Seeder action path planning method and system

Abstract

The present application provides a method and system for planning the action path of a seed drill. The method includes: obtaining the real-time position information of the seed drill; collecting soil images and meteorological values ​​through a small unmanned aerial vehicle system; establishing an obstacle recognition neural network based on the soil images and meteorological values ​​to identify the position information of fixed obstacles and mobile obstacles in the seeding area; using a multi-spectral camera carried by a small unmanned aerial vehicle system to collect image data of the land after seeding, and calculating seeding quality indicators including seed coverage uniformity and seeding depth; based on the real-time position information, obstacle position information and seeding quality indicators, generating a path planning map for the seed drill. The present application improves the accuracy and efficiency of seeding operations by combining multi-source data for intelligent path planning.

Description

Technical Field

[0001] The present application relates to the technical field of agricultural machinery, and in particular to a method and system for planning the action path of a seeder. Background Art

[0002] With the development of agricultural modernization, the application of intelligent seed drills in agricultural production is becoming more and more extensive. The automation level of intelligent seed drills is constantly improving, which puts higher requirements on the planning of their action paths.

[0003] At present, the common planter path planning method is mainly based on GPS navigation technology, which operates through a pre-set path. For example, a straight-line round-trip operation path or a spiral operation path is adopted. Although this method is simple and intuitive, it lacks the ability to perceive and adapt to the working environment in real time.

[0004] Another more advanced path planning method is an autonomous navigation system based on computer vision, which uses cameras to collect image information and identify travel paths and obstacles. This method can achieve a certain degree of environmental perception, but due to the single visual information source, it is easily affected by factors such as weather and lighting in complex farmland environments, affecting recognition accuracy.

[0005] However, the main problem with the existing technology is that it is impossible to achieve real-time monitoring and feedback of the sowing quality, resulting in the seeder being unable to dynamically adjust the travel path according to the actual sowing effect, affecting the overall quality of the sowing operation. Summary of the invention

[0006] In view of this, the present application provides a method and system for planning the action path of a seeder, which solves the problem that the prior art cannot achieve real-time monitoring and feedback of the sowing quality.

[0007] The embodiment of the present application provides a method for planning the action path of a seed drill, comprising: obtaining real-time position information of the seed drill; equipping the seed drill with a small unmanned aerial vehicle system, collecting soil images through an image acquisition device of the small unmanned aerial vehicle system, and collecting meteorological values ​​through sensors of the small unmanned aerial vehicle system; establishing an obstacle recognition neural network based on the soil images and the meteorological values, and identifying the position information of fixed obstacles and mobile obstacles in the seeding area; wherein the fixed obstacles include stones and tree stumps, and the mobile obstacles include agricultural machinery; using a multispectral camera carried by the small unmanned aerial vehicle system to collect image data of the land after seeding, and calculating a seeding quality index based on the image data; wherein the seeding quality index includes seed coverage uniformity and seeding depth; generating a path planning map of the seed drill based on the real-time position information, the position information of the fixed obstacles and the mobile obstacles, and the seeding quality index.

[0008] Optionally, the step of obtaining the real-time position information of the planter includes:

[0009] An RTK-GNSS receiver is installed on the seed drill, and the RTK-GNSS receiver receives a satellite positioning signal;

[0010] Acquire differential correction data through the RTK base station set up in the sowing area;

[0011] The satellite positioning signal is integrated with the differential correction data to obtain the real-time position information of the seed drill;

[0012] The real-time position information is sent to the control system of the seed drill.

[0013] Optionally, the step of establishing an obstacle recognition neural network based on the soil image and the meteorological value includes:

[0014] Use convolutional neural networks to extract features from soil images and identify soil image features including soil type, arable land depth, and surface flatness;

[0015] Temperature stratification is performed based on the values ​​collected by the temperature sensor, humidity stratification is performed based on the values ​​collected by the humidity sensor, and wind speed stratification is performed based on the values ​​collected by the wind speed sensor, so as to obtain meteorological value stratification;

[0016] Based on soil image features and meteorological numerical stratification, deep neural networks are trained to build environmental factor recognition models;

[0017] Based on the environmental element recognition model, an obstacle recognition neural network is established.

[0018] Optionally, the step of identifying the position information of fixed obstacles and mobile obstacles in the sowing area includes:

[0019] Divide the sowing area into grids and use each grid as an identification unit;

[0020] In each recognition unit, the target features in the image are extracted, including color, texture, shape, and size;

[0021] The extracted target features are input into the obstacle recognition neural network, and the obstacle type in each recognition unit is output;

[0022] Use laser radar to obtain three-dimensional point cloud data of obstacles;

[0023] Align the identified obstacle type with the point cloud data and perform data fusion;

[0024] The precise location coordinates of each obstacle are calculated based on the fused data.

[0025] Optionally, the step of collecting image data of the land after sowing and calculating the sowing quality index based on the image data includes:

[0026] Control the multispectral camera to collect images of the land after sowing in the red light band and the near infrared band;

[0027] Calculating a normalized vegetation index based on the land image after sowing, and evaluating seed coverage uniformity according to the normalized vegetation index;

[0028] The image processing algorithm is used to analyze the surface texture characteristics of the land after sowing and evaluate the sowing depth.

[0029] Optionally, the step of calculating a normalized vegetation index based on the post-sowing land image and evaluating the seed coverage uniformity according to the normalized vegetation index comprises:

[0030] Calculate the pixel averages of the land images collected in the red band and near-infrared band respectively;

[0031] The normalized vegetation index is calculated according to the formula NDVI = (NIR-RED) / (NIR+RED), where NIR is the average pixel value corresponding to the near-infrared band and RED is the average pixel value corresponding to the red light band;

[0032] The sowing area is divided into several assessment units, and the mean NDVI value of each assessment unit is calculated;

[0033] The standard deviation of the NDVI mean of all assessment units was calculated as an indicator of seed coverage uniformity.

[0034] Optionally, based on the real-time position information, the position information of the fixed obstacles and the mobile obstacles, and the sowing quality index, the step of generating a path planning map of the seeder includes:

[0035] Based on the real-time position information, the current position coordinates of the planter are obtained as the starting point of the path planning;

[0036] Divide the sowing area into equally spaced planning grids;

[0037] marking the locations of fixed obstacles and movable obstacles on the planning grid;

[0038] Marking the to-be-reseeded area whose sowing quality index is lower than a preset threshold, and treating the unsown area and the to-be-reseeded area as unfinished areas;

[0039] Connect the current position of the planter with the unfinished area to generate multiple candidate paths;

[0040] Based on the A* algorithm, plan the shortest path to the unfinished area by bypassing fixed obstacles and moving obstacles;

[0041] According to the steering radius and operating speed constraints of the seeder, the shortest path is smoothed.

[0042] Optionally, based on the A* algorithm, the shortest path to the unfinished area around fixed obstacles and mobile obstacles is planned, including:

[0043] Establish a moving obstacle trajectory prediction model;

[0044] Based on historical location data, predict the location of mobile obstacles at future time points;

[0045] Convert predicted location information into space-time exclusion zones;

[0046] When the A* algorithm plans the seeder path, the time and space restricted areas are used as dynamic constraints;

[0047] When the deviation between the actual trajectory of a moving obstacle and the predicted trajectory exceeds a threshold, real-time path replanning is triggered.

[0048] Optionally, according to the steering radius and operating speed constraints of the seed drill, smoothing the shortest path includes:

[0049] Calculate the terrain slope at each point on the path based on real-time acquired terrain data;

[0050] Determine the maximum travel speed at the turn according to the terrain slope and the minimum turning radius of the seeder;

[0051] Use Bezier curves to smooth the path at the turns, ensuring that the curvature radius after smoothing is not less than the minimum turning radius of the planter;

[0052] Based on the upper limit of the operating speed, the PID control algorithm is used to achieve a smooth change in the seeder speed to ensure that the speed at the turn does not exceed the maximum driving speed.

[0053] The embodiment of the present application also provides a planter action path planning system, the planter action path planning system comprising:

[0054] An acquisition unit, used for acquiring real-time position information of the seed drill;

[0055] A collection unit, used to equip the seed drill with a small UAV system, collect environmental information through an image collection device of the small UAV system, and collect meteorological information through a sensor of the small UAV system;

[0056] an identification unit, configured to execute a deep learning network model to analyze the working environment based on the environmental information and the meteorological information, and identify the position information of fixed obstacles and mobile obstacles in the sowing area; wherein the fixed obstacles include stones and tree stumps, and the mobile obstacles include agricultural machinery;

[0057] A computing unit, used to collect image data of the red light band and the near infrared band of the land after sowing by using the multi-spectral camera carried by the small UAV system, and calculate the sowing quality index based on the image data; wherein the sowing quality index includes seed coverage uniformity and sowing depth;

[0058] A generating unit is used to generate a path planning map of the seeder based on the real-time position information, the position information of the fixed obstacles and the mobile obstacles, and the sowing quality index, wherein the path planning map is used to guide the seeder to avoid obstacles and complete operations in unsown areas.

[0059] An embodiment of the present application also provides a computer device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned seeder motion path planning method.

[0060] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the above-mentioned planter action path planning method.

[0061] An embodiment of the present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the above-mentioned planter action path planning method.

[0062] This application has the following technical effects: The seeder action path planning method collects soil images and meteorological values ​​through an integrated small drone system, and then establishes an obstacle recognition neural network to accurately identify the location information of fixed and mobile obstacles. At the same time, a multispectral camera is used to collect image data of the land after sowing, and sowing quality indicators are calculated, including seed coverage uniformity and sowing depth. After comprehensive consideration of these data, a path planning map of the seeder is generated, so that the seeder can dynamically adjust the path according to real-time position information, obstacle location information and sowing quality indicators. This method can realize real-time monitoring and feedback of sowing quality, so that the seeder can dynamically adjust the path according to the actual sowing effect, and improve the overall quality of the sowing operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the disclosed embodiments in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0064] Figure 1 It is a flow chart of a method for planning a planter's motion path provided in an embodiment of the present application;

[0065] Figure 2 It is a structural diagram of the RTK-GNSS positioning system in an embodiment of the present application;

[0066] Figure 3 is a schematic diagram of the architecture of an obstacle recognition neural network in an embodiment of the present application;

[0067] Figure 4 It is a flow chart of the method for calculating the sowing quality index in the embodiment of the present application;

[0068] Figure 5 is a schematic diagram of a path planning method in an embodiment of the present application;

[0069] Figure 6 is a schematic diagram of path smoothing processing in an embodiment of the present application;

[0070] Figure 7 It is a structural diagram of the planter motion path planning system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0071] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application is further described in detail below in conjunction with embodiments.

[0072] like Figure 1 As shown, the planter action path planning method includes:

[0073] S1: Get the real-time location information of the planter;

[0074] In order to achieve high-precision positioning of the planter, the present embodiment adopts RTK-GNSS technology. Figure 2 As shown, the present application installs an RTK-GNSS receiver on the seed drill and sets up an RTK base station in the seeding area, and achieves centimeter-level positioning accuracy through the cooperation of the two.

[0075] Specifically, S1 includes:

[0076] S1.1: installing an RTK-GNSS receiver on the seed drill, wherein the RTK-GNSS receiver receives a satellite positioning signal;

[0077] RTK-GNSS receivers are usually installed on the top of the planter or other locations that are convenient for receiving satellite signals. RTK-GNSS receivers use multi-frequency and multi-system receivers that can simultaneously receive signals from multiple satellite navigation systems such as GPS, Beidou, and Galileo. This multi-system reception design significantly increases the number of visible satellites and ensures the reliability and continuity of positioning solutions. The receiver captures navigation signals in the L1, L2 and other frequency bands sent by these satellites through a built-in antenna. These signals contain information such as the satellite's precise ephemeris and pseudorange measurements. When the RTK-GNSS receiver receives satellite signals, it simultaneously obtains pseudorange observations and carrier phase observations. Among them, the pseudorange observations provide initial positioning information with meter-level accuracy, and the carrier phase observations are used for subsequent high-precision positioning corrections.

[0078] S1.2: Obtain differential correction data through the RTK base station set up in the sowing area;

[0079] RTK base stations are key infrastructure for achieving high-precision positioning. Base stations are usually fixedly installed at commanding heights around the sowing area to ensure good communication line of sight with the seeder. The installation location of the RTK base station should be selected in an open area to avoid multipath effects; the distance between the RTK base station and the seeder should usually be kept within 10 kilometers to ensure the positioning accuracy of the RTK base station; the installation height of the RTK base station antenna is generally not less than 2 meters to reduce the interference of ground reflection signals. The RTK base station contains a high-precision GNSS antenna and receiver, which can establish precise reference coordinates through long-term observations. Since the base station is fixed in position, real-time differential correction data can be calculated by comparing the known coordinates with the observed coordinates. These differential correction data mainly contain system error information such as atmospheric delay and satellite orbit error.

[0080] The differential correction data can be obtained in the following way: the base station receiver calculates the integer ambiguity of the carrier phase; fixes the integer ambiguity using the known coordinates of the base station; and generates carrier phase and pseudorange correction data.

[0081] S1.3: fusing the satellite positioning signal with the differential correction data to obtain real-time position information of the seed drill;

[0082] Fusion processing is the core step to improve positioning accuracy. The RTK-GNSS receiver on the planter performs real-time fusion calculations on the received satellite raw observation data and the differential correction data sent by the base station. The common error is eliminated through the differential positioning algorithm, and the carrier phase observation value is used to fix the integer ambiguity.

[0083] The extended Kalman filter algorithm is used to fuse the satellite positioning signal with the differential correction data to obtain centimeter-level real-time position solution results. The state equation of the algorithm includes: position coordinates (X, Y, Z); velocity components (Vx, Vy, Vz); receiver clock error and clock drift.

[0084] S1.4: sending the real-time position information to the control system of the seed drill;

[0085] The high-precision position information obtained by the solution is transmitted to the control system of the planter in real time through standard interfaces (such as RS232, CAN bus, etc.). The control system compares these position data with the preset operation route to achieve accurate navigation and control of the planter's travel path. The update frequency of the position information is usually above 10Hz to ensure the continuity and stability of the control. The position information includes: longitude and latitude coordinates in the WGS84 coordinate system; altitude information; positioning precision factor (PDOP value); RTK solution status (fixed solution / floating point solution).

[0086] S2: Equip the seed drill with a small UAV system to collect environmental and meteorological data;

[0087] S2 may include the following sub-steps:

[0088] S2.1: Collecting soil images by using an image acquisition device of the small UAV system;

[0089] The specific configuration of the small UAV system in this application is as follows: the body adopts a four-rotor configuration, the deadweight does not exceed 2 kg, and the endurance time is not less than 30 minutes; it is equipped with an image acquisition device, including a visible light camera and an infrared camera, with a resolution of not less than 1920×1080 pixels; it is equipped with a temperature sensor, a humidity sensor and a wind speed sensor, and the sampling frequency can reach 1Hz; it has automatic take-off and landing and fixed altitude cruise functions, and the flight altitude can be adjusted within the range of 10-50 meters. A variety of imaging modes are used during the acquisition process, including: 1) Soil surface imaging, through the appropriate aperture (f / 2.8-f / 8), shutter speed (1 / 1000-1 / 60 seconds) and ISO sensitivity (100-800) settings, to ensure that clear soil surface images are obtained under different lighting conditions; 2) Multiple exposure imaging, through the combination of different exposure parameters, the dynamic range of the image is improved to make the soil characteristics more prominent; 3) Continuous shooting mode, to ensure that the image has sufficient overlap, which is convenient for later image stitching and processing. The time interval and flight altitude of image acquisition need to be precisely controlled to ensure the spatial continuity and integrity of the collected data.

[0090] The soil image acquisition strategy specifically includes: using an "S"-shaped flight path to perform a blanket scan of the sowing area; setting the flight altitude at 20 meters to ensure that a single image covers an area of ​​approximately 400 square meters; and setting the overlap rate of adjacent routes to 60% to ensure the accuracy of image stitching.

[0091] S2.2: Collecting meteorological values ​​through sensors of the small UAV system;

[0092] The collection of meteorological data relies on a variety of high-precision sensors on the drone. This sensing system mainly includes: 1) Temperature sensor, using PT100 platinum resistance temperature sensor, with a measurement range of -20℃ to 50℃, an accuracy of ±0.1℃, and a response time of 1 second; 2) Humidity sensor, using high-precision capacitive humidity sensor, with a measurement range of 0-100%RH, an accuracy of ±2%RH, and a temperature compensation range of -20℃ to 50℃; 3) Wind speed sensor, using ultrasonic anemometer, with a measurement range of 0-30m / s, an accuracy of ±0.1m / s, and a sampling frequency of 1Hz.

[0093] These sensors transmit data in real time via 4G / 5G or WiFi networks, using the MQTT protocol to ensure the reliability and real-time nature of data transmission. In addition, the collected meteorological values ​​can also be stored locally to prevent data loss due to communication interruptions.

[0094] Exemplarily, the collected meteorological values ​​are preprocessed as follows:

[0095] 1. Image preprocessing: Denoising: using Gaussian filtering algorithm; Color correction: based on color card calibration; Distortion correction: based on lens parameter model;

[0096] 2. Meteorological data preprocessing: outlier removal: using the 3σ criterion; data smoothing: using sliding average filtering; time synchronization: based on GPS timestamp alignment;

[0097] S3: Establish an obstacle recognition neural network and identify obstacles;

[0098] In the present application embodiment, Figure 3 As shown in the figure, a deep learning method is used to establish an obstacle recognition neural network, and its implementation process is divided into three main stages: feature extraction, model construction, and obstacle recognition.

[0099] S3 specifically includes:

[0100] S3.1: Use convolutional neural network to extract features from soil images and identify soil image features including soil type, tillage depth, and surface flatness;

[0101] The feature extraction of soil images uses a high-performance ResNet50 network architecture. The input is a 1920×1080×3 RGB image, which passes through 5 convolutional layer groups, each of which contains 3 convolutional blocks; residual connections are deployed in each convolutional block, and the gradient vanishing problem is alleviated by skip connections; the last convolutional layer outputs a 512-dimensional feature vector. The soil features that the model can recognize are divided into three categories: 1) Soil type recognition, including: extracting color histogram features; calculating grayscale co-occurrence matrix; extracting texture features; and accurately classifying soil types through deep feature fusion; 2) Cultivated land depth estimation, including: based on image shadow analysis; combining soil color change features; and quantitatively estimating through deep learning regression models; 3) Surface flatness assessment, including: extracting local gradient features; calculating surface roughness indicators; and using multi-scale feature fusion methods to assess flatness.

[0102] S3.2: Perform temperature stratification based on the values ​​collected by the temperature sensor, perform humidity stratification based on the values ​​collected by the humidity sensor, and perform wind speed stratification based on the values ​​collected by the wind speed sensor to obtain meteorological value stratification;

[0103] The hierarchical processing of meteorological data is based on multiple dimensions. Temperature stratification uses PT100 temperature sensor data to divide temperature values ​​into low temperature zone (less than 5°C), normal temperature zone and high temperature zone, and divides the temperature interval at 5°C intervals. Humidity stratification uses capacitive humidity sensor data to divide it into low humidity zone (less than 10%RH), normal humidity zone and high humidity zone, and divides the humidity interval at 10%RH intervals. Wind speed stratification is based on ultrasonic anemometer data and is divided into low wind speed zone (less than 2m / s), normal wind speed zone and high wind speed zone, and divides the wind speed interval at 2m / s intervals. These hierarchical data are fused through the neural network of the Transformer architecture to establish a hierarchical representation of meteorological features.

[0104] S3.3: Based on soil image features and meteorological numerical stratification, deep neural networks are trained to build environmental factor recognition models;

[0105] The image feature extraction branch and the meteorological feature extraction branch are used for dual-stream processing, and the attention mechanism is used to achieve adaptive fusion of features. The model training adopts a layered pre-training strategy, first pre-training each modality separately, and then performing end-to-end joint optimization.

[0106] Specifically, the deep neural network consists of two main parts: image feature extraction branch and meteorological feature extraction branch. The input of the image feature extraction branch is a 1920×1080×3 RGB image, which is extracted through 5 convolution blocks. Each convolution block contains two 3×3 convolution layers, followed by a batch normalization layer and a ReLU activation function. The number of output channels of the first convolution block is 64, and the number of channels doubles after each convolution block, and finally a 1024-dimensional image feature vector is obtained. The input of the meteorological feature extraction branch is three meteorological parameters: temperature, humidity, and wind speed. First, feature transformation is performed through 3 fully connected layers, and the output dimensions of each layer are 128, 256, and 512 respectively. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function, and a Dropout layer (dropout rate is 0.5) is used to prevent overfitting. Finally, a 512-dimensional meteorological feature vector is obtained.

[0107] The 1024-dimensional image features and 512-dimensional meteorological features are first mapped to the same 768-dimensional feature space through a linear transformation layer. Then 8 attention heads are set, and the dimension of each attention head is 96. By calculating the Query, Key, and Value matrices, dynamic weight allocation between features is achieved. Finally, the outputs of the 8 attention heads are concatenated and passed through a 768-dimensional fully connected layer to obtain the fused feature representation.

[0108] S3.4: Based on the environmental element recognition model, establish an obstacle recognition neural network;

[0109] Based on the environmental element recognition model, the obstacle recognition branch is expanded. This branch adopts a two-stage target detection architecture, including a region proposal network (RPN) and a target classification network. The detection capability of obstacles of different scales is enhanced through spatial pyramid pooling.

[0110] Specifically, the obstacle recognition neural network adopts an improved two-stage detection architecture. The first stage is the region proposal network (RPN), which contains a 3×3 sliding window convolution layer with 256 output channels. The output feature map of this layer passes through two parallel 1×1 convolution layers to predict the probability of target existence (2k channels, k is the preset number of anchor boxes) and bounding box regression parameters (4k channels). This application designs anchor boxes of 5 different scales (32×32, 64×64, 128×128, 256×256, 512×512) and 3 different ratios (1:1, 1:2, 2:1), for a total of 15 combinations.

[0111] The second stage is the target classification network. For the candidate regions generated by RPN, fixed-dimensional features are extracted through the spatial pyramid pooling module. This module contains 4 pooling layers with pooling sizes of 1×1, 2×2, 3×3, and 6×6, generating a total of 50-dimensional features. These features pass through two 2048-dimensional fully connected layers, and finally output the target category probability (including 4 categories including background class) and bounding box fine regression parameters (4-dimensional coordinate offset) through two parallel fully connected layers.

[0112] In order to improve the detection performance of small-sized obstacles, this application introduces a feature pyramid structure in the feature extraction network. Specifically, a feature pyramid is constructed from multiple different scale layers of the basic convolutional network (usually the output of the last three convolutional blocks). After the high-level features adjust the number of channels through 1×1 convolution, they are upsampled and fused with the low-level features. Finally, target detection is performed independently on each layer of feature maps, which effectively improves the detection ability of targets of different scales.

[0113] Through the above network structure, the present application realizes accurate identification of environmental elements in the sowing area and precise detection of obstacles, providing a reliable environmental perception basis for subsequent path planning.

[0114] S3.5: Divide the sowing area into grids, and use each grid as an identification unit;

[0115] The grid division adopts an adaptive grid strategy. The sowing area is divided into 5m×5m recognition units to ensure that the detection accuracy is not less than 20%. The grid size is dynamically adjusted according to the terrain characteristics, and a denser division is adopted in complex areas.

[0116] S3.6: In each recognition unit, extract the target features in the image, including color, texture, shape, and size;

[0117] The target feature extraction adopts a multi-scale and multi-level feature extraction strategy, including: 1) color features, extracting RGB and HSV multi-color space features; 2) texture features, using LBP, co-occurrence matrix and other algorithms; 3) shape features, extracting HOG and other gradient features; 4) size features, obtaining actual size information through comparative analysis and reference calibration objects.

[0118] S3.7: Input the extracted target features into the obstacle recognition neural network, and output the obstacle type in each recognition unit;

[0119] The feature input adopts an integrated learning strategy to integrate the prediction results of multiple classifiers. The soft voting mechanism is used to integrate the prediction probabilities of each classifier to improve the stability and accuracy of classification.

[0120] S3.8: Use LiDAR to obtain 3D point cloud data of obstacles;

[0121] The laser radar point cloud acquisition uses a 16-line laser radar with a scanning frequency of 20Hz, an angular resolution of 0.1°, and a scanning range of 360°. The density and integrity of the point cloud data are ensured through high-speed rotation.

[0122] S3.9: Align the coordinates of the identified obstacle type with the point cloud data and perform data fusion;

[0123] Coordinate alignment and data fusion use the ICP (Iterative Closest Point) algorithm, combined with the Kalman filter to optimize the fusion results. The correspondence between the 2D image and the 3D point cloud is established through feature point matching.

[0124] S3.10: Calculate the precise position coordinates of each obstacle based on the fused data;

[0125] The position coordinates are calculated based on the fused multi-source data, and the RANSAC algorithm is used to eliminate outliers, and the accurate three-dimensional coordinates of obstacles are obtained through least squares optimization. A confidence scoring mechanism is established to quantitatively evaluate the reliability of the positioning results.

[0126] S4: Calculate sowing quality index;

[0127] like Figure 4 As shown, the image data of the land after sowing is collected by a multispectral camera, and the sowing quality index is calculated based on the normalized difference vegetation index (NDVI). S4 specifically includes:

[0128] S4.1: Control the multispectral camera to collect images of the land after sowing in the red light band and the near infrared band;

[0129] The imaging system of the multispectral camera adopts a high-precision dual-band imaging solution. The acquisition range of the red light band is 630-690nm, and the near-infrared band is 760-900nm. The camera is equipped with a high-resolution imaging device (1280×960 pixels), with band resolutions of 10nm for red light and 20nm for near-infrared. During the acquisition process, an integration time of 0.1 seconds / pixel is used, and automatic exposure control is used to ensure adaptation to different lighting environments. The shutter speed range is 1 / 2000-1 / 100 seconds. The image acquisition workflow includes: exposure condition optimization (30 seconds), image acquisition (5 seconds / time), cache upload (20 seconds), and the sampling frequency is maintained at 2Hz.

[0130] S4.2: Calculate the pixel averages of the land images collected by the red light band and the near infrared band respectively;

[0131] The pixel average value is calculated using a regional and multi-level statistical method. The acquisition area is divided into multiple calculation units of 10×10 pixels. Each unit contains data from three adjacent frames, and the calculated unit mean is used as the representative value of the area. This method can effectively reduce the impact of local noise and uneven illumination.

[0132] S4.3: Calculate the Normalized Difference Vegetation Index according to the formula NDVI = (NIR-RED) / (NIR+RED), where NIR is the average pixel value corresponding to the near infrared band, and RED is the average pixel value corresponding to the red light band;

[0133] NDVI index calculation is a key step in assessing vegetation coverage. Using the difference in reflectance between the near infrared band (NIR) and the red band (RED), the formula NDVI = (NIR-RED) / (NIR+RED) is used to calculate the combination of ratio and difference. This method can effectively eliminate the influence of external factors such as terrain shadows and atmospheric conditions, and obtain a normalized vegetation index.

[0134] S4.4: Divide the sowing area into several assessment units and calculate the mean NDVI value of each assessment unit;

[0135] The evaluation unit is divided into a 10m x 10m grid system. The mean NDVI is calculated for each evaluation unit, and the coefficient of variation (CV) is used to analyze the spatial distribution pattern. The coefficient of variation is calculated by the ratio of the standard deviation to the mean: CV = standard deviation / mean × 100%. According to the CV value, the coverage uniformity is divided into:

[0136] CV < 10% is a highly uniform area

[0137] 10%≤CV<20% is the moderate uniform area

[0138] CV ≥ 20% is a low uniformity area

[0139] S4.5: Calculate the standard deviation of the NDVI mean of all assessment units as an indicator of seed coverage uniformity;

[0140] The standard deviation statistics adopt the weighted calculation method. First, the NDVI value of each evaluation unit is tested for normality to remove the influence of outliers. Then different weights are assigned according to the area size of each unit, and the overall weighted standard deviation is calculated, which can more accurately reflect the actual uniformity of seed coverage.

[0141] S4.6: Analyze the surface texture characteristics of the land after sowing through image processing algorithms to assess the sowing depth;

[0142] Seeding depth assessment uses a combination of texture analysis and stereo vision:

[0143] 1) Texture feature extraction uses gray level co-occurrence matrix (GLCM) to analyze features such as contrast, entropy, uniformity and energy

[0144] 2) Depth estimation uses the support vector regression (SVR) method and uses the RBF kernel function for feature mapping

[0145] 3) The accuracy of depth assessment is verified by comparing with the measured depth samples. The error range is controlled within 3-5 cm, the medium error range is 2-3 cm or 5-6 cm, and the significant error is less than 2 cm or greater than 6 cm.

[0146] S5: Generate a path planning map of the seed drill;

[0147] In the specific implementation process of path planning, the starting position of the planter must be determined first. The real-time position information obtained by the RTK-GNSS receiver is used as the starting point of the path planning. This position information contains longitude and latitude coordinates with centimeter-level accuracy. In order to facilitate subsequent processing, the longitude and latitude coordinates are converted into XY coordinates in the local plane coordinate system.

[0148] In order to achieve accurate path planning, this application adopts a grid management strategy. The entire sowing area is divided into a planning grid of 5 meters by 5 meters, and each grid unit is assigned a unique identification code. This grid division method is not only convenient for marking the location of obstacles, but also conducive to tracking and managing the sowing quality.

[0149] After the grid division is determined, two types of important information are marked on the planning grid. The first type is obstacle information, including the location of fixed obstacles (such as stones and tree stumps) and mobile obstacles (such as other agricultural machinery). The second type is the area to be reseeded based on the sowing quality index. These areas include areas that have not been sown yet and areas that need to be reseeded because the sowing quality does not meet the standards. These two types of areas are uniformly marked as "unfinished areas".

[0150] In order to deal with the moving obstacles in the scene, this application developed a trajectory prediction model based on deep learning. The model adopts a long short-term memory network (LSTM) structure, and the input includes the historical position sequence, speed and direction information of the obstacle. Through this model, it is possible to predict the possible position of the moving obstacle in the next 5-10 seconds. The prediction results are converted into a spatiotemporal restricted area, that is, an area where the planter is not allowed to pass at a specific time point.

[0151] In the specific path planning process, an improved A algorithm is used. Compared with the traditional A algorithm, the improvement of this application is mainly reflected in the design of the heuristic function. The heuristic function not only takes into account the Manhattan distance from the starting point to the target point, but also introduces the obstacle density factor and the terrain undulation factor, so that the generated path can better adapt to the actual environment. In addition, a dynamic obstacle avoidance mechanism is also implemented. When it is detected that the deviation between the actual trajectory of the moving obstacle and the predicted trajectory exceeds a preset threshold (usually set to 1 meter), the real-time replanning process will be triggered.

[0152] To ensure that the planter can smoothly execute the planned path, path smoothing is also implemented. First, the terrain slope of each point on the path is calculated based on the terrain data obtained in real time. The physical parameters of the planter (such as the minimum turning radius) and the terrain slope information are combined to determine the appropriate maximum driving speed for the turn. Subsequently, the turn is smoothed using a third-order Bezier curve to ensure that the curvature radius of the smoothed path is always greater than the minimum turning radius of the planter.

[0153] Finally, in order to achieve smooth control of the seeder speed, an improved PID control algorithm is used. The algorithm uses a fuzzy adaptive method to dynamically adjust the PID parameters, thereby achieving smooth changes in the seeder speed while ensuring that the speed at the turn does not exceed the maximum safe speed. The output of the controller is processed by speed limiting to ensure that the movement of the seeder is always within a safe range.

[0154] In this way, through the implementation of the above-mentioned technical solution, the seeder path planning method provided in this application can effectively avoid obstacles and ensure the quality and efficiency of the seeding operation.

[0155] Specifically, if Figure 5 As shown, generating the path planning graph of the seeder includes the following sub-steps:

[0156] S5.1: Based on the real-time position information, the current position coordinates of the planter are obtained as the starting point of the path planning;

[0157] The high-precision location data obtained by the RTK-GNSS receiver is used as the starting coordinates of the path planning. The WGS84 coordinate system is used to ensure global consistency. The location information includes longitude, latitude and altitude. The Kalman filter algorithm is used to eliminate positioning noise, and the real-time update frequency is maintained at more than 10Hz.

[0158] S5.2: Divide the sowing area into equally spaced planning grids;

[0159] The planning grid adopts an adaptive grid division strategy. The basic grid size is set to 5m×5m, and can be dynamically adjusted to a smaller size in complex terrain or areas with dense obstacles. The grid index adopts a quadtree structure to facilitate fast search and update.

[0160] S5.3: Mark the locations of fixed obstacles and movable obstacles on the planning grid;

[0161] Obstacle marking uses a multi-layer map structure. Fixed obstacles (such as electric poles and ditches) use static placeholder layers, and mobile obstacles (such as agricultural machinery and personnel) use dynamic placeholder layers. Each obstacle is given a unique ID and its location, size, and type information are recorded.

[0162] S5.4: marking the to-be-reseeded area whose sowing quality index is lower than a preset threshold, and treating the unsown area and the to-be-reseeded area as unfinished areas;

[0163] The determination of the area to be reseeded is based on a comprehensive evaluation of multiple indicators. Parameters such as NDVI value, sowing depth, and soil quality are analyzed. When any indicator is lower than the preset threshold, the area is marked as an area to be reseeded. These areas together with the unseeded areas constitute the target area for task planning.

[0164] S5.5: Connect the current position of the planter with the unfinished area to generate multiple candidate paths;

[0165] Path generation adopts a hierarchical search strategy. First, a global path corridor is constructed, and then multiple candidate paths are generated within the corridor. Path generation takes into account factors such as turning radius and slope constraints, and uses the RRT* (rapid random expansion tree) algorithm to implement path search.

[0166] S5.6: Establish a moving obstacle trajectory prediction model;

[0167] The mobile obstacle prediction model is based on the long short-term memory network (LSTM) architecture. The model input includes the historical trajectory, speed and direction information of the obstacle, and outputs the position prediction at the future moment.

[0168] S5.7: Based on historical position data, predict the position of the mobile obstacle at a future time point;

[0169] The prediction process uses a sliding time window method to predict the movement trajectory within the next 5-10 seconds. The prediction results are updated every 100ms to maintain the real-time nature of the prediction.

[0170] S5.8: Convert the predicted position information into a spatiotemporal exclusion zone;

[0171] The construction of the spatiotemporal restricted area takes into account the dynamic envelope of obstacles. The restricted area range at different times is represented by a probabilistic occupancy grid, and the impact of prediction uncertainty is considered.

[0172] S5.9: When planning the seeder path using the A* algorithm, the time and space restricted areas are used as dynamic constraints;

[0173] Step S5.9 is equivalent to forming an improved version of the A* algorithm, that is, adding a time dimension to the A* algorithm to form a spatiotemporal A* algorithm.

[0174] In the spatiotemporal A* algorithm, the search space is expanded from the two-dimensional plane to the three-dimensional spatiotemporal domain, and the state nodes are expanded from the original position coordinates (x, y) to the spatiotemporal coordinates (x, y, t). Specifically, the implementation of the algorithm includes the following key parts:

[0175] First, the present application designs a new heuristic function h(n). This function not only considers the spatial distance, but also introduces the evaluation of the time dimension: h(n) = w 1 h s (n)+w 2 h t (n). Among them, h s (n) represents the spatial distance term, which is calculated using the improved Euclidean distance; h t (n) represents the time cost term, which is based on the estimated time from the current node to the target point. Weight coefficient w 1 and w 2 The initial values ​​are set to 0.6 and 0.4 respectively through adaptive method and dynamically adjusted according to the real-time planning effect.

[0176] In the design of the cost function g(n), this application proposes a multi-factor comprehensive evaluation method:

[0177] g(n)=g s (n)+g t (n)+g c (n). Among them, g s (n) is the actual path length, g t (n) is the time consumption, g c (n) is the collision risk cost.

[0178] In particular, g c The calculation of (n) takes into account the spatial distance and temporal overlap with the space-time exclusion zone, using the following formula:

[0179] g c (n) = ∑α i *exp(-d i / d 0 )*exp(-|t i -t| / t 0 ).

[0180] Among them, d i is the shortest distance from the current node to the i-th restricted area, t i is the time center of the restricted area, d 0 and t 0is the characteristic scale of distance and time, α i is the risk weight factor.

[0181] In terms of node expansion strategy, this application develops an adaptive spatiotemporal sampling method. In the spatial domain, eight expansion directions are set according to the movement characteristics of the planter, and the step size of each direction is dynamically adjusted according to the current speed, ranging from 0.5 meters to 2 meters. In the time domain, the sampling interval increases with the increase of the prediction time domain, specifically: Δt = t 0 *(1+β*k), where t 0 is the basic time step (set to 0.5 seconds), k is the number of prediction steps, and β is the growth coefficient (set to 0.1).

[0182] In order to handle dynamic constraints, this application designs an efficient collision detection module. This module represents the space-time restricted area as a series of four-dimensional bounding boxes (x_min, x_max, y_min, y_max, t_min, t_max). When the node is expanded, the fast separating axis theorem (SAT) is used to detect whether the current path segment intersects any restricted area. In the specific implementation, the path segment is discretized into a series of space-time points, and each point is detected whether it falls in any restricted area, while considering the actual footprint size of the planter.

[0183] In addition, this application also introduces a path prediction mechanism. When expanding a node, a simplified kinematic model is used to predict the feasible motion range of the planter at the node, and the expansion direction that does not meet the motion constraints is filtered out in advance. The kinematic model takes into account the constraints of the maximum steering angular velocity ωmax and the maximum linear velocity vmax to ensure that the planned path meets the actual motion capability of the planter.

[0184] Finally, in order to improve the real-time performance of the algorithm, this application adopts a hierarchical search strategy. First, a quick path search is performed on a coarse space-time grid to obtain the approximate path direction. Then, local optimization is performed in a fine grid near the path, which significantly reduces the computational complexity while ensuring the planning quality. An adaptive termination condition is also set during the search process. When the cost of the found path is less than the preset threshold, or the search time exceeds the limit, the search is terminated in time.

[0185] The spatiotemporal A* algorithm of the present application can effectively handle path planning problems in dynamic environments and provide a safe and efficient motion trajectory for the seed drill.

[0186] S5.10: When the deviation between the actual trajectory of the moving obstacle and the predicted trajectory exceeds a threshold, the real-time path replanning is triggered;

[0187] The trigger conditions for path replanning include prediction errors exceeding 1 meter or angle deviations exceeding 15 degrees. Replanning uses a local path correction strategy to maintain the overall continuity of the path.

[0188] S5.11: Calculate the terrain slope at each point on the path based on the terrain data acquired in real time;

[0189] Terrain analysis uses DTM (digital terrain model) data to calculate the three-dimensional coordinates and normal vector of each grid point to obtain local slope information.

[0190] S5.12: Determine the maximum driving speed at the turn according to the terrain slope and the minimum turning radius of the planter;

[0191] Turning speed planning is based on the vehicle dynamics model. It takes into account factors such as center of mass offset and sideslip angle, and calculates the upper limit of safe speed in combination with the minimum turning radius.

[0192] S5.13: Use Bezier curve to smooth the path at the turning point, and ensure that the curvature radius after smoothing is not less than the minimum turning radius of the planter;

[0193] like Figure 6 As shown in the figure, the path smoothing adopts the third-order Bezier curve, and the selection of control points ensures the curvature continuity. The control point positions are adjusted through iterative optimization so that the smoothed path meets the minimum turning radius constraint.

[0194] S5.14: Based on the upper limit of the operating speed, the PID control algorithm is used to achieve a smooth change in the seeder speed to ensure that the speed at the turn does not exceed the maximum driving speed.

[0195] The PID control parameters adopt an adaptive adjustment strategy, and the proportional, integral, and differential coefficients are dynamically adjusted according to the speed error. Acceleration constraints are established to ensure the smoothness of speed changes.

[0196] The steps of analyzing the surface texture characteristics of the land after sowing by image processing algorithms and evaluating the sowing depth include:

[0197] The frequency distribution of soil surface texture characteristics was analyzed using Fourier transform;

[0198] The sowing depth index is calculated based on the frequency distribution characteristics.

[0199] In the process of seeding depth assessment, this application innovatively uses a soil surface texture analysis method based on Fourier transform. Specifically, the soil surface image collected by the multispectral camera is first preprocessed, including image enhancement, noise removal, and contrast adjustment, to highlight the texture characteristics of the soil surface. The preprocessed image is converted into a grayscale image to ensure the accuracy of subsequent analysis.

[0200] In the Fourier transform analysis stage, the two-dimensional fast Fourier transform (2D-FFT) is used to convert the image from the spatial domain to the frequency domain. This transformation makes the periodic texture characteristics of the soil surface appear as different frequency components in the frequency domain. By analyzing the distribution characteristics of the frequency spectrum, the texture information related to the sowing depth can be effectively extracted. Specifically, the deep sowing area usually shows a spectrum feature dominated by low frequencies, while the shallow sowing area shows more high-frequency components.

[0201] In order to improve the accuracy of the analysis, the present application first performs polar coordinate transformation on the spectrum graph, and then calculates the energy distribution in different directions. This method can effectively capture the directional texture characteristics caused by the sowing operation. The present application can also establish a mapping relationship database between frequency distribution and sowing depth, and obtain a reliable depth estimation model through statistical analysis of a large amount of experimental data.

[0202] This application innovatively introduces the spectrum energy ratio method in the calculation of the sowing depth index. Specifically, the frequency spectrum is divided into a low frequency band (0-0.2 times the Nyquist frequency), a mid-frequency band (0.2-0.6 times the Nyquist frequency) and a high frequency band (0.6-1.0 times the Nyquist frequency), and the energy proportions of these frequency bands are calculated respectively. Through the ratio relationship of the energy of these frequency bands, a quantitative evaluation index of the sowing depth is established. After experimental verification, the depth estimation error of this method is controlled within the range of ±0.5 cm, which meets the actual production needs.

[0203] In addition, considering the influence of factors such as soil type and water content on texture characteristics, this application can also dynamically adjust the parameters of spectrum analysis through synchronously collected soil water content data and soil type information to ensure accurate depth estimation results under different environmental conditions. This compensation mechanism significantly improves the adaptability and reliability of this method in practical applications.

[0204] In addition, if Figure 7 As shown, the present application also provides a planter action path planning system, comprising:

[0205] An acquisition unit 10 is used to acquire real-time position information of the seed drill;

[0206] A collection unit 20 is used to equip the seed drill with a small UAV system, collect environmental information through an image collection device of the small UAV system, and collect meteorological information through a sensor of the small UAV system;

[0207] The identification unit 30 is used to execute a deep learning network model to analyze the working environment based on the environmental information and the meteorological information, and identify the location information of fixed obstacles and mobile obstacles in the sowing area; wherein the fixed obstacles include stones and tree stumps, and the mobile obstacles include agricultural machinery;

[0208] The computing unit 40 is used to collect image data of the red light band and the near infrared band of the land after sowing by using the multi-spectral camera carried by the small UAV system, and calculate the sowing quality index based on the image data; wherein the sowing quality index includes seed coverage uniformity and sowing depth;

[0209] The generating unit 50 is used to generate a path planning map of the seeder based on the real-time position information, the position information of the fixed obstacles and the mobile obstacles, and the sowing quality index, wherein the path planning map is used to guide the seeder to avoid obstacles and complete operations in unsown areas.

[0210] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present disclosure to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for planning a planter's motion path, characterized in that: The following steps are involved: Get the real-time location information of the planter; A small unmanned aerial vehicle system is equipped on the seed drill, and soil images are collected by an image acquisition device of the small unmanned aerial vehicle system, and meteorological values ​​are collected by a sensor of the small unmanned aerial vehicle system; Based on the soil image and the meteorological value, an obstacle recognition neural network is established to identify the position information of fixed obstacles and mobile obstacles in the sowing area; wherein the fixed obstacles include stones and tree stumps, and the mobile obstacles include agricultural machinery; The multispectral camera carried by the small UAV system is used to collect image data of the land after sowing, and the sowing quality index is calculated based on the image data; wherein the sowing quality index includes seed coverage uniformity and sowing depth; Based on the real-time position information, the position information of the fixed obstacles and the mobile obstacles, and the sowing quality index, a path planning map of the seeder is generated.

2. The method according to claim 1, characterized in that The steps to obtain the real-time location information of the planter include: An RTK-GNSS receiver is installed on the seed drill, and the RTK-GNSS receiver receives a satellite positioning signal; Acquire differential correction data through the RTK base station set up in the sowing area; The satellite positioning signal is integrated with the differential correction data to obtain the real-time position information of the seed drill; The real-time position information is sent to the control system of the seed drill.

3. The method according to claim 1, characterized in that The step of establishing an obstacle recognition neural network based on the soil image and the meteorological value comprises: Use convolutional neural networks to extract features from soil images and identify soil image features including soil type, arable land depth, and surface flatness; Temperature stratification is performed based on the values ​​collected by the temperature sensor, humidity stratification is performed based on the values ​​collected by the humidity sensor, and wind speed stratification is performed based on the values ​​collected by the wind speed sensor, so as to obtain meteorological value stratification; Based on soil image features and meteorological numerical stratification, deep neural networks are trained to build environmental factor recognition models; Based on the environmental element recognition model, an obstacle recognition neural network is established.

4. The method according to claim 3, characterized in that The steps of identifying the position information of the fixed obstacles and the mobile obstacles in the sowing area include: Divide the sowing area into grids and use each grid as an identification unit; In each recognition unit, the target features in the image are extracted, including color, texture, shape, and size; The extracted target features are input into the obstacle recognition neural network, and the obstacle type in each recognition unit is output; Use laser radar to obtain three-dimensional point cloud data of obstacles; Align the identified obstacle type with the point cloud data and perform data fusion; The precise location coordinates of each obstacle are calculated based on the fused data.

5. The method according to claim 1, characterized in that: The steps of collecting image data of the land after sowing and calculating the sowing quality index based on the image data include: Control the multispectral camera to collect images of the land after sowing in the red light band and the near infrared band; Calculating a normalized vegetation index based on the land image after sowing, and evaluating seed coverage uniformity according to the normalized vegetation index; The image processing algorithm is used to analyze the surface texture characteristics of the land after sowing and evaluate the sowing depth.

6. The method according to claim 5, characterized in that The steps of calculating the normalized vegetation index based on the land image after sowing and evaluating the seed coverage uniformity according to the normalized vegetation index include: Calculate the pixel averages of the land images collected in the red band and near-infrared band respectively; The normalized vegetation index is calculated according to the formula NDVI = (NIR-RED) / (NIR+RED), where NIR is the average pixel value corresponding to the near-infrared band and RED is the average pixel value corresponding to the red light band; The sowing area is divided into several assessment units, and the mean NDVI value of each assessment unit is calculated; The standard deviation of the NDVI mean of all assessment units was calculated as an indicator of seed coverage uniformity.

7. The method according to claim 1, characterized in that Based on the real-time position information, the position information of the fixed obstacles and the mobile obstacles, and the sowing quality index, the step of generating a path planning map of the sowing machine includes: Based on the real-time position information, the current position coordinates of the planter are obtained as the starting point of the path planning; Divide the sowing area into equally spaced planning grids; marking the locations of fixed obstacles and movable obstacles on the planning grid; Marking the to-be-reseeded area whose sowing quality index is lower than a preset threshold, and treating the unsown area and the to-be-reseeded area as unfinished areas; Connect the current position of the planter with the unfinished area to generate multiple candidate paths; Based on the A* algorithm, plan the shortest path to the unfinished area by bypassing fixed obstacles and moving obstacles; According to the steering radius and operating speed constraints of the seeder, the shortest path is smoothed.

8. The method according to claim 7, characterized in that Based on the A* algorithm, the shortest path to the unfinished area is planned to bypass fixed obstacles and mobile obstacles, including: Establish a moving obstacle trajectory prediction model; Based on historical location data, predict the location of mobile obstacles at future time points; Convert predicted location information into space-time exclusion zones; When the A* algorithm plans the seeder path, the time and space restricted areas are used as dynamic constraints; When the deviation between the actual trajectory of a moving obstacle and the predicted trajectory exceeds a threshold, real-time path replanning is triggered.

9. The method according to claim 7, characterized in that: According to the steering radius and operating speed constraints of the planter, smoothing the shortest path includes: Calculate the terrain slope at each point on the path based on real-time acquired terrain data; Determine the maximum travel speed at the turn according to the terrain slope and the minimum turning radius of the seeder; Use Bezier curves to smooth the path at the turns, ensuring that the curvature radius after smoothing is not less than the minimum turning radius of the planter; Based on the upper limit of the operating speed, the PID control algorithm is used to achieve a smooth change in the seeder speed to ensure that the speed at the turn does not exceed the maximum driving speed.

10. A planter action path planning system, characterized in that: include: An acquisition unit, used for acquiring real-time position information of the seed drill; A collection unit, used to equip the seed drill with a small UAV system, collect environmental information through an image collection device of the small UAV system, and collect meteorological information through a sensor of the small UAV system; an identification unit, configured to execute a deep learning network model to analyze the working environment based on the environmental information and the meteorological information, and identify the position information of fixed obstacles and mobile obstacles in the sowing area; wherein the fixed obstacles include stones and tree stumps, and the mobile obstacles include agricultural machinery; A computing unit, used to collect image data of the red light band and the near infrared band of the land after sowing by using the multi-spectral camera carried by the small UAV system, and calculate the sowing quality index based on the image data; wherein the sowing quality index includes seed coverage uniformity and sowing depth; A generating unit is used to generate a path planning map of the seeder based on the real-time position information, the position information of the fixed obstacles and the mobile obstacles, and the sowing quality index, wherein the path planning map is used to guide the seeder to avoid obstacles and complete operations in unsown areas.

Images