Garlic harvester automatic row alignment control method and system based on machine vision

Abstract

The application discloses a garlic harvester automatic alignment control method and system based on machine vision, and belongs to the field of agricultural machinery automatic control, which comprises the following steps: obtaining image data based on machine vision, generating a plant density distribution map by using a deep learning network; constructing a density-velocity coupling dynamics model based on the plant density distribution map, solving the optimal advancing speed of the garlic harvester under the current plant density condition by using a nonlinear mapping function and an optimization algorithm; compensating visual perception based on the optimal advancing speed to obtain compensated visual perception information; constructing a multi-time scale predictive control framework based on the compensated visual perception information and the optimal advancing speed to obtain a control strategy; and optimizing the density-velocity coupling dynamics model and the control strategy based on actual operation data to realize automatic alignment control of the garlic harvester. The application replaces the traditional multiple sets of speed control and path detection strategies, reduces the complexity of the control system, and improves the adaptability of the system.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of automatic control technology of agricultural machinery, and particularly relates to a garlic harvester automatic row alignment control method and system based on machine vision. BACKGROUND

[0002] With the development of modern agriculture towards intelligence and automation, the automatic control technology of agricultural machinery based on machine vision has attracted widespread attention. In the garlic harvesting operation, the use of machine vision to identify garlic stalk lodging and the boundary to be harvested has become an important research direction. Related technologies also include the study of flexible toothed active support mechanism combined with garlic growth posture, the study of weed and mulch entanglement and blockage detection and early warning method in the garlic harvesting process, the development of double-layer garlic precise flexible positioning cutting mechanism with active weed removal system, and the use of gravity sensor to study automatic unloading device, integrated development of automatic bag changing device and yield estimation system, etc.

[0003] The existing mechanical garlic combined harvesting technology has the problems of difficult row alignment, missed picking and high damage rate, and lacks an automatic unloading device. Specifically, the existing automatic row alignment control technology of garlic harvester based on machine vision mainly has the following technical problems:

[0004] The traditional garlic harvester usually adopts a fixed speed operation mode, which cannot adapt to the change of plant density in the field. In high density areas, fixed high speed operation will lead to decreased tracking accuracy and increased garlic damage rate. In low density areas, fixed low speed operation will result in low operation efficiency, wasting energy and time. In the existing technology, the machine vision system and the speed control system are usually independent of each other, lacking effective coordination mechanism. The vision system cannot automatically adjust its parameter configuration according to the current speed, resulting in decreased vision recognition accuracy at high speed. When the speed of the harvester changes, the blur degree of image acquisition, exposure conditions and field of view will change accordingly. The existing technology lacks a compensation mechanism for the influence of speed change on vision perception, making it difficult for the control system to cope with variable speed scenarios. The parameters of the traditional row alignment control system are fixed and cannot be continuously optimized according to actual operation data, making it difficult to adapt to complex and variable field environments. The existing technology usually adopts a single control strategy, which cannot simultaneously consider short-term precision control and long-term path planning, resulting in poor control effect under complex terrain and uneven plant distribution conditions.

[0005] Therefore, it is necessary to improve the intelligent level of garlic planting and harvesting, innovate and develop unmanned production equipment for the whole process of garlic, and promote industrialization and demonstration, so as to improve the unmanned level of garlic production and provide an important guarantee for the sustainable and healthy development of garlic industry. SUMMARY

[0006] To solve the above technical problems, the application provides a machine vision-based automatic row control method and system for a garlic harvester to solve the problems existing in the prior art.

[0007] To achieve the above object, the application provides a machine vision-based automatic row control method for a garlic harvester, comprising:

[0008] Based on the image data obtained by machine vision, the plant density features of the current area and the front area are analyzed in real time by using a deep learning network to generate a plant density distribution map;

[0009] Based on the plant density distribution map, a density-velocity coupling dynamics model is constructed, and the optimal travel speed of the garlic harvester under the current plant density condition is solved by a nonlinear mapping function and an optimization algorithm;

[0010] Based on the optimal travel speed, adaptive image preprocessing, dynamic adjustment of the field of view range, and dynamic adjustment of the feature extraction parameters are used to compensate for the visual perception to obtain compensated visual perception information;

[0011] Based on the compensated visual perception information and the optimal travel speed, a multi-time scale predictive control framework is constructed, and a control strategy is obtained based on the multi-time scale predictive control framework;

[0012] Based on actual operation data, the density-velocity coupling dynamics model and the control strategy are continuously optimized to realize automatic row control of the garlic harvester.

[0013] Optionally, based on the image data obtained by machine vision, the process of analyzing the plant density features of the current area and the front area in real time by using a deep learning network to generate a plant density distribution map comprises:

[0014] Based on a multi-scale convolutional neural network, the collected images are subjected to feature extraction to generate multi-scale features;

[0015] The multi-scale features are fused through a feature pyramid network, and the key area features are highlighted in combination with an attention mechanism to generate a preliminary density feature map;

[0016] The preliminary density feature map is processed by a density regression network to generate a plant density distribution map within the current field of view range.

[0017] Optionally, the process of obtaining the optimal travel speed of the garlic harvester under the current plant density condition comprises:

[0018] Based on the plant density distribution map, a system adaptability index containing an efficiency weight coefficient, a speed penalty coefficient, an accuracy weight coefficient, and an acceleration compensation coefficient is constructed, and a density-velocity coupling dynamics model is determined based on the system adaptability index;

[0019] The plant density is normalized by using a nonlinear mapping function, and the interaction between the plant density and the traveling speed is quantified by a density-speed interaction function;

[0020] Based on the interaction between the plant density and the traveling speed, a gradient descent method is used to optimize and correct the parameters of a density-speed coupling dynamics model;

[0021] Based on the optimized parameters, a numerical optimization method is used to solve the traveling speed that makes the system adaptability index optimal, as the optimal traveling speed under the current plant density condition.

[0022] Optionally, the expression of the system adaptability index is:

[0023] ;

[0024] In the formula, is the system adaptability index; is the plant density; is the traveling speed; , , , respectively represent the efficiency weight coefficient, the speed penalty coefficient, the precision weight coefficient, and the acceleration compensation coefficient; is the plant density normalization mapping function; is the density-speed interaction function; is the acceleration, and · represents the dot product.

[0025] Optionally, the expression for quantifying the interaction between the plant density and the traveling speed by the density-speed interaction function is:

[0026] ;

[0027] In the formula, is the density-speed interaction function; is the plant density; is the traveling speed; , , respectively represent the density linear influence coefficient, the speed linear influence coefficient, and the density-speed quadratic interaction coefficient, and · represents the dot product.

[0028] Optionally, based on the optimal traveling speed, an adaptive image preprocessing, a field of view range dynamic adjustment, and a feature extraction parameter dynamic adjustment method are used to compensate for the visual perception, and the process of obtaining the compensated visual perception information includes:

[0029] A speed-image blur mapping model is established based on the optimal traveling speed and the image quality;

[0030] Adaptive pre-processing the collected images based on the current traveling speed and the speed-image blur mapping model to obtain image quality evaluation results;

[0031] Dynamic adjustment of feature extraction parameters of the density estimation network based on the current traveling speed and the image quality evaluation results to obtain compensated visual perception information.

[0032] Optionally, the process of establishing the speed-image blur mapping model comprises:

[0033] Establishing a mapping relationship between speed and blur kernel standard deviation;

[0034] Establishing a relationship model between speed and image contrast;

[0035] Establishing a relationship model between speed and texture degradation;

[0036] Establishing a speed-image blur mapping model based on the mapping relationship between speed and blur kernel standard deviation, the relationship model between speed and image contrast, and the relationship model between speed and texture degradation.

[0037] Optionally, the process of constructing a multi-time scale predictive control framework comprises:

[0038] Dividing the control domain into short-term, medium-term and long-term time domains according to the time domain;

[0039] Constructing a hierarchical prediction model based on the short-term, medium-term and long-term time domains;

[0040] Obtaining control decisions of different time scales based on the hierarchical prediction model;

[0041] Coordinating control decisions of different time scales through a hierarchical decision structure and a constraint hardening mechanism to obtain final control instructions.

[0042] Optionally, the process of continuously optimizing the density-speed coupled dynamics model and the control strategy based on actual operation data to realize automatic row control of the garlic harvester comprises:

[0043] Collecting and storing plant density distribution, optimal traveling speed, visual compensation parameters and control execution effect data in the operation process in real time to construct a labeled operation dataset;

[0044] Updating the parameters of the density-speed coupled dynamics model using an incremental learning method and giving new weights through a window weighting strategy;

[0045] Based on a reinforcement learning framework, taking row tracking deviation and operation efficiency as reward signals to optimize the multi-time scale predictive control framework;

[0046] Based on the new weights and the optimized multi-time-scale predictive control framework, a knowledge base containing the characteristics of different operation scenarios is established. Based on the knowledge base containing the characteristics of different operation scenarios, adaptive migration of control parameters is achieved through similarity calculation, and optimization experience is shared during multi-machine collaborative operation.

[0047] The present invention also provides a method for automatically controlling rows of a garlic harvester based on machine vision, which is used to implement the method. The system includes:

[0048] The density perception module is used to analyze the plant density characteristics of the current area and the area ahead in real time based on image data acquired by machine vision using a deep learning network to generate a plant density distribution map;

[0049] A dynamic modeling module is used to construct a density-speed coupling dynamic model based on the plant density distribution map, and solve the optimal travel speed of the garlic harvester under the current plant density conditions through a nonlinear mapping function and an optimization algorithm;

[0050] a visual compensation module for compensating visual perception based on the optimal travel speed by using adaptive image preprocessing, dynamic adjustment of the visual field, and dynamic adjustment of feature extraction parameters to obtain compensated visual perception information;

[0051] a multi-scale control module, configured to construct a multi-time-scale predictive control framework based on the compensated visual perception information and the optimal travel speed, and obtain a control strategy based on the multi-time-scale predictive control framework;

[0052] The self-learning optimization module is used to continuously optimize the density-speed coupling dynamic model and control strategy based on actual operation data to achieve automatic row control of the garlic harvester.

[0053] Compared with the prior art, the present invention has the following advantages and technical effects:

[0054] The present invention provides a method for automatic row control of a garlic harvester based on machine vision. Through a deep learning network, plant density characteristics are analyzed in real time and a density distribution map is generated, achieving accurate perception of the field environment. The travel speed is dynamically adjusted based on a density-speed coupling dynamics model, effectively solving the problems of insufficient precision and low efficiency caused by the traditional fixed-speed operation mode. An adaptive visual compensation mechanism ensures the stability of the image recognition system at different speeds. A multi-time-scale predictive control framework is adopted to coordinate short-term precise control with long-term path planning, significantly improving the accuracy and adaptability of row control. Through a continuous learning optimization mechanism, the system can continuously accumulate experience from actual operations and improve itself, so that the control performance gradually improves over time. This method realizes intelligent control of garlic harvesting operations, effectively reduces the missed picking rate and damage rate, and improves operation efficiency and adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0056] Figure 1 This is a flow chart of a method for automatic row alignment control of a garlic harvester based on machine vision according to an embodiment of the present invention;

[0057] Figure 2 A line graph comparing the recognition accuracy of the traditional density estimation method of an embodiment of the present invention and the deep learning density estimation network of this technical solution under different lighting conditions;

[0058] Figure 3 A bar chart comparing system adaptability indicators under different plant densities and travel speeds according to an embodiment of the present invention;

[0059] Figure 4 A dual-axis combination diagram showing changes in image quality at different travel speeds and improvement effects brought about by the compensation mechanism according to an embodiment of the present invention;

[0060] Figure 5 A boxplot showing the distribution of row tracking deviations of a garlic harvester before and after adopting a multi-time-scale predictive control framework according to an embodiment of the present invention;

[0061] Figure 6 This is an area chart showing the performance improvement trend of the self-learning optimization system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0063] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0064] like Figure 1 As shown, this embodiment provides a method for automatic row control of a garlic harvester based on machine vision, comprising the following steps:

[0065] Step 1, analyze the image data obtained by machine vision through a deep learning network to estimate the plant density characteristics of the current area and the area ahead in real time, and generate a plant density distribution map. The image data obtained by machine vision is analyzed by a deep learning network to estimate the plant density characteristics of the current area and the area ahead in real time, and generate a plant density distribution map. Specifically, the following sub-steps are included.

[0066] Step 1.1, multi-scale feature extraction. A multi-scale convolutional neural network is applied to extract features from the collected images, generating feature maps of different scales. The network contains four convolutional layers, each followed by batch normalization and ReLU activation function, respectively extracting local texture features, medium-scale shape features and global spatial distribution features. Through multi-scale feature extraction, the system can capture both microscopic morphology and macroscopic distribution characteristics of garlic plants.

[0067] The detailed structure of the multi-scale convolutional neural network is as follows: input layer: receiving the original RGB image (size 640x480 pixels) from the camera; first convolution unit: containing 32 3x3 convolution kernels, extracting local texture features, followed by batch normalization layer and ReLU activation function, and finally down-sampling through 2x2 max-pooling layer; second convolution unit: containing 64 3x3 convolution kernels, extracting medium complexity shape features, also followed by batch normalization layer and ReLU activation function, and down-sampling through 2x2 max-pooling layer; third convolution unit: containing 128 3x3 convolution kernels, capturing more abstract features, followed by batch normalization layer and ReLU activation function, and down-sampling through 2x2 max-pooling layer; fourth convolution unit: containing 256 3x3 convolution kernels, extracting high-level semantic features and global spatial relationships, followed by batch normalization layer and ReLU activation function.

[0068] The multi-scale convolutional neural network can focus on both local detail features (such as leaf texture and shape) and global distribution patterns (such as row spacing and plant arrangement) of garlic plants by adjusting the receptive field size of different levels of features, providing comprehensive feature representation for subsequent density estimation.

[0069] Step 1.2, density feature aggregation and analysis. The multi-scale features are fused through a feature pyramid network, combined with an attention mechanism to highlight key area features, generating a preliminary density feature map. During feature fusion, an adaptive weight allocation strategy is used to automatically adjust the weights of different scale features according to the current scene conditions, improving the system's adaptability to various plant density patterns.

[0070] The specific implementation of the feature pyramid network is as follows: a top-down path: high-level semantic features (output of the fourth convolution unit) are restored to the same spatial resolution as low-level features through upsampling; a lateral connection: the upsampled high-level features are connected with low-level features of the corresponding resolution (element-level addition or channel-level connection is adopted); and a fused feature processing: the connected features are subjected to 1x1 convolution for channel number adjustment and feature fusion.

[0071] The attention mechanism part adopts a spatial attention module, and the structure thereof comprises: convolution dimension reduction: 1x1 convolution is used to reduce the number of feature map channels; attention map calculation: a spatial attention weight map is generated through a series of convolution operations; and feature weighting: the attention weight map is applied to the original features to highlight key region information.

[0072] The adaptive weight distribution strategy dynamically adjusts the importance of features of different scales based on the complexity of the current scene, the lighting conditions and the plant density distribution, thereby improving the adaptability of the system to a changing environment.

[0073] Step 1.3, density distribution map generation. The fused feature map is processed by a density regression network to generate a plant density distribution map within the current field of view, and the density change trend of the front area is predicted according to a historical image sequence. The density regression network adopts a full convolution structure, and outputs a density distribution map matching the size of the input image, with each pixel value representing an estimated plant density value of the corresponding area.

[0074] The detailed structure of the density regression network is as follows: a feature decoder: comprising four transpose convolution layers, gradually restoring the spatial resolution of the feature map; a first transpose convolution layer: 128 3x3 convolution kernels, a step of 2, and up-sampling the feature map; a second transpose convolution layer: 64 3x3 convolution kernels, a step of 2, and continuing to up-sample; a third transpose convolution layer: 32 3x3 convolution kernels, a step of 2, and further up-sampling; a fourth transpose convolution layer: 16 3x3 convolution kernels, a step of 1, and fine-tuning the features; an output layer: a 1x1 convolution layer, converting the feature map into a single-channel density map, with each pixel value representing an estimated plant density of the corresponding position; and an activation function: a ReLU activation function is adopted to ensure that the density estimation result is non-negative.

[0075] The density change trend prediction of the front area adopts a recurrent neural network structure, and specifically: input: a sequence of 5 consecutive density distribution maps; feature extraction: 3D convolution is used to extract spatio-temporal features; time series modeling: LSTM (Long Short-Term Memory Network) units are used to process time series information; and prediction output: a fully connected layer generates a density change trend map within the next 3 seconds.

[0076] To verify the effectiveness of the plant density estimation network in the technical solution, the following experiments are performed: Figure 2As shown, the accuracy of the traditional density estimation method and the deep learning density estimation network of the present technical solution under different lighting conditions is compared, the density estimation accuracy of the present technical solution is better than that of the traditional method under various lighting conditions, especially under weak light conditions, the accuracy is improved more obviously, which verifies the effectiveness of multi-scale feature extraction and attention mechanism.

[0077] Step 1.4, density change rate calculation. Based on the time series analysis of the density distribution map of the continuous frames, the spatial gradient and change rate of the plant density are calculated to provide the basis for subsequent speed adjustment. By analyzing the density change between continuous frames through the optical flow algorithm, the system can predict the density change trend of the front area and realize forward-looking control.

[0078] Step 2, construct a density speed coupling dynamics model based on the plant density distribution map, and solve the optimal travel speed of the garlic harvester under the current plant density through a nonlinear mapping function and an optimization algorithm. A mathematical relationship model between plant density, travel speed and control accuracy is constructed, which can quantitatively describe the optimal travel speed under different plant density conditions.

[0079] Further, the process of obtaining the optimal travel speed of the garlic harvester under the current plant density includes: constructing a system adaptability index including efficiency weight coefficient, speed penalty coefficient, accuracy weight coefficient and acceleration compensation coefficient based on the plant density distribution map, and determining the density speed coupling dynamics model based on the system adaptability index; using a nonlinear mapping function to normalize the plant density, and quantifying the interaction between plant density and travel speed through a density speed interaction function; based on the interaction between plant density and travel speed, using gradient descent method to optimize and correct the parameters of the density speed coupling dynamics model; based on the optimized parameters, the travel speed that optimizes the system adaptability index is solved through numerical optimization method as the optimal travel speed under the current plant density. Specifically, the following sub-steps are included.

[0080] Step 2.1, system adaptability index definition. The system adaptability index is defined to comprehensively evaluate the overall performance of the system under the condition of a specific plant density and travel speed :

[0081] ;

[0082] Wherein is the system adaptability index; is the plant density; is the travel speed; , , , respectively represent the efficiency weight coefficient, speed penalty coefficient, precision weight coefficient and acceleration compensation coefficient; is the plant density normalization mapping function; is the density-speed interaction function; is the acceleration. The system adaptability index comprehensively considers the two dimensions of control accuracy and operation efficiency, and the larger the value represents the better system performance.

[0083] Step 2.2, nonlinear mapping function determination. The nonlinear mapping function in the model is determined and : The plant density is normalized, and the density values in different ranges are mapped to the interval, which is convenient for subsequent processing. The specific form is:

[0084] ;

[0085] wherein, is the plant density normalization mapping function; is the plant density; is the base number of natural logarithm; is the adjustment parameter of the slope of the control function; is the reference density value.

[0086] The interactive influence of plant density and travel speed is described in the form of a quadratic function:

[0087] ;

[0088] wherein, is the density-speed interaction function; is the plant density; is the travel speed; , , respectively represent the density linear influence coefficient, speed linear influence coefficient and density-speed quadratic interaction coefficient.

[0089] Step 2.3, parameter optimization and correction. Through machine learning method, based on the actual operation data, the model parameters , , , and , , , are optimized and corrected, so that the model can accurately reflect the performance characteristics of the specific garlic harvester. The optimization adopts gradient descent method, and the parameters are updated by minimizing the mean square error between the predicted speed and the actual optimal speed.

[0090] To verify the effectiveness of the density-velocity coupled dynamics model, the system adaptability index analysis results under different plant densities and travel speeds are shown in FIG. 2, and the system adaptability index Figure 3 shows an obvious regular distribution under different density-speed combinations, proving that the density-speed coupled model can accurately quantify the system performance under different conditions, providing a reliable theoretical basis for solving the optimal speed.

[0091] Step 2.4, optimal speed solving algorithm. Based on the established density-speed coupled dynamics model, for any given plant density , the optimal speed is solved, which is the optimal speed under the current density condition. The solving process uses numerical optimization methods to find the optimal solution under the condition of meeting the physical constraints of the garlic harvester.

[0092] The calculation process of the optimal speed is as follows:

[0093] Objective function setting: the system adaptability index defined in step 2.1 is taken as the optimization goal, and the speed that maximizes its value is sought.

[0094] Constraint condition: set the harvester physical constraint conditions, including the minimum and maximum allowed speed range and the maximum acceleration limit, to ensure that the solution meets the actual operation requirements.

[0095] Gradient solving: use the nonlinear mapping functions and determined in step 2.2 to take the partial derivative of the system adaptability index , obtaining the gradient information about the speed .

[0096] Iterative optimization: use the gradient ascent method for iterative solving, starting from the initial speed value, gradually adjusting the speed value along the gradient direction until it converges to the optimal solution or reaches the maximum iteration number.

[0097] Constraint processing: check the speed value obtained by iterative solving to ensure that the final optimal speed is within the allowed physical range.

[0098] Through the above calculation process, the system can quickly calculate the corresponding optimal speed according to the real-time detected plant density, realizing the dynamic coordination of density and speed.

[0099] ​​Step 3, based on the output results of the optimal travel speed and the density-velocity coupling dynamics model, set the compensation algorithm of the influence of speed change on visual perception, and ensure the stable recognition performance of the visual system through dynamic adjustment of adaptive image preprocessing and feature extraction parameters. Based on the optimal speed instruction output in step 2 and the current speed change trend, develop a compensation algorithm for the influence of speed change on visual perception, compensate for visual perception, and obtain compensated visual perception information to ensure that the visual system can maintain stable recognition performance at different travel speeds.

[0100] Further, the process of obtaining compensated visual perception information includes: establishing a speed-image blur mapping model based on the optimal travel speed and image quality; performing adaptive preprocessing on the collected images based on the current travel speed and the speed-image blur mapping model to obtain image quality evaluation results; and dynamically adjusting the feature extraction parameters of the density estimation network based on the current travel speed and the image quality evaluation results to obtain compensated visual perception information. Specifically, the following sub-steps are included:

[0101] Step 3.1, speed-image relationship modeling. Analyze the relationship between travel speed and image quality, establish a speed-image blur mapping model, and quantify the image change characteristics at different speeds.

[0102] The establishment process of the speed-image blur mapping model is as follows: first, define the image blur evaluation function :

[0103] ;

[0104] wherein, is the comprehensive evaluation index of image blur; is the travel speed; is the speed-dependent blur kernel standard deviation; is the image contrast index; is the texture degradation index; is the blur weight coefficient; is the contrast weight coefficient; is the texture weight coefficient.

[0105] Second, establish the mapping relationship between speed and blur kernel standard deviation:

[0106] ;

[0107] wherein, is the speed-dependent blur kernel standard deviation; is the travel speed; is the static blur reference value; is the exposure time; is the lens focal length; is the angle between the moving direction and the imaging plane; is a sinusoidal function.

[0108] The relationship model between speed and image contrast is established:

[0109]

[0110] wherein, is the image contrast index; is the moving speed; is the reference contrast in static state; is the contrast decay coefficient; is the base of natural logarithm.

[0111] The relationship model between speed and texture degradation is established:

[0112]

[0113] wherein, is the texture degradation index; is the moving speed; is the texture degradation parameter; is the texture degradation exponential parameter. Through the control variable experiment, the image samples under different speeds are collected, and the rules of image definition, contrast, texture characteristics and other parameters changing with speed are analyzed to construct the prediction model.

[0114] Step 3.2, adaptive image preprocessing. According to the current moving speed and the speed-image relationship model established in step 3.1, the collected images are adaptively preprocessed, including dynamically adjusting the exposure time, applying the motion blur compensation filter, adaptively enhancing the contrast, etc., to improve the image quality. The preprocessing parameters are automatically determined by the speed-image relationship model, realizing the dynamic optimization of image quality.

[0115] Step 3.3, dynamic adjustment of field of view. According to the moving speed instruction output by step 2, the field of view and focus point of the vision system are dynamically adjusted. When the speed is , it is defined as high-speed running mode, at this time the front detection distance is expanded to 8 to 12 meters, the field of view angle is adjusted to 45 to 60 degrees, and the long-distance path planning is focused on; when the speed is , it is defined as medium-speed running mode, the front detection distance is set to 4 to 8 meters, and the field of view angle is 30 to 45 degrees, balancing the long-distance and short-distance information; when the speed is ​​When the speed is defined as the low-speed driving mode, the near-distance detail analysis is enhanced, the detection distance is focused on the range of 2 to 4 meters, the field of view angle is 15 to 30 degrees, and the local precision control ability is improved. Through the adjustable focal length camera or the multi-camera array, the dynamic adjustment of the field of view range is realized, and the optimal visual information can be obtained at different speeds.

[0116] Step 3.4, adaptive adjustment of feature extraction parameters. Based on the current speed and the image quality evaluation result output by step 3.2, the feature extraction parameters of the density estimation network are dynamically adjusted, including the convolution kernel size, the pooling strategy, the attention mechanism weight, etc., so that the network can adapt to the image characteristics at different speeds. Through the feedback mechanism, the parameter configuration is continuously evaluated and optimized to realize the stability of feature extraction.

[0117] As shown in Figure 4 , the influence of speed change on visual perception performance and the effect of compensation mechanism are shown. Without compensation, the image quality decreases significantly with the increase of speed, while after using the compensation mechanism of the present technical solution, the image quality remains stable within a wide speed range, proving the effectiveness of the compensation measures such as adaptive image preprocessing, dynamic adjustment of field of view range and adaptive adjustment of feature extraction parameters.

[0118] Step 4, based on the output results of the compensation algorithm, i.e. the compensated visual perception information and the output results of the optimal driving speed, a multi-time scale predictive control framework is constructed, and a control strategy is obtained based on the multi-time scale predictive control framework to coordinate short-term accurate control and long-term path planning. Based on the optimal speed sequence output by step 2 and the optimized visual perception information of step 3, a multi-time scale predictive control framework is introduced to coordinate short-term accuracy control and long-term path planning, so that the garlic harvester can work stably and efficiently in complex environments. Specifically, the following sub-steps are included.

[0119] Step 4.1, control domain division. The control problem is divided into a short-term accurate control domain of millisecond level and a long-term planning control domain of second level, corresponding to near-distance row deviation correction and long-distance path planning respectively. The short-term control domain mainly deals with immediate deviation correction to ensure that the harvester accurately tracks the garlic row at the current position; the long-term planning control domain focuses on the path direction and plant density changes ahead to prepare for future speed and direction adjustment.

[0120] The control domain is divided into three regions according to the spatial domain, and the specific implementation includes: near-field region (0 to 2 meters): responsible for immediate accurate control, response time less than 50 milliseconds; mid-field region (2 to 5 meters): responsible for medium-term path planning, response time 50 to 500 milliseconds; far-field region (5 to 10 meters): responsible for long-term strategy planning, response time 0.5 to 3 seconds.

[0121] The control domain is divided by time domain, and the specific implementation of the control delay and prediction window division includes: short-term time domain: 0 to 200 milliseconds, corresponding to current state feedback control; medium-term time domain: 200 milliseconds to 2 seconds, corresponding to forward-looking path planning; long-term time domain: 2 seconds to 10 seconds, corresponding to long-term level speed planning.

[0122] Control weight distribution: control instructions in different space-time domains dynamically allocate weights according to the current state, for example, increasing the weight of far-field control in high-speed mode, and increasing the weight of near-field control in high-precision mode.

[0123] Step 4.2, hierarchical prediction model construction. Construct prediction models of different time scales, including: short-term model: based on the current state and control input, predict the system state change within 50 to 200 milliseconds, for immediate accurate control; medium-term model: combine the current trajectory and plant density distribution, predict the optimal path and speed adjustment within 0.5 to 2 seconds; long-term model: use the global density distribution map to plan the optimal travel strategy within 3 to 10 seconds.

[0124] The prediction model uses a model predictive control framework to continuously update the control strategy through a rolling optimization approach.

[0125] The specific structure of the hierarchical prediction model is as follows:

[0126] (1) Short-term prediction model. Dynamics model: use a simplified vehicle kinematics model, including position, heading angle, speed, and steering angle state variables; sampling frequency: 50 Hz (update every 20 milliseconds); prediction window: 10 frames (predict 200 milliseconds in the future).

[0127] Objective function:

[0128] ;

[0129] Where, is the short-term prediction objective function; is the number of short-term prediction window steps; denotes the summation operator; is the lateral deviation weight coefficient; is the lateral deviation of the th step; is the heading deviation weight coefficient; is the heading deviation of the th step; is the steering smoothing weight coefficient; is the steering angle change of the th step.

[0130] Constraint conditions:

[0131] ;

[0132] ;

[0133] ;

[0134] wherein, is a maximum steering angle limit; is a maximum steering angle change limit; is a maximum travel speed limit; , , denote absolute values of steering angle, steering angle change and travel speed, respectively;

[0135] (2) Mid-term prediction model. Dynamics model: combined vehicle kinematics and simplified dynamics model, considering sideslip and acceleration limit; sampling frequency: 10 Hz (updated every 100 milliseconds); prediction window: 20 frames (predicting 2 seconds into the future).

[0136] Objective function:

[0137] ;

[0138] wherein, is a mid-term prediction objective function; is a mid-term prediction window step number; denotes a summation operator; is a trajectory accuracy weight coefficient; is a trajectory tracking error at step ; is a path smoothness weight coefficient; is a path curvature at step ; is a comfort weight coefficient; is a lateral acceleration at step ;

[0139] Constraint conditions:

[0140] ;

[0141] ;

[0142] ;

[0143] wherein, is a maximum lateral acceleration limit; is a maximum path curvature limit; is a minimum travel speed limit; is a maximum travel speed limit; , , denote the absolute values of lateral acceleration, path curvature, and travel speed, respectively;

[0144] (3) Long-term prediction model. Dynamics model: simplified point-mass model, mainly focusing on speed planning; sampling frequency: 1 Hz (updated once per second); prediction window: 10 frames (predicting 10 seconds into the future).

[0145] Objective function:

[0146] ;

[0147] wherein, is the long-term prediction objective function; is the number of long-term prediction window steps; denotes the summation operator; is the system suitability index at step is the plant density at step is the travel speed at step is the travel speed at step is the travel speed at step is the travel speed at step is the speed smoothing weight coefficient; is the travel speed at step

[0148] Constraints:

[0149] ;

[0150] ;

[0151] ;

[0152] wherein, is the minimum travel speed limit; is the travel speed at step is the maximum travel speed limit; is the absolute value of the travel speed at step and the travel speed at step ; is the travel speed at step is the maximum allowed speed change amount; denotes the summation operator; is the number of long-term prediction window steps; is the time interval at step is the maximum allowed job time.

[0153] ​​​​The rolling optimization process of the prediction model includes: predicting the future state → solving the optimization problem → executing the first step of the optimal control sequence → obtaining the new state → re-predicting and optimizing, and repeating the above steps to achieve closed-loop control.

[0154] Step 4.3, multi-scale control strategy coordination. A multi-scale control strategy coordination mechanism is designed to ensure that control decisions at different time scales are coordinated and consistent. A hierarchical decision structure is adopted, with long-term planning providing target trajectories and speed range constraints for medium-term control, medium-term control providing reference paths and speed change trends for short-term control, and short-term control executing specific control instructions to achieve seamless layered control.

[0155] Detailed implementation of the control strategy coordination mechanism:

[0156] Hierarchical information transmission mechanism: a bidirectional information flow design is adopted, including top-down transmission, medium-term to short-term transmission, and bottom-up feedback; among them, the top-down transmission process includes the transmission of the speed target interval and the general path direction set by the long-term planning to the medium-term controller; the medium-term to short-term transmission process includes the generation of the reference path by the medium-term controller to provide the short-term controller for tracking; the bottom-up feedback process includes the actual execution effect of the short-term controller feedback to the upper controller for adjustment and optimization.

[0157] Constraint hardening mechanism: as the control level goes from high to low, the constraint conditions gradually harden, from soft constraints in long-term planning, to more explicit hard constraint ranges in medium-term control, and finally to precise control in short-term control; among them, the target speed and path direction in the soft constraints of long-term planning can be adjusted moderately according to the situation.

[0158] Buffer management: the system sets control buffers between the controllers at each level, including a speed buffer to store the speed trajectory generated by the long-term planning for the medium-term controller to reference, and a path buffer to store the path points generated by the medium-term planning for the short-term controller to track, to ensure smooth transition of control signals.

[0159] Weight distribution algorithm: dynamically distribute the weights of each controller according to the current operation state, including plant density weight, speed weight, and environment weight. Among them, the plant density weight includes increasing the weight of the short-term controller in high-density areas to ensure accurate row tracking, and increasing the weight of the long-term controller in low-density areas to improve operation efficiency; the speed weight includes increasing the weight of the medium-term controller at high speed to avoid sharp changes in direction, and increasing the weight of the short-term controller at low speed to improve tracking accuracy; the environment weight includes automatically increasing the weight of the short-term controller under complex terrain or adverse weather conditions to ensure operation safety.

[0160] Conflict resolution strategy: When different level controllers generate conflicting control instructions, a hierarchical arbitration mechanism is adopted, including safety priority, precision priority, efficiency balance, and smooth transition. Among them, safety priority includes that safety constraints have the highest priority, and any control instruction that may cause danger is immediately rejected; precision priority includes that the precision of line tracking is prioritized under the premise of safety, and the accurate instruction of short-term controller is superior to the rough instruction of long-term planning; efficiency balance includes selecting the control strategy that can improve the overall operation efficiency under the condition that the precision requirement is met; smooth transition includes using gradual transition method to avoid control instruction mutation when switching control strategy.

[0161] Adaptive coordination mechanism: The system automatically adjusts the coordination strategy according to the actual operation effect, and realizes the continuous optimization of system performance through performance monitoring, parameter tuning and strategy learning. Among them, performance monitoring includes real-time monitoring of the execution effect of each level controller, including tracking accuracy, response speed, stability and other indicators; parameter tuning includes automatically adjusting the weight distribution parameters and conflict resolution threshold according to the performance monitoring results; strategy learning includes learning the optimal coordination strategy from historical operation data through machine learning method and continuously updating the coordination algorithm parameters.

[0162] Step 4.4, control smooth transition and constraint processing. Implement control smooth transition and constraint processing mechanism to ensure the continuity and physical feasibility of control instructions. Control smooth transition handles the mutation of control variables by designing a smoothing function to avoid sudden acceleration or sudden turning of the harvester; constraint processing considers the physical limitations of garlic harvester (such as maximum turning angle, maximum acceleration, etc.) to ensure that the generated control instructions are within the physically feasible range.

[0163] The specific implementation of control smooth transition and constraint processing includes:

[0164] Smooth function design; speed smoothing, S-shaped speed curve is adopted, smooth acceleration and deceleration is realized by jerk limitation, the speed change function is defined as:

[0165] ;

[0166] Among them, is the instantaneous speed at time ; is the initial speed; is the target speed; is a natural constant; is a smoothing coefficient, controlling the steepness of the transition, usually taking a value of 3-5; is the current time variable; is the midpoint of the transition time.

[0167] Jerk constraint is:

[0168] ;

[0169] where, is the displacement is the third derivative of time , i.e. jerk; is the absolute value operator; is the maximum jerk limit, usually set to 0.5 m / s³.

[0170] Steering smoothing, exponential moving average filter is used to process steering command, reduce steering jitter, filter function is:

[0171] ;

[0172] where, is the smoothed steering angle at step ; is the discrete time step index; is the smoothing factor, control the filter strength; is the original steering command at step ; the smoothing factor ranges from 0.1 to 0.3, the higher the density , the smaller the value.

[0173] Initial conditions:

[0174] ;

[0175] where, is the smoothed steering angle at initial time; is the original steering command at initial time;

[0176] Physical constraint handling includes: (1) speed constraint: considering the maximum / minimum speed limit of garlic harvester; (2) acceleration constraint: dynamically adjusting the maximum acceleration according to vehicle performance and ground conditions; (3) steering constraint: considering the maximum steering angle and steering rate, ensuring that the steering command is within the feasible range; (4) obstacle avoidance constraint: ensuring that the control command will not cause the machine to collide with the plant or leave the garlic row; (5) constraint softening technique: for multiple constraints that are difficult to satisfy simultaneously, a relaxation variable and penalty function method is used to find the best compromise solution in the optimization process; (6) safety check module: all control commands are verified by the safety check module before execution, to ensure that the control command will not cause dangerous situations.

[0177] As Figure 5As shown, the distribution of row tracking deviation of the garlic harvester before and after using the multi-time scale predictive control framework is shown. After using the technical solution, the row tracking deviation is reduced and the distribution is more concentrated. The median of the deviation is reduced from ±5.2 cm of the traditional method to ±1.8 cm, proving the effect of the multi-time scale predictive control framework in improving control accuracy.

[0178] Step 5, based on the density distribution map, the output result of the compensation algorithm, the output result of the optimal speed, and the multi-time scale predictive control framework, a self-learning optimization system is established to optimize the density-speed mapping relationship by continuously collecting operation data. Based on the output data of the complete control process of steps 1 to 4, the self-learning optimization system is established to continuously optimize the density-speed mapping relationship by continuously collecting actual operation data, so that the system performance gradually improves over time. Based on the actual operation data, the density-speed coupling dynamics model and the control strategy are continuously optimized to realize automatic row control of the garlic harvester, which specifically includes the following sub-steps.

[0179] Step 5.1, operation data collection and labeling. During the operation of the harvester, key operation data is automatically collected and stored, including the plant density distribution generated in step 1, the optimal speed calculated in step 2, the visual compensation effect of step 3, the control instruction execution situation of step 4, and information such as row accuracy and environmental conditions, forming a complete operation data set. The system automatically labels the row control performance indicators under different density-speed combinations to provide training data for subsequent model optimization.

[0180] The operation data collection system is composed of multiple layers, including a data collection layer, a data preprocessing layer, an automatic labeling system, and a data management system. The data collection layer includes integrated vision system images, GPS positioning, IMU attitude, steering angle sensors, speed sensors, and other multi-source data. The vision data sampling frequency is 10 Hz, and the other sensor data is 50 Hz. Distributed caching and persistent storage are combined to ensure data integrity. The data preprocessing layer includes implementing timestamp unification of different data sources, identifying and marking abnormal data points, and lossy compression processing for large-capacity data such as images. The automatic labeling system includes automatically calculating performance indicators such as row deviation mean, standard deviation, speed stability, and energy consumption, classifying data according to environmental conditions such as light, weather, and terrain, and assigning a quality score to each group of data that affects the weight of subsequent learning. The data management system includes supporting the addition of new data to the database in an incremental manner, implementing a hierarchical storage strategy that supports long-term storage of core data and regular cleaning of auxiliary data, and providing interfaces that support efficient time series and semantic queries.

[0181] Step 5.2, Model Adaptive Update. Based on the collected actual operation data, the parameters of the density-velocity coupling dynamic model are updated regularly to enable the model to more accurately reflect the optimal operation state of the current harvester in a specific field. The model update adopts an incremental learning method, giving higher weight to new data, ensuring that the model can adapt to environmental changes in a timely manner.

[0182] The model adaptive update system adopts a comprehensive technical framework, including three core components: incremental learning framework, parameter update strategy, and adaptive adjustment mechanism. The incremental learning framework automatically adjusts the update frequency according to the operation time, adopts a learning rate adjustment mechanism that gradually decreases with the improvement of system stability, and prevents overfitting through L2 regularization to preserve the model's generalization ability; the parameter update strategy adopts a window weighting method, i.e., recent data are given higher weight and historical data weight decays exponentially over time, combined with batch update technology, while the verification check mechanism ensures the quality of the update; the adaptive adjustment mechanism realizes the dynamic optimization of the system through the environment perception function and performance monitoring system. Among them, the process of automatically adjusting the update frequency includes triggering usually every 1-2 hours of operation or accumulating new data exceeding the threshold; the batch update technology includes grouping parameters and adopting an alternating update strategy to ensure system stability; the verification check mechanism includes verifying the updated model with the historical "golden data set", and rolling back the update if performance declines; the environment perception function includes detecting environmental changes such as weather and terrain changes, triggering the update of the corresponding parameter group; the performance monitoring system includes; includes continuous monitoring of system performance, triggering emergency update when performance indicators decline beyond the threshold.

[0183] Step 5.3, Control Strategy Evolution. Through reinforcement learning technology, the control strategy is continuously optimized and evolved, and the system can learn from successful control decisions and avoid repeating failed strategies. Real-time feedback obtained during the harvesting process (such as row tracking deviation, speed change, plant damage rate, etc.) is used as a reward signal for reinforcement learning to guide the optimization direction of the control strategy.

[0184] The control policy evolution framework adopts a comprehensive technical architecture, including four key components that work together: a policy representation system, a reinforcement learning algorithm system, an exploration and exploitation balance mechanism, and a safety constraint learning framework. Specifically, the policy representation system expresses the control policy through parameterized functions (such as deep neural networks), defines the state space containing key variables such as plant density, speed, row deviation, and path curvature, and the action space including target speed adjustment and steering angle adjustment. The reinforcement learning algorithm system combines the TRPO (Trust Region Policy Optimization) algorithm to ensure policy update stability, uses the experience replay mechanism to improve sample utilization efficiency, and uses reward shaping technology to convert long-term goals (work efficiency and accuracy) into immediately measurable reward signals. The exploration and exploitation balance mechanism achieves local exploration through Gaussian noise, directional exploration based on historical data, and gradually reduces the exploration proportion as the system performance improves. The safety constraint learning framework clearly defines the safe operation boundary of the control policy, maximizes the reward while meeting the safety constraint conditions, and provides a quick fallback strategy in case of exploration failure, ensuring that the system remains safe and controllable during optimization.

[0185] Step 5.4, knowledge base accumulation and sharing. Build a job knowledge base to record the optimal control policies and parameter configurations for different fields, different garlic varieties, and different weather conditions. The knowledge base supports knowledge transfer across fields and seasons, enabling the system to quickly adapt to new operating environments without starting from scratch. In the scenario of multiple harvesters working together, it supports experience sharing between different machines, accelerating the learning process of the entire fleet.

[0186] The knowledge base system adopts a comprehensive technical architecture, including four interrelated core components, namely, knowledge representation structure, knowledge migration mechanism, distributed knowledge sharing architecture, and knowledge base maintenance mechanism. The knowledge representation structure builds a complete knowledge system through scene description sub, strategy parameter group, performance index, and credibility score; the knowledge migration mechanism realizes cross-scene knowledge application through similarity calculation, parameter adaptation, and migration verification; the distributed knowledge sharing architecture consists of central knowledge base, local knowledge cache, asynchronous update, and conflict resolution, ensuring efficient knowledge sharing; and the knowledge base maintenance mechanism guarantees high-quality operation of the knowledge base through knowledge compression, knowledge elimination, and continuous verification. Among them, the scene description sub includes environmental factors such as terrain features, plant density distribution, and variety characteristics; the strategy parameter group includes optimal control parameters corresponding to specific scenes; the performance index includes the actual performance of the parameter group in a specific scene; the credibility score includes the reliability of the parameter group based on data volume and consistency evaluation; the similarity calculation process includes calculating environmental similarity based on scene description sub; parameter adaptation includes automatically adjusting parameters according to environmental differences; migration verification includes verifying the effectiveness of migrated knowledge in a new environment and recording adaptive adjustments; the central knowledge base includes storing the comprehensive experience of all harvesters; the local knowledge cache includes each harvester maintaining its own commonly used scene knowledge; asynchronous update includes synchronizing local knowledge with the central knowledge base periodically; conflict resolution includes selecting the optimal solution based on performance index and credibility score; knowledge compression includes merging similar parameter settings to reduce redundancy; knowledge elimination includes removing knowledge items that have not been used for a long time and have poor performance; and continuous verification includes periodically verifying the effectiveness of knowledge items and updating the credibility score.

[0187] As shown in Figure 6 , the performance improvement trend of the self-learning optimization system is shown, and the system performance continuously improves with the extension of use time. The performance improves faster at the beginning of the job, and then tends to be stable and maintains a slow upward trend. After 30 days of continuous use, the comprehensive performance of the system is improved by about 18% compared with the initial state, verifying the effectiveness and continuous improvement ability of the self-learning optimization system, and proving the long-term benefits of knowledge base accumulation and sharing mechanism.

[0188] An automatic row control system of a garlic harvester based on machine vision, for executing the automatic row control method of a garlic harvester based on machine vision described above, comprising:

[0189] A density perception module for analyzing image data through a deep learning network, estimating plant density characteristics in real time and generating a density distribution map.

[0190] A density perception module for analyzing plant density characteristics of the current area and the front area in real time based on image data obtained by machine vision, generating a plant density distribution map using a deep learning network.

[0191] A dynamics modeling module is configured to construct a density-velocity coupled dynamics model based on the plant density distribution map, and to solve the optimal traveling speed of the garlic harvester under the current plant density condition through a nonlinear mapping function and an optimization algorithm.

[0192] A vision compensation module is configured to compensate the visual perception based on the optimal traveling speed, using adaptive image preprocessing, dynamic adjustment of the field of view range, and dynamic adjustment of feature extraction parameters, to obtain compensated visual perception information.

[0193] A multi-scale control module is configured to construct a multi-time scale predictive control framework based on the compensated visual perception information and the optimal traveling speed, and to obtain a control strategy based on the multi-time scale predictive control framework.

[0194] A self-learning optimization module is configured to continuously optimize the density-velocity coupled dynamics model and the control strategy based on actual operation data, to realize automatic row control of the garlic harvester.

[0195] The system can adjust the optimal traveling speed in real time according to the plant density through the density-velocity coupled dynamics model, reduce the speed to improve the accuracy in high-density areas, and increase the speed to improve the efficiency in low-density areas; the system can maintain stable recognition performance under different light conditions, different plant densities, and different speeds through the compensation mechanism of vision perception for speed changes; the system can simultaneously consider short-term accuracy control and long-term path planning through the multi-time scale predictive control framework, thereby improving the accuracy and stability of the row control; the harvester can continuously learn and optimize from actual operation through the self-learning optimization system, and the system performance is continuously improved with the use time; the system replaces the traditional multiple sets of speed control and path detection strategies through the unified density-speed coordination framework, thereby reducing the complexity of the control system, improving the system adaptability, and reducing the difficulty of system maintenance and configuration.

[0196] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for automatic row control of a garlic harvester based on machine vision, characterized in that: The following steps are involved: Based on the image data obtained by machine vision, a deep learning network is used to analyze the plant density characteristics of the current area and the area ahead in real time to generate a plant density distribution map; A density-speed coupling dynamic model is constructed based on the plant density distribution map, and the optimal travel speed of the garlic harvester under the current plant density condition is solved by a nonlinear mapping function and an optimization algorithm; Based on the optimal travel speed, adaptive image preprocessing, dynamic adjustment of the field of view, and dynamic adjustment of feature extraction parameters are used to compensate for visual perception, thereby obtaining compensated visual perception information; constructing a multi-time-scale predictive control framework based on the compensated visual perception information and the optimal travel speed, and obtaining a control strategy based on the multi-time-scale predictive control framework; Based on actual operation data, the density-speed coupling dynamic model and control strategy are continuously optimized to achieve automatic row control of the garlic harvester.

2. The automatic row control method for garlic harvester based on machine vision according to claim 1, wherein Based on image data acquired by machine vision, a deep learning network is used to analyze plant density characteristics in the current area and the area ahead in real time. The process of generating a plant density distribution map includes the following: Extract features from the collected images based on a multi-scale convolutional neural network to generate multi-scale features; The multi-scale features are fused through the feature pyramid network, and the key area features are highlighted in combination with the attention mechanism to generate a preliminary density feature map; The preliminary density feature map is processed by a density regression network to generate a plant density distribution map within the current field of view.

3. The automatic row control method for garlic harvester based on machine vision according to claim 1, wherein The process of obtaining the optimal travel speed of the garlic harvester under the current plant density conditions includes: constructing a system adaptability index including an efficiency weight coefficient, a speed penalty coefficient, an accuracy weight coefficient, and an acceleration compensation coefficient based on the plant density distribution map, and determining a density-speed coupling dynamics model based on the system adaptability index; A nonlinear mapping function was used to normalize plant density, and the interaction between plant density and travel speed was quantified using a density-speed interaction function. Based on the interactive effect between plant density and travel speed, the gradient descent method is used to optimize and correct the parameters of the density-speed coupling dynamic model; Based on the optimized parameters, the numerical optimization method is used to solve the travel speed that optimizes the system adaptability index, which is used as the optimal travel speed under the current plant density conditions.

4. The automatic row control method for garlic harvester based on machine vision according to claim 3, wherein The expression of the system adaptability index is: ； Where, is the system adaptability index; is the plant density; is the travel speed; 、 、 、 They represent the efficiency weight coefficient, speed penalty coefficient, accuracy weight coefficient, and acceleration compensation coefficient respectively; is the normalized mapping function of plant density; is the density-velocity interaction function; is the acceleration, and · represents the dot product.

5. The automatic row control method for garlic harvester based on machine vision according to claim 4 is characterized in that, The expression for quantifying the interactive effect of plant density and travel speed through the density-speed interaction function is: ； Where, is the density-velocity interaction function; is the plant density; is the travel speed; 、 、 They represent the density linear influence coefficient, velocity linear influence coefficient, and density-velocity quadratic interaction coefficient, respectively. · represents the dot product.

6. The automatic row control method for garlic harvester based on machine vision according to claim 1, characterized in that: Based on the optimal travel speed, adaptive image preprocessing, dynamic adjustment of the field of view, and dynamic adjustment of feature extraction parameters are used to compensate for visual perception, and the process of obtaining compensated visual perception information includes: establishing a speed image ambiguity mapping model based on the optimal traveling speed and image quality; Adaptively preprocessing the collected image based on the current traveling speed and the speed image ambiguity mapping model to obtain an image quality assessment result; The feature extraction parameters of the density estimation network are dynamically adjusted based on the current travel speed and the image quality evaluation result to obtain compensated visual perception information.

7. The method for automatic row control of a garlic harvester based on machine vision according to claim 6, wherein: The process of establishing the velocity image ambiguity mapping model includes: Establish a mapping relationship between speed and blur kernel standard deviation; Model the relationship between speed and image contrast; Establish a model of the relationship between speed and texture degradation; A velocity image fuzziness mapping model is established based on the mapping relationship between velocity and blur kernel standard deviation, the relationship model between velocity and image contrast, and the relationship model between velocity and texture degradation.

8. The method for automatically controlling rows of a garlic harvester based on machine vision according to claim 7, wherein: The process of building a multi-timescale predictive control framework includes: The control domain is divided into short-term domain, medium-term domain and long-term domain according to the time domain; Constructing a hierarchical prediction model based on the short-term time domain, the medium-term time domain, and the long-term time domain; Obtaining control decisions at different time scales based on the hierarchical prediction model; The control decisions at different time scales are coordinated through a hierarchical decision structure and constraint hardening mechanism to obtain the final control instructions.

9. The automatic row control method for garlic harvester based on machine vision according to claim 8, characterized in that: The process of continuously optimizing the density-speed coupling dynamics model and control strategy based on actual operation data to achieve automatic row control of the garlic harvester includes: Collect and store plant density distribution, optimal travel speed, visual compensation parameters, and control execution effect data during the operation in real time to build an annotated operation dataset; An incremental learning method is used to update the parameters of the density-velocity coupled dynamics model, and new weights are assigned through a window weighting strategy. Based on the reinforcement learning framework, the multi-timescale predictive control framework is optimized using line tracking deviation and operation efficiency as reward signals; Based on the new weights and the optimized multi-time-scale predictive control framework, a knowledge base containing the characteristics of different operation scenarios is established. Based on the knowledge base containing the characteristics of different operation scenarios, adaptive migration of control parameters is achieved through similarity calculation, and optimization experience is shared during multi-machine collaborative operation.

10. A garlic harvester automatic row control system based on machine vision, characterized in that: For implementing the method according to claim 1, the system comprises: The density perception module is used to analyze the plant density characteristics of the current area and the area ahead in real time based on image data acquired by machine vision using a deep learning network to generate a plant density distribution map; A dynamic modeling module is used to construct a density-speed coupling dynamic model based on the plant density distribution map, and solve the optimal travel speed of the garlic harvester under the current plant density conditions through a nonlinear mapping function and an optimization algorithm; a visual compensation module for compensating visual perception based on the optimal travel speed by using adaptive image preprocessing, dynamic adjustment of the visual field, and dynamic adjustment of feature extraction parameters to obtain compensated visual perception information; a multi-scale control module, configured to construct a multi-time-scale predictive control framework based on the compensated visual perception information and the optimal travel speed, and obtain a control strategy based on the multi-time-scale predictive control framework; The self-learning optimization module is used to continuously optimize the density-speed coupling dynamic model and control strategy based on actual operation data to achieve automatic row control of the garlic harvester.

Images