanism is introduced after the original ASPP module; in addition, the hole convolution method of the ASPP module is replaced with a depth-wise separable convolution method; finally, in the decoding layer, the feature information of the 1st, 3rd, and 6th bottleneck layers of the shallow features are fused, and the CBAM attention mechanism is introduced to allocate feature weights, which are finally fused with the deep features to realize the detection of complex roads in the tea garden.

[0032] Optionally, constructing the target navigation path in the three-dimensional space includes:

[0033] The tea garden road image information is combined with the depth image information and projected into a three-dimensional space to form three-dimensional point cloud information;

[0034] Preprocessing the three-dimensional point cloud information;

[0035] Based on the RANSAC algorithm, the preprocessed three-dimensional point cloud information is fitted, three points are randomly selected as the minimum sample set, the plane is determined, the distance from the point to the plane is calculated to determine the inner point, and the judgment is continuously iterated by setting a threshold. Finally, the plane whose inner point value is greater than the preset threshold is selected as the fitting result to obtain the target navigation path of the mobile robot in the three-dimensional space.

[0036] Optionally, preprocessing the three-dimensional point cloud information includes:

[0037] The three-dimensional point cloud information is filtered by using bilateral filtering, and then the point cloud is downsampled by using a random downsampling method.

[0038] Optionally, the tea garden mobile robot moves based on the target navigation path including:

[0039] The target navigation path is transmitted to the ROS operating system of the tea garden mobile robot to obtain the current position information of the tea garden mobile robot;

[0040] The adjustment mode of the tea garden mobile robot is preset, and the adjustment mode includes: fine-tuning steering mode and in-situ steering mode; when the steering angle of the tea garden mobile robot is less than or equal to 15°, the fine-tuning steering mode is adopted, otherwise the in-situ steering mode is adopted;

[0041] The tea garden mobile robot selects the target point (P x ,P y ), by defining the motor output speed of the left and right wheels and Depend on and Where r is the radius of the left and right wheels, V r is the linear velocity of the right wheel, V l For the left and right linear speeds, get the current speed of the mobile robot and angular velocity Based on the turning radius of the robot When V r =-V l When L is the wheelbase, the turning radius of the robot is R=0, and the tea garden mobile robot adopts the in-situ steering method. The robot turns to the same straight line as the target point and continues to move in a straight line. When the steering angle is less than or equal to 15°, the tea garden mobile robot adopts the fine-tuning steering method and adjusts the body posture in combination with the turning radius of the robot.

[0042] Set the current speed V of the mobile robot c and angular velocity W c The input is sent to the control system of the mobile robot, and the mobile robot controls the turning radius and steering angle to achieve autonomous following of the target navigation path of the mobile robot in the tea garden.

[0043] Optionally, real-time monitoring of tea leaves includes:

[0044] Collect tea disease image data in different time periods and weather environments, preprocess and annotate the images, and establish a tea disease VOC dataset;

[0045] Using the tea disease voc dataset, the improved YOLOv5 model is improved to obtain a tea disease detection model;

[0046] The tea disease detection model is used to perform real-time detection of tea diseases, output disease detection frames and categories, and report abnormalities.

[0047] Optionally, improving the YOLOv5 model includes:

[0048] Firstly, the CBAM attention mechanism is introduced before the Neck network and the Head network; then, the SPPF module is used to replace the original SPP spatial pyramid pooling as the backbone network of the network model, and serial MaxPool processing and fused convolution are adopted; secondly, the CA attention mechanism is added after each layer of CSP processing of the backbone network; finally, the multi-scale FPN feature network is used to replace the original neck network.

[0049] The beneficial effects of the present invention are:

[0050] The traditional manual operation method of detecting tea diseases is inefficient and costly. In addition, the existing methods focus more on the path planning and path following problems of flat sections or single obstacles, ignoring the terrain and obstacle randomness of the tea garden environment, resulting in the inability of mobile machines to perform autonomous navigation operations efficiently and stably. The present invention obtains two-dimensional tea garden navigation path image information through the improved Deeplabv3Plus algorithm, projects the two-dimensional space into the three-dimensional space, and presents a more realistic tea garden operation environment through the visual conversion of the space; finally, the target navigation path is fitted and autonomous following is achieved to realize the autonomous navigation of the mobile robot. At the same time, on the basis of autonomous navigation, the improved YOLOv5 algorithm is used to monitor tea diseases in real time. This method can truly present the complex operation scenes of the tea garden, more effectively improve the navigation accuracy of the mobile robot and the stability of autonomous following, and at the same time can replace traditional manual operations, reduce labor costs, improve the detection rate and efficiency of tea diseases, ensure the quality of tea production, provide scientific and accurate protection measures for tea, and further provide theoretical and practical support for improving the level of rural intelligence and building rural smart agriculture. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0052] Figure 1 This is a schematic diagram of the autonomous navigation inspection structure of a tea garden mobile robot according to an embodiment of the present invention;

[0053] Figure 2 This is a framework diagram of the autonomous navigation inspection method for a tea garden mobile robot according to an embodiment of the present invention;

[0054] Figure 3 A schematic diagram of an image of a tea garden navigation path in a two-dimensional space obtained according to an embodiment of the present invention;

[0055] Figure 4 A block diagram of a target path autonomous following technology in a three-dimensional space according to an embodiment of the present invention;

[0056] Figure 5 A schematic diagram of real-time detection of tea diseases according to an embodiment of the present invention;

[0057] Figure 6 A schematic diagram of the process flow of the autonomous following method of a mobile robot in a tea garden according to an embodiment of the present invention;

[0058] Figure 7 Schematic diagram of the original Deeplabv3Plus network structure of an embodiment of the present invention;

[0059] Figure 8 Schematic diagram of the improved Deeplabv3Plus network structure of an embodiment of the present invention;

[0060] Fig. 9 Schematic diagram of the original YOLOv5 network structure of an embodiment of the present invention;

[0061] Fig.10 This is a schematic diagram of the improved YOLOv5 network structure of an embodiment of the present invention;

[0062] Fig.11 A schematic diagram of maximum pooling layer calculation in an embodiment of the present invention;

[0063] Fig.12 A schematic diagram of an autonomous following model of a target path according to an embodiment of the present invention;

[0064] Among them, 1. Navigation visual sensor; 2. Disease visual sensor; 3. Power box; 4. Controller; 5. Driving wheel; 6. Track. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0067] This embodiment proposes a new automated inspection method for the existing tea garden production management method, improves the accuracy of tea disease detection, provides scientific and accurate tea protection measures, ensures the production quality of tea, realizes automated tea garden inspection, and further provides theoretical and practical support for building smart agriculture in rural areas and improving the level of intelligent agricultural production. The design is as follows:

[0068] (1) Using a mobile robot in a tea garden as a carrier, installing navigation vision sensors and disease vision sensors, and adopting autonomous navigation inspection operations to replace traditional manual management methods and tea disease detection methods;

[0069] (2) The semantic segmentation method of visual navigation is used, combined with 3D point cloud information, to present the tea garden working environment more realistically, fit the autonomous navigation target path, provide a stable and efficient target navigation path, and realize the autonomous navigation function of the tea garden mobile robot;

[0070] (3) A deep learning target detection algorithm is used to perform real-time detection of tea diseases, output disease types and report abnormalities, improve the shortcomings of traditional manual disease detection, provide scientific tea protection measures, realize the real-time disease detection function of the mobile robot, and finally complete the autonomous inspection task of the mobile robot in the tea garden.

[0071] The workflow of this embodiment is:

[0072] The tea garden mobile robot is mainly composed of a navigation vision sensor 1, a disease vision sensor 2, a power box 3, a controller 4, a driving wheel 5 and a crawler 6 (such as Figure 1As shown). The designed autonomous navigation inspection method of the tea garden mobile robot mainly includes four steps, namely: obtaining tea garden road image information in two-dimensional space, constructing a target navigation path in three-dimensional space, autonomous following of the target navigation path, and real-time detection of tea diseases. When the tea garden mobile robot is operating, the two sides of the tracked chassis are on both sides of the tea bush. The ROS system is used as the operating system of the tea garden mobile robot. The navigation vision sensor is combined to obtain the target path image, and the active wheel is driven by the motor to drive the track movement to drive the tea garden mobile robot forward. During the advancement of the tea garden mobile robot, the disease visual sensor monitors the tea leaves in real time. When tea diseases are detected, the tea garden mobile robot system will report the disease category and provide scientific and accurate protective measures (such as Figure 2 As shown), it can ensure the production quality of tea, and at the same time provide scientific theoretical and practical support for building smart agriculture in rural areas and improving the level of intelligent agricultural production.

[0073] This embodiment provides an autonomous navigation inspection method for a tea garden mobile robot, including:

[0074] Obtain tea garden road image information in two-dimensional space;

[0075] Based on the tea garden road image information, construct a target navigation path in three-dimensional space;

[0076] The tea garden mobile robot moves based on the target navigation path, and monitors the tea leaves in real time through disease visual sensing during the movement.

[0077] Furthermore, obtaining the tea garden road image information in two-dimensional space includes:

[0078] Collect tea bush image data in tea gardens;

[0079] Based on the tea bush image of the tea garden, construct a tea garden navigation path VOC dataset;

[0080] Based on the tea garden navigation path VOC dataset, the improved Deeplabv3Plus model is trained and tested to obtain a road segmentation model;

[0081] Based on the road segmentation model, the tea garden road image information in two-dimensional space is acquired.

[0082] Further improvements to the Deeplabv3Plus model include:

[0083] The lightweight Mobilenetv2 network is used to replace the original Xecption network as the backbone network of the network model; a separate feature processing branch is added to the ASPP module, and the ECA attention mechanism is introduced in the branch to extract features. At the same time, the CBAM attention mechanism is introduced after the original ASPP module; in addition, the hole convolution method of the ASPP module is replaced with a depth-wise separable convolution method; finally, in the decoding layer, the feature information of the 1st, 3rd, and 6th bottleneck layers of the shallow features are fused, and the CBAM attention mechanism is introduced to allocate feature weights, which are finally fused with the deep features to realize the detection of complex roads in the tea garden.

[0084] Specifically, in this embodiment, Figure 3 As shown, the specific workflow for obtaining the tea garden road image information in two-dimensional space is as follows:

[0085] 1. The tea bushes in the tea garden were selected as the navigation path of the mobile robot in the tea garden. The visual sensor was installed on the mobile robot in the tea garden to collect videos of tea bushes in different time periods and weather environments. The rosbag function of the ROS system was used to extract tea bush images from the collected videos in the form of every 2 frames. The Labelme software was used to preprocess and annotate the images, and a tea garden navigation path VOC dataset was established.

[0086] 2. The lightweight Mobilenetv2 network is used to replace the original Xecption network with larger parameters as the backbone network of the network model. The backbone network is divided into 17 bottleneck layers and outputs two feature layer branches. The shallow features are output in the 1st, 3rd and 6th bottleneck layers respectively, and the deep features are output in the 17th bottleneck layer. At the same time, the lightweight network model can reduce the number of parameters of the model operation and speed up the model operation.

[0087] 3. The deep features include 2 branches, one directly enters the multi-scale fusion ASPP network structure to perform deep separable convolutions, where ASPP includes 5 branches, consisting of 1x1 convolution, 3x3 convolutions with expansion rates of 6, 12, and 18, and pooling layers. After being processed by the 3x3 deep separable convolution with an expansion rate of 6, it is divided into two branches, one enters the deep separable convolution with an expansion rate of 12, and the other branch is directly output; the convolution layer with an expansion rate of 16 will also generate a branch that enters the convolution layer with an expansion rate of 18, which can fully and richly extract features from the image information, and then enter the CBAM attention mechanism for weight allocation, improving the segmentation ability of the network; the other branch directly enters the ECA attention mechanism for feature extraction, which can better retain deep feature information and weight allocation, improve the effect of feature recognition, and then merge with the branch processed by the CBAM attention mechanism;

[0088] 4. After the 1st, 3rd, and 6th bottleneck layers of the shallow features are processed, they are fused into the CBAM attention mechanism for feature extraction and fused with the processed deep features to improve the correlation of information in different receptive fields and the ability of feature extraction;

[0089] 5. The tea garden navigation path voc dataset is randomly divided into a training set and a validation set in a ratio of 9:1. At the same time, the image size of the tea garden navigation path voc dataset is adjusted to 512x512 pixels. The improved Deeplabv3Plus algorithm is used to train and test the network model. The input image is processed by Mobilenetv2, and enters each bottleneck layer layer by layer for corresponding convolution operations. After completion, the deep features are output directly into the ASPP structure for deep separable convolution operations, and the three shallow features are output for fusion and CBAM attention mechanism weight allocation. Finally, they are fused with the deep features for 3x3 convolution operations, and the complex scenes of the tea garden are semantically segmented to obtain the tea garden navigation path image information in two-dimensional space, as shown in the attached figure. Figure 8 shown.

[0090] Further, constructing the target navigation path in the three-dimensional space includes:

[0091] The tea garden road image information is combined with the depth image information and projected into a three-dimensional space to form three-dimensional point cloud information;

[0092] Preprocessing the three-dimensional point cloud information;

[0093] The preprocessed three-dimensional point cloud information is fitted based on the RANSAC algorithm, three points are randomly selected as the minimum sample set, the plane is determined, the distance from the point to the plane is calculated to determine the inner point, and the judgment is continuously iterated by setting the threshold, and finally the one with more inner points is selected as the fitting result to obtain the target navigation path of the mobile robot in the three-dimensional space.

[0094] Specifically, in this embodiment, Figure 4 As shown, the specific workflow of constructing the target navigation path in three-dimensional space is as follows:

[0095] 1. Combine the acquired two-dimensional tea garden navigation path image information with the depth image information and project it into three-dimensional space to form dense three-dimensional point cloud information;

[0096] 2. Process the 3D point cloud information and use bilateral filtering to filter the point cloud. By considering the distance from the central pixel to the neighboring pixel, the point cloud can be smoothed to improve the quality of the point cloud information and better retain the edge information of the navigation path.

[0097] 3. Use a random downsampling method to downsample the point cloud, and perform random point downsampling operations by controlling the number of sampling points;

[0098] 4. Fit the point cloud of the navigation path based on the RANSAC algorithm, randomly select three points as the minimum sample set, determine the plane, calculate the distance from the point to the plane to determine the inner point, and continuously iterate the judgment by setting the threshold. Finally, the one with more inner points is selected as the fitting result to obtain the three-dimensional space target navigation path of the mobile robot.

[0099] Specifically, in this embodiment, the specific workflow of the tea garden mobile robot moving based on the target navigation path is as follows:

[0100] 1. Transmit the target navigation path to the ROS operating system of the tea garden mobile robot to obtain the current position information of the mobile robot;

[0101] 2. Combined with the characteristics of the tea garden operation scene, the tea garden mobile robot mainly moves in a straight line. When encountering obstacles or the width of the tea bushes is inconsistent, the tea garden mobile robot needs to adjust its own driving direction, which is mainly divided into fine-tuning steering and in-situ steering. When the robot's steering angle is less than or equal to 15°, the fine-tuning steering method is used; otherwise, the in-situ steering method is used;

[0102] 3. The robot selects the target point (P x ,P y ), by defining the motor output speed of the left and right wheels and Depend on and Where r is the radius of the left and right wheels, and the current speed of the mobile robot can be obtained and angular velocity The turning radius of the robot It can be seen that when V r =-V l When the turning radius R of the robot is 0, the tea garden mobile robot adopts the in-situ steering method. The robot turns to the same straight line as the target point and continues to move in a straight line. When the steering angle is less than or equal to 15°, the robot adopts the fine-tuning steering method and adjusts the body posture in combination with the turning radius of the robot. Fig.12 As shown;

[0103] 4. The calculated parameters are input into the control system of the mobile robot. The mobile robot controls the turning radius and steering angle to achieve autonomous following of the target navigation path of the mobile robot in the tea garden.

[0104] Furthermore, real-time monitoring of tea leaves includes:

[0105] Collect tea disease image data in different time periods and weather environments, preprocess and annotate the images, and establish a tea disease VOC dataset;

[0106] Using the tea disease voc dataset, the improved YOLOv5 model is improved to obtain a tea disease detection model;

[0107] The tea disease detection model is used to perform real-time detection of tea diseases, output disease detection frames and categories, and report abnormalities.

[0108] Further improvements to the YOLOv5 model include:

[0109] First, the CBAM attention mechanism is introduced before the Neck network and the Head network; then, the SPPF module is used to replace the original SPP spatial pyramid pooling as the backbone network of the network model, and serial MaxPool processing and fused convolution are used; secondly, the CA attention mechanism is added after each layer of CSP processing in the backbone network; finally, the multi-scale FPN feature network is used to replace the original neck network; among them, the serial MaxPool uses the maximum pooling layer, such as Fig.11 shown.

[0110] Specifically, in this embodiment, Figure 5 As shown, the specific workflow of real-time detection of tea diseases is as follows:

[0111] 1. Use the disease visual sensor to collect tea disease image data in different time periods and weather environments, use LabelImg software to preprocess and annotate the images, and establish a tea disease VOC dataset;

[0112] 2. Based on the deep learning target detection algorithm, the YOLOv5 network model is improved. SPPF spatial pyramid pooling is used in the backbone network. The image is processed by serial MaxPool and then fused convolution is performed. Fig.12 As shown, the computational efficiency is improved;

[0113] 3. After each layer of CSP processing in the backbone network, the CA attention mechanism is added to improve the network's ability to recognize disease characteristics, and then connected to the Neck network for corresponding processing;

[0114] 4. Using a multi-scale FPN feature network as the neck network, the network can not only propagate from top to bottom, but also from bottom to top, enhancing the detection effect of feature information;

[0115] 5. Before the Neck network and the Head network, the CBAM attention mechanism is introduced to improve the network's detection and positioning capabilities and identification of regions of interest, thereby enhancing the detection effect of small targets;

[0116] 6. The tea disease VOC dataset is used to train and test the network model. The image is passed into the network model for slicing and a series of convolution operations, including CBS, CSP and CAP. CBS includes convolution layer, BN layer and BiLu activation function. Finally, three feature maps of different scales are output. Each feature map corresponds to a prediction box of different scales. Tea diseases are detected in real time, and the disease detection box and category are output, as well as abnormal reports, as shown in the attached figure. Fig.10 shown.

[0117] The autonomous navigation and inspection method of the mobile robot in the tea garden proposed in this embodiment mainly includes four steps, namely: obtaining the image information of the tea garden road in two-dimensional space, autonomously following the target navigation path in three-dimensional space, and real-time detection of tea diseases. First, the complex roads in the tea garden are segmented by the improved Deeplabv3Plus algorithm to obtain the semantic information in the two-dimensional space. The semantic information in the two-dimensional space is used and combined with the depth image to be projected into the three-dimensional space to form a three-dimensional tea garden road point cloud; secondly, the tea garden road point cloud is processed, and the RANSAC algorithm is combined to fit the efficient and stable target navigation path, and the pure pursuit algorithm is used to autonomously follow the target navigation path; finally, on the basis of autonomous navigation, the improved YOLOv5 algorithm is used to monitor tea diseases in real time, realize the autonomous navigation and inspection operation of the mobile robot in the tea garden, ensure the production quality of tea, and at the same time provide scientific theoretical and practical support for building smart agriculture in rural areas and improving the level of intelligent agricultural production. The specific effects are shown in the attached figure. Figure 6 shown.

[0118] For the original Deeplabv3Plus algorithm (attached Figure 7 ), combined with the characteristics of complex road detection in tea gardens, the Deeplabv3Plus algorithm is improved. First, the lightweight Mobilenetv2 network is used to replace the original Xecption network with large parameters as the backbone network of the network model to reduce the number of model parameters and speed up the operation; secondly, a separate feature processing branch is added to the ASPP module, and the ECA attention mechanism is introduced in the branch to extract features. At the same time, the CBAM attention mechanism is introduced after the original ASPP module; in addition, the hole convolution method of the ASPP model is replaced with the depth-separable convolution method to improve the correlation of information in different receptive fields; finally, in the decoding layer, the 1st, 3rd, and 6th bottleneck layer feature information of the shallow features are fused, and the CBAM attention mechanism is introduced to distribute the feature weights, and finally fused with the deep features to achieve the detection of complex roads in the tea garden.

[0119] For the original YOLOv5 algorithm (attached Fig. 9 ), combined with the characteristics of tea plantation tea diseases, the YOLOv5 algorithm is improved. First, the CBAM attention mechanism is introduced before the Neck network and the Head network to improve the network's detection and positioning and identification of regions of interest, and enhance the detection effect of small targets; in addition, the SPPF module is used to replace the original SPP spatial pyramid pooling as the backbone network of the network model, and serial MaxPool processing and fused convolution are used to improve the computational efficiency; secondly, after each layer of CSP processing in the backbone network, the CA attention mechanism is added to improve the network's ability to recognize disease characteristics; finally, the multi-scale FPN feature network is used to replace the original neck network, and the network can not only propagate from top to bottom, but also from bottom to top, enhancing the detection effect of feature information and realizing the monitoring of tea diseases.

[0120] At present, the traditional manual operation method of detecting tea diseases is inefficient and costly. In addition, the existing methods focus more on the path planning and path following problems of flat sections or single obstacles, ignoring the terrain and obstacle randomness of the tea garden environment, resulting in the inability of mobile machines to perform autonomous navigation operations efficiently and stably. This embodiment obtains two-dimensional tea garden navigation path image information through the improved Deeplabv3Plus algorithm, and projects the two-dimensional space into a three-dimensional space, which is also one of the characteristics of this method. Through the visual conversion of space, a more realistic tea garden operation environment is presented; finally, the target navigation path is fitted and autonomous following is achieved to achieve autonomous navigation of the mobile robot. At the same time, on the basis of autonomous navigation, the improved YOLOv5 algorithm is used to monitor tea diseases in real time. This method can truly present the complex operation scenes of the tea garden, more effectively improve the navigation accuracy of the mobile robot and the stability of autonomous following, and at the same time can replace traditional manual operations, reduce labor costs, improve the detection rate and efficiency of tea diseases, ensure the quality of tea production, provide scientific and accurate protection measures for tea, and further provide theoretical and practical support for improving the level of rural intelligence and building rural smart agriculture.

[0121] This embodiment is based on the semantic segmentation method of visual navigation. By improving the Deeplabv3Plus algorithm, the image information of complex roads in the two-dimensional tea garden is obtained, the operation speed and detection efficiency of the network model are improved, and a solid foundation is provided for the subsequent processing of three-dimensional space road point cloud.

[0122] With the development of deep learning methods, disease detection based on deep learning is widely used and has good detection effects. This embodiment uses a target detection algorithm based on a deep learning method to detect tea diseases through an improved YOLOv5 algorithm, which can not only further improve the detection speed and the real-time and accuracy of disease detection, but also effectively identify the characteristics of tea diseases, without human subjective factors, and can provide accurate and scientific protective measures, further solving the problem of reducing labor costs and insufficient power.

[0123] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. An autonomous navigation inspection method for a mobile robot in a tea garden, characterized in that: include: Obtain tea garden road image information in two-dimensional space; Based on the tea garden road image information, construct a target navigation path in three-dimensional space; The tea garden mobile robot moves based on the target navigation path, and monitors the tea leaves in real time through disease visual sensing during the movement; Acquiring the tea garden road image information in two-dimensional space includes: Collect tea bush image data in tea gardens; Based on the tea bush image of the tea garden, construct a tea garden navigation path VOC dataset; Based on the tea garden navigation path VOC dataset, the improved Deeplabv3Plus model is trained and tested to obtain a road segmentation model; Based on the road segmentation model, obtaining the tea garden road image information in two-dimensional space; Improvements to the Deeplabv3Plus model include: The lightweight Mobilenetv2 network is used to replace the original Xecption network as the backbone network of the network model; a separate feature processing branch is added to the ASPP module, and the ECA attention mechanism is introduced in the branch to extract features. At the same time, the CBAM attention mechanism is introduced after the original ASPP module; in addition, the hole convolution method of the ASPP module is replaced with a depth-wise separable convolution method; finally, in the decoding layer, the feature information of the 1st, 3rd, and 6th bottleneck layers of the shallow features are fused, and the CBAM attention mechanism is introduced to allocate feature weights, which are finally fused with the deep features to realize the detection of complex roads in the tea garden.

2. The autonomous navigation inspection method of a mobile robot in a tea garden according to claim 1, characterized in that: The collection of tea garden tea bush image data includes: The tea bushes in the tea garden are selected as the preset navigation path of the mobile robot in the tea garden. The visual sensor is installed on the mobile robot in the tea garden to collect video data of the tea bushes in the tea garden in different time periods and weather environments. The tea garden tea bush image data is extracted from the collected tea garden tea bush video data.

3. The autonomous navigation inspection method of a mobile robot in a tea garden according to claim 1, characterized in that: Constructing the target navigation path in three-dimensional space includes: The tea garden road image information is combined with the depth image information and projected into a three-dimensional space to form three-dimensional point cloud information; Preprocessing the three-dimensional point cloud information; Based on the RANSAC algorithm, the preprocessed three-dimensional point cloud information is fitted, three points are randomly selected as the minimum sample set, the plane is determined, the distance from the point to the plane is calculated to determine the inner point, and the judgment is continuously iterated by setting a threshold. Finally, the plane whose inner point value is greater than the preset threshold is selected as the fitting result to obtain the target navigation path of the mobile robot in the three-dimensional space.

4. The autonomous navigation inspection method of a mobile robot in a tea garden according to claim 3, characterized in that: Preprocessing the three-dimensional point cloud information includes: The three-dimensional point cloud information is filtered by using bilateral filtering, and then the point cloud is downsampled by using a random downsampling method.

5. The autonomous navigation inspection method of a mobile robot in a tea garden according to claim 1, characterized in that: The tea garden mobile robot moves based on the target navigation path including: The target navigation path is transmitted to the ROS operating system of the tea garden mobile robot to obtain the current position information of the tea garden mobile robot; The adjustment mode of the tea garden mobile robot is preset, and the adjustment mode includes: fine-tuning steering mode and in-situ steering mode; when the steering angle of the tea garden mobile robot is less than or equal to 15°, the fine-tuning steering mode is adopted, otherwise the in-situ steering mode is adopted; The tea garden mobile robot selects the target point (P x ,P y ), by defining the motor output speed of the left and right wheels and Depend on and Where r is the radius of the left and right wheels, V r is the right wheel linear speed, V l For the left and right linear speeds, get the current speed of the mobile robot and angular velocity Based on the turning radius of the robot When V r =-V l When L is the wheelbase, the turning radius of the robot is R=0, and the tea garden mobile robot adopts the in-situ steering method. The robot turns to the same straight line as the target point and continues to move in a straight line. When the steering angle is less than or equal to 15°, the tea garden mobile robot adopts the fine-tuning steering method and adjusts the body posture in combination with the turning radius of the robot. Set the current speed V of the mobile robot c and angular velocity W c The input is sent to the control system of the mobile robot, and the mobile robot controls the turning radius and steering angle to achieve autonomous following of the target navigation path of the mobile robot in the tea garden.

6. The autonomous navigation inspection method of a mobile robot in a tea garden according to claim 1, characterized in that: Real-time monitoring of tea includes: Collect tea disease image data in different time periods and weather environments, preprocess and annotate the images, and establish a tea disease VOC dataset; Using the tea disease voc dataset, the improved YOLOv5 model is improved to obtain a tea disease detection model; The tea disease detection model is used to perform real-time detection of tea diseases, output disease detection frames and categories, and report abnormalities.

7. The autonomous navigation inspection method of a mobile robot in a tea garden according to claim 6, characterized in that: Improvements to the YOLOv5 model include: Firstly, the CBAM attention mechanism is introduced before the Neck network and the Head network; then, the SPPF module is used to replace the original SPP spatial pyramid pooling as the backbone network of the network model, and serial MaxPool processing and fused convolution are adopted; secondly, the CA attention mechanism is added after each layer of CSP processing of the backbone network; finally, the multi-scale FPN feature network is used to replace the original neck network.

Images