Residual fusion attention mechanism-based rhizome crop field visual identification algorithm and system

Abstract

The invention provides a rhizome crop field visual identification algorithm and system based on a residual fusion attention mechanism, and the system comprises a four-layer structure: the first layer is a sensing layer which comprises two industrial area-array cameras which form a binocular vision system for collecting a front field image; the second layer is a control layer, a Google fourth-generation tensor processor TPU-v4 is used as a main control core of the tracked vehicle, and an ROS 2 is used as a main control system and is responsible for operating a related network algorithm to carry out image processing and path planning; the third layer is a driving layer which mainly comprises a driving controller and is used for receiving a signal of the main control core and sending a driving signal to each executing mechanism; and the fourth layer is an execution layer, comprises a crawler chassis and related mechanical parts and is used for driving the vehicle to run.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and machine learning, and in particular to the technology of applying visual recognition in the field of agriculture, and more particularly to a visual recognition algorithm and application of a root crop field based on a residual fusion attention mechanism. Technical Background

[0002] With the continuous improvement of agricultural automation and intelligence, the application of machine vision technology in the field of automatic navigation in agricultural production is becoming more and more extensive.

[0003] Due to the unique terrain characteristics of ridge-field-ridge-ridge arrangement in the root and tuber crop fields, agricultural vehicles need to drive accurately along the ridges to successfully complete operations such as sowing, fertilization, pruning, and harvesting. Currently, the commonly used crop field visual recognition technology mainly relies on traditional image processing methods and simple machine learning models. Although these methods can meet basic recognition needs to a certain extent, they have the following shortcomings when facing the above challenges.

[0004] Due to the characteristics of root crops, the aboveground parts are relatively short and lack features, making them difficult to identify and capture in complex backgrounds (especially weeds with similar appearance). Most existing algorithm architectures are relatively traditional and cumbersome. When dealing with such complex scenes, a lot of computing resources are wasted, resulting in slow recognition response speeds. At the same time, they are unable to accurately extract truly useful image features, which is not conducive to subsequent decision-making. How to quickly and effectively distinguish and identify the ridges of crops and the furrows where weeds grow is a major challenge. At the same time, general visual cameras are more sensitive to operating environment conditions such as lighting, weather, and color, and the imaging quality is poor. Conventional visual algorithms are easily affected by similar interference, which affects the recognition accuracy and has poor adaptability to the environment. How to design an efficient and reliable preprocessing method to improve the image quality captured by visual cameras in various complex working conditions and the recognition coordination of visual algorithms is a difficult point.

[0005] How to improve the recognition accuracy of the visual system while reducing the amount of calculation and increasing the response speed has become a key issue that needs to be solved urgently. By solving these key pain points, we can further promote the application of visual recognition technology on mobile platforms, especially small and medium-sized agricultural vehicles, making field automation operations more efficient. Summary of the invention

[0006] The purpose of the present invention is to provide a root crop field visual recognition algorithm and system based on the residual fusion attention mechanism, so as to overcome the pain points of the existing agricultural vehicle technical solutions in field automated operation navigation, such as low recognition accuracy and reliability, large amount of calculation, slow speed and low efficiency, which make it difficult to continue to develop a series of mechanized and automated operations such as automated sowing and harvesting of root crop fields.

[0007] In order to achieve the above purpose, the present invention adopts the following technical solution, which mainly includes the following steps:

[0008] Step 1: Use Google's fourth-generation tensor processor TPU-v4 as the main control core of the tracked vehicle, ROS 2 as the main control system, and use two industrial array cameras to form a binocular vision system as the environmental perception unit;

[0009] Step 2: Reset the tracked vehicle to the starting point of the operation, that is, the beginning of the first ridge, and use two industrial array cameras to form a binocular vision system to collect visual images of the Panax notoginseng field in front, and transmit the obtained data back to the TPU-v4 main control core;

[0010] Step 3: For the field images captured by the two industrial area array cameras, the main control core first performs a series of preprocessing operations to obtain a high-standard and high-resolution basic base map of the Panax notoginseng field. The preprocessing operation process is as follows:

[0011] In order to offset the color deviation caused by factors such as lighting and lens when the image is taken, the image is light compensated. The formula is as follows, where avg R is the average value of color component R, avg Gray is the average gray value of the image, c(R) is the R value of the original image, and f is the maximum value of the three components R, G, and B of all pixels in the image. The same is true for components G and B:

[0012]

[0013] In order to compensate for the color difference between the captured image and the real scene, the image is interpolated and corrected. The formula is as follows, where the coefficients b, i, and j are determined by the spectral characteristics of the camera parameters, the light source conditions, and the spectral characteristics of the filter:

[0014]

[0015] After color correction, smoothing filtering is performed to eliminate the noise caused by camera quality factors. The formula is as follows, where g(x,y) is the center element of the window, f(x+i,y+j) is the pixel of the noisy image, w(i,j) is the weight value, and the window size is (2k+1)×(2k+1). Arranging the n×n weights into a matrix is ​​called a weighted template. The weighted mean filter formula and weighted template formula are as follows:

[0016]

[0017] Step 4: After image preprocessing is completed, the main control core uses the convolutional neural network DenseNet to perform advanced feature extraction on the preprocessed base map, segmenting and identifying the target ridges and the furrows on both sides of the target ridges. The formula is as follows, where X i , X0, X1, X i-1 Represents a feature map, and [] represents concatenating all output feature layers by channel combination:

[0018] X l =H L ([X0,X1,...,X l-1 ]) (5)

[0019] Among them, H represents the combination of nonlinear transformation BN+ReLU+Conv(1×1)+Conv(3×3), that is, after normalizing the input feature map, the nonlinear factor is introduced through the ReLU activation function, so that the input of each layer has a stable distribution, which accelerates the training process and alleviates the gradient disappearance. Then a 1×1 convolution layer is used to change the depth of the feature map to generate a new feature map, and finally a 3×3 convolution layer is used to extract spatial features. As the network continues to deepen, all the previous layers are densely connected to the later layers, so that feature reuse is achieved and extraction is facilitated.

[0020] At the same time, the residual fusion attention network mechanism is introduced to mine deep feature maps with higher abstraction and semantic expression ability, improve recognition reliability while reducing computational complexity and improving robustness. The implementation method is as follows: Design a stacked network structure composed of multiple layers of attention modules. Each attention module includes two branches: TRUNK branch and MASK branch. The TRUNK branch uses the DAS full convolution algorithm for feature processing and uses separable convolution to improve recognition efficiency. The specific steps are to first compress the input features and then process them by offsetting the dynamic grid. The formula is as follows, where X represents the input features and W represents the depth separable convolution:

[0021] X c =GELU(InstanceNorm(XW1))(6)

[0022]

[0023] The number of channels of the feature map is changed to the original input through layer normalization and Sigmoid activation function. The formula is as follows:

[0024] A=σ(LayerNorm(deform(X C ))) (8)

[0025] The output result A of this formula is the attention gate that controls the information flow from the feature map. The element value in each gate tensor is between 0 and 1, which determines the part of the feature map that is emphasized or filtered. Finally, the original input tensor is multiplied by the attention tensor obtained in the previous step to obtain the final output feature result X out .

[0026] At the same time, in another branch, the MASK branch, the processing of the feature map adopts a bottom-up and top-down parallel structure, that is, the forward downsampling process and upsampling process. While obtaining the global features, the high-dimensional features that have completed the upsampling process are combined with the low-dimensional features that have not completed the downsampling process, thereby enhancing the expressiveness of the feature map. Each completed combination is an attention module. Finally, a Sigmoid activation layer is added to be the output of the MASK branch. Among them, three different activation functions are used to implement three attention mechanisms. Among them, the mixed attention (f1) performs Sigmoid weighting on all spatial positions of all channels without other restrictions. Among them, the spatial attention (f2) uses Ridge regression, that is, L2 regularization for constraint. By reducing the weight size, it helps the MASK branch to better focus on specific areas in the image, reduce the model complexity on the basis of retaining features, reduce the overfitting phenomenon of deep networks, and significantly improve the generalization ability of the model. Among them, the channel attention (f3) uses L2 normalization for constraint, pays more attention to specific channels, and outputs a one-dimensional vector with the same length as the number of channels as the feature weight. The three different attention mechanism formulas are as follows, where x i,c represents the value of the i-th spatial position and the c-th channel in the input feature map, mean c and std c Respectively represent the mean and standard deviation of the cth channel in the input feature map, ||x i || represents the L2 norm of all channels at the i-th spatial position in the feature map:

[0027]

[0028]

[0029] The results of the TRUNK branch and the MASK branch are combined in a certain way to obtain the final output of the attention module. The formula is as follows, where (1+M i,c (x)) represents the weighting of the input features, and M i,c (x) is the value in the interval [0,1] learned by the attention module. The purpose of adding the two is to adaptively adjust the weights according to the input features of different parts, while effectively solving the problem that conventional Sigmoid will reduce the feature value. i,cRepresents the cth feature map of the ith channel of the input feature extracted by the convolutional neural network:

[0030] H i,c (x) = (1 + M i,c (x))*F i,c (x) (12)

[0031] The above attention modules are stacked and iterated. In addition to the normal head-to-tail connection between the modules, an additional jump connection is branched to allow the input to bypass the convolution process and pass directly to the output. A 1×1 convolution layer is used to adjust the dimension. This forms a residual jump connection structure, which alleviates the problem of gradient vanishing in deep networks and improves the network's ability to extract the most effective features.

[0032] Finally, a global average pooling layer and a fully connected layer are used to obtain the final output layer. Through the convolutional neural network DenseNet and the residual fusion attention network mechanism, the image classification and recognition results are finally obtained, that is, the target ridges and the furrows on both sides of the target ridges are identified and marked;

[0033] Step 5: Based on the preprocessing, a high-standard and high-resolution base map of the Panax notoginseng field is obtained, and the target ridges and furrows on both sides are identified and marked by the convolutional neural network, and the distance between the ridges and the position relationship between the vehicle's current position (and the camera shooting position) and the starting point of the furrow are calculated. This step is completed using the principle of triangulation. The formula is as follows, where f is the focal length of the camera in pixels, c is x and c y are the principal point positions of the two cameras, B is the distance between the principal point positions of the two cameras, d is the pixel difference between the two feature points in the horizontal / vertical direction, Z is the object depth, and X and Y are the horizontal and vertical coordinates of the object under test in the camera coordinate system, respectively:

[0034]

[0035] Step 6: According to the calculation results, if the distance between the furrows on both sides is not enough to support the normal passage of the vehicle, the parking alarm is manually controlled. If the vehicle is allowed to pass, a path is planned from the current position to the starting point of the target ridge, and the end position of the path is required to make the two tracks of the tracked vehicle face the furrows on both sides of the target ridge. After the planning is completed, the corresponding command signal is sent to each driver of the drive layer;

[0036] Step 7: After receiving the signal, each driver in the driving layer sends a driving instruction to the corresponding power component in the execution layer, drives the vehicle along the predetermined route to the starting point of the target ridge, and then controls the vehicle to move along the furrows on both sides.

[0037] Move forward at a constant speed;

[0038] Step 8: After the vehicle completes the operation of the target ridge, the ridge is recorded as "completed", and the main control core and drive layer send a command signal to control the vehicle to turn around and reset and repeat the above steps 2 to 7 to operate the next ridge until all ridges are completed. During this period, if there is an impassable road or an abnormal driving situation caused by recognition error, the vehicle will stop and alarm and be handed over to manual remote control to improve the robustness of the system.

[0039] As a preferred solution of the present invention, in step 1, Google's fourth-generation tensor processor TPU-v4 is selected as the main control core of the tracked vehicle. As a latest processor designed for deep learning, TPU can achieve low latency and low power consumption when processing large deep learning models such as visual recognition due to its targeted optimized architecture and large-scale distributed parallel computing. It is very suitable for small and medium-sized agricultural machinery that cannot be equipped with large computers;

[0040] As a preferred solution of the present invention, in step 1, two industrial area array cameras are installed in front of the vehicle to form a binocular vision system. Such a layout can collect the most valuable information while protecting the camera from damage caused by environmental interference.

[0041] As a preferred solution of the present invention, in step 3, a series of preprocessing operations are applied to the captured original image, including light compensation, color interpolation, color correction, and smoothing filtering, so as to improve the image quality and reduce the adverse effects of environmental factors, so as to facilitate the subsequent feature recognition and extraction operations.

[0042] As a preferred solution of the present invention, in step 4, the pre-processed high-quality and high-resolution image is recognized and processed using a DenseNet convolutional neural network and a residual fusion attention network fused with a DAS neural network. The new network architecture can effectively solve the problem of gradient vanishing and signal attenuation that is common in traditional deep networks, while enhancing feature contrast and ignoring unnecessary calculation areas, reducing calculation costs, and greatly improving the efficiency, accuracy, and feature discrimination and decision-making capabilities of image recognition.

[0043] The disclosed embodiment includes a modular residual fusion attention network architecture, which is built by stacking attention modules. The two branches of each attention module respectively use DAS neural network and three (hybrid, spatial, channel) attention mechanisms, and use a modular architecture combining residual modules and attention modules, that is, an additional jump connection is added between each attention module for deep features to be directly transferred to subsequent levels. It can reduce the problem of gradient vanishing signal attenuation, enhance feature contrast while ignoring unnecessary calculation areas, and greatly improve the efficiency and accuracy of image recognition and the ability to distinguish and decide features.

[0044] The beneficial effects of the present invention are:

[0045] ① The present invention is suitable for the automated operation and navigation of small and medium-sized agricultural vehicles along ridges and furrows in root and tuber crop fields;

[0046] ② The image preprocessing process of the present invention can effectively improve the quality of the original image, especially in hilly and mountainous areas where root crops are planted, and the image quality problems caused by lighting, fog, etc., or the shooting equipment itself, are convenient for further recognition and processing of the image;

[0047] ③ The system described in the present invention, due to its unique branch and structural combination form, combined with the characteristics of the processor carrier TPU, can reduce the amount of calculation and calculation cost while ensuring the recognition accuracy, improve the recognition efficiency, and is more suitable for small and medium-sized agricultural vehicles that are inconvenient to carry high-computing power computers;

[0048] ④ The present invention helps realize the automated operation and automatic driving of small and medium-sized agricultural vehicles in root and tuber crop fields, improves the automation level and efficiency of field operations, saves manpower and material resources, and has broad market prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a control flow chart of the present invention;

[0050] Figure 2 It is the overall structural framework diagram of the present invention;

[0051] Figure 3 This is the residual fusion attention network architecture diagram included in the present invention. DETAILED DESCRIPTION

[0052] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments are within the scope of protection of the present invention.

[0053] Example:

[0054] like Figure 1As shown, the present invention provides a root crop field visual recognition algorithm and system based on the residual fusion attention mechanism. The system of the present invention includes a four-layer structure. The first layer is the perception layer, which includes two industrial area array cameras for collecting images of the field in front; the second layer is the control layer, which uses Google's fourth-generation tensor processor TPU-v4 as the main control core of the tracked vehicle, and ROS 2 as the main control system, which is responsible for running related network algorithms for image processing and path planning; the third layer is the drive layer, which mainly includes a drive controller for receiving signals from the main control core and sending drive signals to each actuator; the fourth layer is the execution layer, which includes a track chassis and related mechanical components for driving the vehicle.

[0055] The algorithm of the present invention comprises the following steps:

[0056] Step (1), using Google's fourth-generation tensor processor TPU-v4 as the main control core of the tracked vehicle, ROS 2 as the main control system, and using two industrial array cameras to form a binocular vision system as the environment perception unit;

[0057] Step (2), reset the tracked vehicle to the starting point of the operation, that is, the beginning of the first ridge. Use two industrial array cameras to form a binocular vision system to collect visual images of the Panax notoginseng field in front, and transmit the obtained data back to the TPU-v4 main control core;

[0058] Step (3), for the field images captured by the two industrial array cameras, the main control core first performs a series of preprocessing operations, including light compensation, color interpolation and color correction, and smoothing filtering. Through a series of preprocessing processes, high-quality and high-resolution field images are finally obtained;

[0059] Step (4), the main control core uses the convolutional neural network DenseNet to perform advanced feature extraction on the preprocessed basic base map, segment and identify the target ridges, and the furrows on both sides of the target ridges. At the same time, the residual fusion attention network mechanism is introduced to mine deep feature maps with higher abstraction and semantic expression capabilities, thereby improving recognition reliability while reducing computational complexity and improving robustness. Through the convolutional neural network DenseNet and the residual fusion attention network mechanism, the classification and recognition results of the image are finally obtained, that is, the target ridges and the furrows on both sides of the target ridges are identified and marked;

[0060] Step (5) is to obtain a high-standard and high-resolution base map of the Panax notoginseng field based on preprocessing, and to identify the marked target ridges and furrows on both sides through a convolutional neural network, and calculate the distance between ridges and the position relationship between the vehicle's current position (and the camera shooting position) and the starting point of the furrow. This step is completed using the principle of triangulation;

[0061] Step (6), according to the calculation results, if the distance between the furrows on both sides is not enough to support the normal passage of the vehicle, the parking alarm is manually controlled. If the vehicle is allowed to pass, a path is planned from the current position of the vehicle to the starting point of the target ridge, and the end position of the path is required to make the two tracks of the tracked vehicle face the furrows on both sides of the target ridge. After the planning is completed, the corresponding command signal is sent to each driver of the drive layer;

[0062] Step (7), after receiving the signal, each driver of the driving layer sends a driving instruction to the corresponding power component of the execution layer, drives the vehicle to travel along the predetermined route to the starting point of the target ridge, and then controls the vehicle to move forward at a constant speed along the ridges on both sides at a predetermined speed;

[0063] Step (8), after the vehicle completes the operation of the target ridge, the ridge is recorded as "completed", and the main control core and the drive layer send a command signal to control the vehicle to turn around and reset, repeating the above steps (2) to (7) to operate the next ridge until all ridges are completed. During this period, if an impassable road appears, or if the vehicle is driving abnormally due to an identification error, the vehicle will stop and alarm and be handed over to manual remote control to improve the robustness of the system.

Claims

1. A root crop field visual recognition system based on residual fusion attention mechanism, characterized by: It consists of four layers. The first layer is the perception layer, which includes two industrial area array cameras for collecting images of the field ahead; the second layer is the control layer, which uses Google's fourth-generation tensor processor TPU-v4 as the main control core of the tracked vehicle, and ROS2 as the main control system, responsible for running related network algorithms for image processing and path planning; the third layer is the drive layer, which mainly includes a drive controller for receiving signals from the main control core and sending drive signals to each actuator; the fourth layer is the execution layer, including the track chassis and related mechanical components to drive the vehicle.

2. A root crop field visual recognition algorithm based on residual fusion attention mechanism: It is characterized by The steps include: Step 1: Use Google's fourth-generation tensor processor TPU-v4 as the main control core of the tracked vehicle, ROS2 as the main control system, and use two industrial array cameras to form a binocular vision system as the environmental perception unit; Step 2: Reset the tracked vehicle to the starting point of the operation, that is, the beginning of the first ridge, and use two industrial array cameras to form a binocular vision system to collect visual images of the Panax notoginseng field in front, and transmit the obtained data back to the TPU-v4 main control core; Step 3: For the field images captured by the two industrial array cameras, the main control core first performs a series of preprocessing operations to obtain a high-standard and high-resolution basic base map of the Panax notoginseng field. The preprocessing operation process is as follows: in order to offset the color deviation problem caused by factors such as lighting and lens when the image is captured, the image is light compensated. In order to compensate for the color difference between the captured image and the real scene, the image is color interpolated and color corrected. After color correction, smoothing filtering is performed to eliminate the noise generated by camera quality factors. Through the above preprocessing process, a high-quality and high-resolution field image is finally obtained for subsequent recognition; Step 4: After the image preprocessing is completed, the main control core uses the convolutional neural network DenseNet to perform advanced feature extraction on the preprocessed base map, segment and identify the target ridges and the furrows on both sides of the target ridges. The formula is as follows, where X i , X0, X1, X i-1 Represents a feature map, and [] represents concatenating all output feature layers by channel combination: X l =H L ([X0,X1,...,X l-1 ]) (5) Among them, H L Represents the combination of nonlinear transformations "BN+ReLU+Conv(1×1)+Conv(3×3)", that is: after normalizing the input feature map; introducing nonlinear factors through the ReLU activation function, so that the input of each layer has a stable distribution, accelerating the training process and alleviating the gradient disappearance; then using a 1×1 convolution layer to change the depth of the feature map to generate a new feature map; finally, using a 3×3 convolution layer to extract spatial features. As the network continues to deepen, all the previous layers are densely connected to the later layers, thereby achieving feature reuse and facilitating extraction; At the same time, the residual fusion attention network mechanism is introduced to mine deep feature maps with higher abstraction and semantic expression ability, improve recognition reliability while reducing computational complexity and improving robustness. The implementation method is as follows: a stacked network structure composed of multiple layers of attention modules is designed. Each attention module includes two branches: TRUNK branch and MASK branch. The TRUNK branch uses the DAS full convolution algorithm for feature processing and uses separable convolution to improve recognition efficiency. The specific steps are to first compress the input features and then process them by offsetting the dynamic grid. The formula is as follows, where X represents the input features and W represents the depth separable convolution: X c =GELU(InstanceNorm(XW1)) (6) The number of channels of the feature map is changed to the original input through layer normalization and Sigmoid activation function. The formula is as follows: A=σ(LayerNorm(deform(X C ))) (8) The output result A of this formula is the attention gate that controls the information flow from the feature map. The element value in each gate tensor is between 0 and 1, which determines the part of the feature map that is emphasized or filtered. Finally, the original input tensor is multiplied by the attention tensor obtained in the previous step to obtain the final output feature result; At the same time, in another branch, the MASK branch, the feature map is processed in a bottom-up and top-down parallel structure, that is, the forward downsampling process and upsampling process. While obtaining the global features, the high-dimensional features that have completed the upsampling process are combined with the low-dimensional features that have not completed the downsampling process, thereby enhancing the expression ability of the feature map. Each completed combination is an attention module, and finally a Sigmoid activation layer is added to be the output of the MASK branch. Among them, three different activation functions are used to implement three attention mechanisms. Among them, the mixed attention (f1) is for all spatial positions of all channels. Sigmoid weighting is performed without other restrictions. Spatial attention (f2) is constrained by Ridge regression, that is, L2 regularization. By reducing the weight size, the MASK branch can better focus on specific areas in the image, reduce model complexity on the basis of retaining features, alleviate the overfitting phenomenon of deep networks, and significantly improve the generalization ability of the model. Channel attention (f3) is constrained by L2 normalization, paying more attention to specific channels, and outputting a one-dimensional vector with the same length as the number of channels as feature weighting. The formulas of three different attention mechanisms are as follows, where x i,c represents the value of the i-th spatial position and the c-th channel in the input feature map, mean c and std c Respectively represent the mean and standard deviation of the cth channel in the input feature map, ||x i || represents the L2 norm of all channels at the i-th spatial position in the feature map: The results of the TRUNK branch and the MASK branch are combined in a certain way to obtain the final output of the attention module. The formula is as follows, where (1+M i,c (x)) represents the weighting of the input features, and M i,c (x) is the value in the interval [0,1] learned by the attention module. The purpose of adding the two is to adaptively adjust the weights according to the input features of different parts, while effectively solving the problem that conventional Sigmoid will reduce the eigenvalue. i,c Represents the cth feature map of the ith channel of the input feature extracted by the convolutional neural network: H i,c (x)=(1+M i,c (x))*F i,c (x) (12) The above attention modules are stacked and iterated. In addition to the normal head-to-tail connection between the modules, an additional jump connection is branched to allow the input to bypass the convolution process and pass directly to the output. A 1×1 convolution layer is used to adjust the dimension. This forms a residual jump connection structure, which alleviates the problem of gradient vanishing in deep networks and improves the network's ability to extract the most effective features. Finally, a global average pooling layer and a fully connected layer are used to obtain the final output layer. Through the convolutional neural network DenseNet and the residual fusion attention network mechanism, the image classification and recognition results are finally obtained, that is, the target ridges and the furrows on both sides of the target ridges are identified and marked; Step 5: Based on the preprocessing, a high-standard and high-resolution base map of the Panax notoginseng field is obtained, and the target ridges and furrows on both sides are identified and marked through the convolutional neural network, and the distance between the ridges and the position relationship between the vehicle's current position (and the camera shooting position) and the starting point of the furrow are calculated. This step is completed using the principle of triangulation; Step 6: According to the calculation results, if the distance between the furrows on both sides is not enough to support the normal passage of the vehicle, the parking alarm is manually controlled. If the vehicle is allowed to pass, a path is planned from the current position of the vehicle to the starting point of the target ridge, and the end position of the path is required to make the two tracks of the tracked vehicle face the furrows on both sides of the target ridge. After the planning is completed, the corresponding command signal is sent to each driver of the drive layer; Step 7: After receiving the signal, each driver in the driving layer sends a driving instruction to the corresponding power component in the execution layer, drives the vehicle along the predetermined route to the starting point of the target ridge, and then controls the vehicle to move along the furrows on both sides. Move forward at a constant speed; Step 8. After the vehicle completes the operation on the target ridge, the ridge is recorded as "completed". The main control core and the drive layer send a command signal to control the vehicle to turn around and reset, and repeat the above steps 2 to 7 to operate on the next ridge until all ridges are completed. During this period, if an impassable road appears, or if an identification error causes the vehicle to drive abnormally, the vehicle will stop and alarm and be handed over to manual remote control to improve the robustness of the system.

3. The algorithm according to claim 2, characterized in that: In step 1, the two industrial area array cameras are two high-definition camera lens modules installed in front of the vehicle chassis with a spacing of B. They are used to collect images in front of the vehicle during operation and form a binocular vision system to provide a reference for distance measurement calculations.

4. The algorithm according to claim 2, characterized in that: The DenseNet convolutional neural network and the residual fusion attention network mechanism are used to recognize and process the pre-processed high-quality and high-resolution images. Among them, the DenseNet convolutional neural network has a unique dense connection architecture, that is, the output of the lth layer will not only be used as the input of the l+1 layer, but also as the input of all subsequent layers. Each layer only learns a small number of features, and the later layers can access the features of all previous layers, which makes the processing more efficient and also makes the feature types more diverse. Among them, the residual fusion attention network is built by stacking attention modules. The two branches of each attention module use DAS neural network and three attention mechanisms respectively, and use a modular architecture combining residual modules and attention modules, that is, adding an additional jump connection between each attention module for deep features to be directly transferred to subsequent layers, which can reduce the problem of gradient vanishing signal attenuation, enhance feature contrast while ignoring unnecessary calculation areas, and greatly improve the efficiency and accuracy of image recognition and the ability to distinguish and decide features.

5. The algorithm according to claim 2, characterized in that: A feedback mechanism is introduced to realize closed-loop control. When an abnormal error occurs during the operation process or the system itself, the area where the operation has been completed normally is recorded. After returning to normal, the process of re-setting the positioning can be avoided, and the operation can be automatically continued from the last normal operation location, avoiding complex operation settings and unnecessary repeated operations.

Images