A tractor automatic driving path tracking preview point selection system based on a deep learning network

Abstract

The application discloses a tractor automatic driving path tracking preview point selection system based on a deep learning network, and relates to the technical field of automatic driving path tracking. The system comprises the following steps: step one, building a kinematics model of a tractor test platform; and step two, performing path tracking by using a look-ahead point search network (LPSNet) algorithm. The system utilizes powerful deep learning technology, combines a traditional pure path tracking algorithm, and firstly takes the position and attitude information of a vehicle at a current time and the information of part of a path as input to directly obtain a predicted preview point by using the end-to-end advantage of a neural network. Then, the predicted preview point is used to further determine a path preview point. Finally, the path preview point is brought into the traditional pure path tracking algorithm to obtain a front wheel steering angle of the vehicle at a next time, and the vehicle is controlled to complete path tracking work.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving path tracking technology, specifically to a tractor autonomous driving path tracking preview point selection system based on a deep learning network. Background Technology

[0002] Path tracking control is a key step in realizing tractor automatic navigation technology. It is widely used in agricultural production processes such as tillage, sowing, fertilization, spraying, and harvesting, and is one of the core technologies in precision agriculture. It can reduce crop losses, improve work efficiency, and lower labor costs. However, with the continuous development of automation, precision, and intelligence in agricultural production, the requirements for path tracking accuracy are constantly increasing.

[0003] Pure path tracking algorithms are a typical tracking algorithm based on the geometric relationship between vehicles and paths, widely used in tractor path tracking algorithms. They are characterized by simplicity, accuracy, and efficiency. However, traditional algorithms suffer from the problem of difficulty in determining the aiming distance. Therefore, many scholars have conducted in-depth research on determining the aiming distance, including studies on the influence of vehicle speed on the aiming distance. However, since vehicle behavior does not change proportionally to speed, this method cannot completely solve the problem.

[0004] In summary, vehicle speed, tracking accuracy, steering frequency, vehicle stability, path curvature, and other factors all affect the aiming distance. Pure path tracking algorithms still face the following problems in understanding the relationship between aiming distance and these influencing factors: first, it is difficult to describe clearly and accurately using only mathematical expressions; second, the aiming distance does not fully consider the relationship between the vehicle and the path; and third, a fixed aiming distance needs to be manually specified through experiments or intuitive experience. Summary of the Invention

[0005] This invention provides a path tracking and preview point selection system for autonomous tractor driving based on a deep learning network, in order to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A path tracking preview point selection system for tractor autonomous driving based on deep learning networks, comprising the following steps:

[0008] Step 1: Construction of the kinematic model of the tractor test platform;

[0009] Step 2: Perform path tracking using the Pre-aiming Point Search Network (LPSNet) algorithm. The specific steps are as follows:

[0010] S1. Building the neural network;

[0011] S2, Data Collection;

[0012] S3, the training process of network parameters;

[0013] S4. Extract the aiming point.

[0014] A further improvement to the technical solution of this invention lies in the following: the specific method of step one is to simplify the tractor into a two-wheeled bicycle model in the XOY inertial coordinate system, and obtain the kinematic model of the tractor based on the relationship between the tractor, the tracking path, and the turning radius in the model:

[0015]

[0016]

[0017] Where x, y, δ, φ, L, v, and γ are the abscissa of the tractor, the ordinate of the vehicle, the front wheel steering angle, the vehicle's heading angle, the vehicle's wheelbase, the vehicle speed, and the actual turning radius, respectively. t+Δt It is the target steering curvature of the tractor.

[0018] A further improvement of the technical solution of the present invention is that: S1 further includes constructing a pre-aiming point search network (LPSNet) based on a deep learning neural network (DLNN). The pre-aiming point search network (LPSNet) is mainly composed of FC layers, and each layer is activated by the ReLU function. When the tractor is working, the information obtained at each moment is input into the pre-aiming point search network LPSNet, so as to obtain the pre-aiming point information required to complete the adjustment of vehicle attitude control.

[0019] A further improvement of the technical solution of the present invention is that: S2 further includes establishing a simulation path composed of straight line segments and circular arc segments to simulate most of the situations in which the tractor travels in the field, and the path is composed of discrete points.

[0020] A further improvement to the technical solution of the present invention is that: S3 further includes using the MSE loss function to reflect the distance error between the pre-aiming point predicted by the network and the expected pre-aiming point, specifically expressed as follows:

[0021]

[0022] x in the formula predict ,y predict These are the predicted x and y coordinates of the aiming point obtained through the aiming point search network LPSNet, respectively. expect ,x expect These are the expected values.

[0023] A further improvement of the technical solution of the present invention is that: S4 further includes, in the curved segment, drawing a straight line by connecting the predicted aiming point and the curvature center of the current road segment, the intersection of the straight line and the path is the aiming point; in the straight segment, drawing a perpendicular line from the predicted aiming point to the path, the intersection of the perpendicular line and the path is the aiming point; after determining the aiming point, combining it with the pure path tracking algorithm, the front wheel steering angle control variable for path tracking of the vehicle can be obtained, thereby controlling the steering of the vehicle.

[0024] A further improvement of the technical solution of the present invention is that the pure path tracking algorithm is achieved by calculating the steering angle of the front wheels. The pure path tracking algorithm takes the center point of the rear axle of the vehicle as the tangent point and the vehicle body as the tangent line. By controlling the steering angle of the front wheels, it continuously tries to make the center of the rear axle of the tractor smoothly pass the pre-aiming point along a curve.

[0025] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows:

[0026] 1. This invention provides a path tracking preview point selection system for autonomous tractor driving based on deep learning networks. Utilizing powerful deep learning technology combined with traditional pure path tracking algorithms, it achieves a more intelligent preview behavior, similar to that of a human driving a vehicle, thus enabling higher-precision path tracking. First, using the vehicle's current position and attitude information, along with partial path information, as input, the system leverages the end-to-end advantage of neural networks to directly obtain predicted preview points. Next, the predicted preview points are used to further determine the path preview points. Finally, the path preview points are fed into a traditional pure path tracking algorithm to obtain the vehicle's front wheel steering angle at the next moment, and then the vehicle is controlled to complete the path tracking task.

[0027] 2. This invention provides a path tracking preview point selection system for tractor autonomous driving based on a deep learning network. Regardless of the initial lateral deviation and initial heading angle deviation, the online path tracking algorithm based on the LPSNet preview point search network can quickly eliminate lateral position deviation and heading deviation, and approach the path as quickly as possible with a smooth curve with a small or no overshoot. This shows that the online performance of the LPSNet preview point search network algorithm is better than other algorithms.

[0028] 3. This invention provides a path tracking and aiming point selection system for tractor autonomous driving based on a deep learning network. The system was validated through simulation experiments and field trials. Simulation experiments included comparative tracking experiments on complex paths, and performance tests at three different angles under two different initial position deviations and two different speeds. Field trials included a U-shaped path comparison experiment and a performance comparison analysis. The results show that the path tracking effect and performance of this system are superior to traditional pure path tracking algorithms and MPC algorithms, making it more suitable for agricultural production and other high-precision working conditions, and meeting the requirements of tractor navigation line driving. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of a two-wheeled bicycle according to the present invention;

[0030] Figure 2 This is a schematic diagram illustrating the relationship between the tractor, the tracking path, and the turning radius of the present invention;

[0031] Figure 3 This is a schematic diagram of the deep learning neural network model of the present invention;

[0032] Figure 4 This is a schematic diagram of the pre-aiming point search network (LPSNet) of the present invention;

[0033] Figure 5 This is a schematic diagram illustrating the principle of vehicle position and attitude information data processing according to the present invention;

[0034] Figure 6 A schematic diagram of the model for determining the aiming point in this invention;

[0035] Figure 7 This is a schematic diagram of an overview model of a complex path in the field navigation line autonomous driving environment of the present invention;

[0036] Figure 8 This is a partially enlarged schematic diagram of region a in this invention;

[0037] Figure 9 This is a partially enlarged schematic diagram of region b of the present invention;

[0038] Figure 10 This is a partially enlarged schematic diagram of region d in this invention;

[0039] Figure 11 This is a partially enlarged schematic diagram of region g in this invention;

[0040] Figure 12 This is a schematic diagram of the test results of the online effect of the present invention under the condition that the initial position deviation is 2m and the driving speed is 3m / s;

[0041] Figure 13 This is a schematic diagram of the test results of the online effect of the present invention under the condition that the initial position deviation is 2m and the driving speed is 5m / s;

[0042] Figure 14 This is a schematic diagram of the test results of the online effect of the present invention under the condition that the initial position deviation is 3m and the driving speed is 3m / s;

[0043] Figure 15 This is a schematic diagram of the test results of the online effect of the present invention under the condition that the initial position deviation is 3m and the driving speed is 5m / s;

[0044] Figure 16 This is a schematic diagram of the control framework of the navigation line tracking automatic driving tractor system of the present invention;

[0045] Figure 17 This is a schematic diagram showing the test results of the U-shaped path and the comparison of errors throughout the entire process of the present invention;

[0046] Figure 18 This is a schematic diagram showing a partial enlarged view of the curve in the U-shaped path test results of this invention;

[0047] Figure 19 This is a schematic diagram of the tractor aiming distance variation curve based on the LPSNet algorithm for aiming point search network in the U-shaped path test of this invention;

[0048] Figure 20 This is a magnified schematic diagram of the online process at the beginning of the U-shaped path and the local enlarged lateral error of the present invention;

[0049] Figure 21 This is a schematic diagram of the pure path tracing algorithm model of the present invention. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to embodiments:

[0051] Example 1

[0052] This invention provides a path tracking preview point selection system for tractor autonomous driving based on deep learning networks. The method for selecting preview points for tractor autonomous driving path tracking based on deep learning includes the following steps:

[0053] Step 1: Construction of the kinematic model of the tractor test platform. Since the tractor operates relatively slowly in the field, the mechanical factors of the environment on the vehicle chassis will not be discussed; only the geometric factors during the vehicle's steering process will be studied. The tractor will be simplified into a two-wheeled bicycle model in the XOY inertial coordinate system (e.g., ...). Figure 1As shown in the figure, it is assumed that the tractor travels at a low speed in a field with good road conditions, so most of the dynamic influencing factors can be ignored. The dynamic model of the tractor is based on the following assumptions: (1) The tractor is a rigid body; (2) The rotation angles of the left and right wheels of the front axle are equal; (3) The left and right swaying and pitching motions of the tractor are ignored; (4) Lateral slip is ignored.

[0054] The relationship between the tractor, the tracking path, and the turning radius (e.g.) Figure 2 As shown in the figure, based on the relationship between the tractor, the tracking path, and the turning radius in the model, the kinematic model of the tractor is obtained:

[0055]

[0056]

[0057] Where x, y, δ, φ, L, v, and γ are the abscissa of the tractor, the ordinate of the vehicle, the front wheel steering angle, the vehicle's heading angle, the vehicle's wheelbase, the vehicle speed, and the actual turning radius, respectively. t+Δt It is the target steering curvature of the tractor;

[0058] Step 2: Perform path tracking using the Pre-aiming Point Search Network (LPSNet) algorithm. The specific steps are as follows:

[0059] S1. Building the neural network: Based on deep learning neural networks (DLNN), construct the pre-aiming point search network (LPSNet).

[0060] Deep learning neural networks (DLNNs) are multi-layer feedforward neural networks trained through error backpropagation. A simplified diagram of its simplest single-hidden-layer network model is shown below (e.g., Figure 3 (As shown). The essence of the backpropagation (BP) learning algorithm is to load the relationship between input and output into the multilayer perceptron (MLP) topology so as to fully train the network while taking into account past input data, so as to achieve the goal of enabling the network to have predictive capabilities using a given past training dataset.

[0061] Preview Point Search Network (LPSNet) (e.g.) Figure 4 As shown, the network mainly consists of fully connected (FC) layers, each activated by the ReLU function. The ReLU activation function used in the proposed Leading Point Search Network (LPSNet) has lower computational cost compared to other activation functions. Furthermore, for the neural network trained via gradient descent based on backpropagation proposed in this application, ReLU helps avoid the vanishing gradient phenomenon, which significantly hinders the learning process.

[0062] Figure 4In the diagram, Vehicle XC represents the global X-coordinate of the rear axle center of the vehicle at the current position, Vehicle YC represents the global Y-coordinate of the rear axle center of the vehicle at the current position, Vehicle HDG represents the heading angle of the vehicle at the current position, Vehicle Velocity represents the vehicle's speed at the current position, and Ref. 300path Point COOR represents the coordinates of approximately 300 path points in the forward direction from the point closest to the vehicle at the current position. This information can be obtained using the BeiDou Navigation Satellite System. In the diagram, Goal point XC represents the predicted X-coordinate of the aiming point, and Goal point YC represents the predicted Y-coordinate of the aiming point. When the tractor is working, the acquired information is input into the aiming point search network LPSNet at each moment to obtain the aiming point information needed to adjust the vehicle's attitude control.

[0063] S2. Data collection: A simulation path consisting of straight line segments and circular arc segments is established to simulate most of the situations in which the tractor travels in the field. The path is composed of discrete points. This simulation path is established in the Python 3.8 environment using PyCharm software when collecting the dataset.

[0064] The principle of vehicle position and attitude information data processing in the LPSNet pre-aiming point search network (e.g.) Figure 5 As shown in the diagram, W refers to the length of the tractor's working area, H refers to the width of the tractor's working area, v refers to the tractor's speed, and φ refers to the tractor's heading angle. The coordinate information has been normalized here. The x-coordinate is the percentage of the distance between the tractor and the left side of the working area relative to the length of its left neighbor, and the y-coordinate is the percentage of the distance between the tractor and the top side of the working area relative to its top neighbor. Compared to directly using coordinates, this method enhances the network's generalization ability.

[0065] Each data entry in the dataset refers to the coordinates of 300 points taken in the forward direction from the nearest path point to the vehicle. The coordinates of each point are composed of the following format: The discrete points, combined with the vehicle's positioning coordinates (x, y) within the working area. vehicle ,y vehicle Furthermore, considering the potential for misjudgment when relying solely on a single vehicle coordinate system, which might result in the vehicle being in the same position across different paths, the heading angle parameter φ of the vehicle at its current position is added to each valid example data, along with a velocity variable v. Therefore, each data instance is structured as a one-dimensional tensor with 1 row and 604 columns. The aiming point required for the tractor's steering control during operation is the pre-y corresponding to each input.

[0066] S3. The training process of network parameters uses the MSE loss function to reflect the distance error between the predicted target point and the expected target point, as shown below:

[0067]

[0068] x in the formula predict ,y predict These are the predicted x and y coordinates of the aiming point obtained through the aiming point search network LPSNet, respectively. expect ,x expect These are the expected values. During training, the batch size used was 64, the momentum was 0.9, and the decay rate was 0.0005.

[0069] S4. Extracting the Pre-aiming Point: To further improve the path tracking accuracy of the tractor, this algorithm does not directly use the predicted pre-aiming point, but instead performs a pre-aiming point extraction operation: On curved sections, a straight line is drawn connecting the predicted pre-aiming point and the center of curvature of the current road segment. The intersection of this straight line and the path is the pre-aiming point (e.g., ...). Figure 6 (1) As shown in the figure; on the straight line segment, draw a perpendicular line from the predicted aiming point to the path, and the intersection of the perpendicular line and the path is the aiming point (as shown in the figure). Figure 6 (2) As shown, after determining the aiming point, the pure path tracking algorithm can be combined to obtain the front wheel steering angle control variable for the vehicle to perform path tracking and then control the vehicle's steering.

[0070] The pure path tracking algorithm for tractors achieves path tracking by calculating the steering angle of the front wheels. This algorithm uses the center point of the rear axle as the tangent point and the vehicle body as the tangent line. By controlling the steering angle of the front wheels, it continuously attempts to smoothly guide the center point of the rear axle along a curve past a pre-aimed point. Figure 21 As shown:

[0071] Where (x) i ,y i (x, y) are the coordinates of the target point, (x, y) are the coordinates of the center of the rear wheel axle, R is the turning radius, and A is the intersection of the straight line parallel to the heading through B and the radius OC.

[0072] The following mathematical expression can be derived:

[0073]

[0074] In the formula (x i ,y i ) represents the coordinates of the aiming point.

[0075]

[0076] We can obtain the following from geometric relationships:

[0077]

[0078]

[0079] Combining equations (4) and (5), we can obtain the following expression:

[0080]

[0081] Substituting formulas (2) and (6) into formula (7) yields the following expression:

[0082]

[0083] Meanwhile, the curvature of the turning radius of a vehicle when turning can be expressed as follows:

[0084]

[0085] From formulas (8) and (9), the aiming distance L can be obtained. d The aiming distance is the sole variable determining the tracking performance, but determining it is relatively difficult. If the aiming distance is set too small, the vehicle will oscillate back and forth near the tracking path; if the aiming distance is sufficiently small, the tractor may experience instability that threatens its driving safety. Therefore, the aiming distance of the model used in the subsequent experiments in this paper is set to satisfy L... d >1.2m. If the pre-aiming distance is set too large, it will cause a chamfering phenomenon. Since the research in this application only needs to determine the pre-aiming distance under the premise of ensuring vehicle safety, there is no requirement for the upper limit of the pre-aiming distance.

[0086] Example 2

[0087] Based on the deep learning network-based path tracking preview point selection system for tractor autonomous driving proposed in this application, a simulation experiment was conducted, and the simulation results were analyzed.

[0088] I. Simulation Environment: To verify the path tracking algorithm proposed in this paper, several simulation tests were designed using PyCharm software in Python 3.8 and TensorFlow environment. These tests compared traditional pure path tracking algorithms based on different aiming distances with the Model Predict Control (MPC) algorithm. The computer configuration used was: Windows 10 Home Basic 64-bit, eight cores / sixteen threads, 16GB RAM, NVIDIA GeForce RTX 2060 graphics card, 6GB VRAM.

[0089] II. Path tracing under complex paths.

[0090] a) Simulation experiment: The automatic driving route of the tractor's field navigation line consists of straight and curved segments. Therefore, the demonstration of the path tracking effect is mainly achieved by establishing a complex path simulating the automatic driving environment of the field navigation line. This path includes straight segments with different inclination angles, curved segments with different curvatures, and the tractor's entry process when starting work. Figure 7 In the overview diagram, region a represents the upper line section when the tractor starts working, and regions b, c, d, e, f, and g are curve segments with curvatures of 0.14, 0.17, 0.14, 0.14, 0.17, and 0.14, respectively. During tracking, the tractor's speed on straight sections is set to simulate normal field operation at 5 m / s, and on curved sections it is set to 3 m / s. The tracking effects of several algorithms are compared... Figure 7 As shown, the Overview picture is an overview of the tracking effect of the entire path and an error comparison picture of several algorithms, while the rest are magnified local pictures of the marked areas on the path and error comparison pictures of the current road segment.

[0091] b. Simulation result analysis.

[0092] Table 1 shows the lateral error analysis of the tractor after entering the path under complex conditions. It can be seen that the overall lateral error, mean, and maximum value of the LPSNet algorithm based on the pre-aiming point search network are all smaller than those of the pure path tracking algorithm and the Model Predict Control (MPC) algorithm based on different pre-aiming distances. The data shows that the tracking performance of the LPSNet algorithm is better than the other algorithms.

[0093] Table 1: Comparison of Overall Tracking Error Data under Complex Paths

[0094]

[0095] exist Figure 7 The complex path shown includes curves with curvatures of 0.17 and 0.14. For these two curves with different curvatures, a comparative analysis of the lateral error of path tracking is shown in Table 2. It can be seen that the average and maximum error values ​​of the path tracking algorithm under different curvatures are smaller than those of other algorithms, indicating that the pre-aiming point search network LPSNet has better tracking performance in curves than other algorithms.

[0096] Table 2: Comparison of tracking error data under different curvature curves

[0097]

[0098] III. Performance Comparison After Launch.

[0099] a. Simulation Experiment: The alignment performance of the tractor navigation system is one of the important factors affecting the quality and efficiency of tractor operations in farmland. The three main factors for evaluating alignment performance are alignment speed, alignment stability, and overshoot. Two initial position deviations were set in the alignment effect test: 2m and 3m. For each initial position deviation, three initial heading deviation angles were set: 0°, 90°, and 180°. Simultaneously, two travel speeds were set for each alignment effect condition: 3m / s and 5m / s. The alignment effect test results for the initial position deviation of 2m and the travel speed of 3m / s are as follows: Figure 12 As shown; the test results for the online performance under the conditions of an initial position deviation of 2m and a driving speed of 5m / s are as follows. Figure 13 As shown; the test results for the online performance under the conditions of an initial position deviation of 3m and a driving speed of 3m / s are as follows. Figure 14 As shown; the test results for the online performance under the conditions of an initial position deviation of 3m and a driving speed of 5m / s are as follows. Figure 15 As shown.

[0100] b. Comparison of online performance.

[0101] In the process of tractor path tracking, the performance evaluation is mainly based on the following aspects. First, the path tracking distance refers to the integral length of the path from point P to the starting point of the path, provided that the tractor's lateral deviation from point P is less than a set threshold T. In the path tracking performance test conducted in this application, the path tracking distance is visually represented as the difference in longitudinal coordinates between point P and the starting point of the path. Second, the path tracking time refers to the time taken for the tractor to reach point P. In addition to path tracking distance and time, overshoot is also introduced to evaluate the tractor's path tracking performance at the initial position. Overshoot refers to the maximum lateral distance the tractor deviates from the path after its first intersection with the path during the path tracking process. The path tracking performance data statistics for an initial position deviation of 2m are shown in Table 3, and the data statistics for an initial position deviation of 3m are shown in Table 4.

[0102] Table 3. Comparison of online performance when the initial position deviation is 2m

[0103]

[0104] Table 4. Comparison of online performance when the initial position deviation is 3m

[0105]

[0106]

[0107] from Figure 12 , Figure 13 , Figure 14 , Figure 15 As shown in Tables 3 and 4, with an initial position deviation of 2m, the pure path tracking algorithm based on the LPSNet pre-aiming network approaches the path with a shorter pre-aiming distance and less pre-aiming time, while maintaining good tracking performance. Regarding overshoot, the LPSNet algorithm outperforms the other three traditional pure path tracking algorithms based on different pre-aiming distances. Although the MPC algorithm did not exhibit significant overshoot, it consistently maintained a certain lateral error in subsequent tracking tests. Therefore, overall, with an initial lateral error of 2m and different initial headings and speeds, the path tracking algorithm based on the LPSNet pre-aiming network demonstrates better pre-aiming performance.

[0108] Based on the above analysis, the pure path tracking algorithm based on the LPSNet pre-aiming point search network increases the online time and online distance with increasing initial position deviation. However, overall, it can quickly eliminate lateral deviation under different initial position deviation conditions, and the online curve changes smoothly, indicating that the algorithm can effectively improve the tractor's online speed and stability. As shown in the figure, under the two initial lateral deviations, when the initial heading deviation is 0° (a relatively large heading deviation), the MPC algorithm will produce a large overshoot and oscillation at a vehicle speed of 5 m / s. Traditional pure path tracking algorithms based on several different pre-aiming distances, because the lateral deviation of the vehicle's position at the start is already greater than the preset pre-aiming distance, and this trend increases over a period of time, require setting a larger pre-aiming distance at the start stage. Path tracking can only be performed at a fixed pre-aiming distance when the vehicle's lateral error is less than or equal to the preset pre-aiming distance. Therefore, traditional pure path tracking algorithms based on several different pre-aiming distances will produce a larger lateral error and a larger online distance, which further increases the online time. Overall, regardless of the initial lateral deviation and initial heading angle deviation, the online path tracking algorithm based on the LPSNet pre-aiming network can quickly eliminate lateral position deviation and heading deviation, and approach the path as quickly as possible with a smooth curve with small or no overshoot. This shows that the online performance of the LPSNet pre-aiming network algorithm is better than the other algorithms.

[0109] Example 3

[0110] Based on the deep learning network-based path tracking preview point selection system for tractor autonomous driving proposed in this application, field tests were conducted, and the test results were analyzed.

[0111] I. Construction of field test platform.

[0112] In field tests, the tractor test platform was used to track a reference path in the wild to verify the path tracking control algorithm. The system control framework of the self-navigating tractor operates as follows: Figure 16 As shown in Table 5, the hardware of the perception layer is the Beidou system. The Beidou receiving antenna is fixed to the front of the tractor, providing high-precision positioning information to the tractor through RTK differential principles. The hardware of the decision layer is an IPC (Industrial Personal Computer), which makes real-time decisions on actuator commands based on the tractor's current state. The hardware of the execution layer consists of an automatic steering device and a speed control device. The automatic steering device uses a DC motor to drive the SCU (Hydraulic Steering Control Unit) to control the tractor's automatic steering. The speed control device controls the clutch, brake, and throttle through an actuator DC motor with position feedback to achieve speed control. The decision layer and perception layer transmit positioning data via RS232 serial communication at a frequency of 20Hz. Simultaneously, the decision layer and execution layer transmit data via a CAN bus at a communication rate of 500Kbps. The basic parameters of the tractor test platform are shown in Table 5.

[0113] Table 5

[0114]

[0115]

[0116] The tractor used for the field test platform was a Huanghai Jinma 554. We established path tracking experiments on both dry and wet roads. The test location was a certain place. The latitude and longitude of the test site were: longitude: 117.58548080, latitude: 32.13491429. The road conditions for both dry and wet roads were clay soil. The road conditions were bumpy. The weather on the wet road was cloudy, with a temperature of 14 degrees Celsius and a humidity of 25%.

[0117] II. The tractor test platform uses the Beidou device to feed back the real-time position of the tractor and performs automatic navigation on the reference path based on the path tracking algorithm. Considering different initial states, turning radii and driving speeds, experiments were conducted according to the following path planning strategies: (1) An inverted U-shaped path was set and the curvatures of the two curves on the inverted U-shaped path were set to 0.11 and 0.14, respectively. (2) The starting point was set to face the path direction and offset by 2m. (3) During the driving process, the straight section was set to 5m / s and the curve section to 3m / s. Figure 17 The image shows a comparison of the tracking results of the experiment and a comparison of the error curves for the entire journey. Figure 18 (a) and (b) show magnified views of the path tracking results at the first and second curves and the starting point, respectively, corresponding to... Figure 17Regions b and c in the text.

[0118] III. Analysis of Field Test Results.

[0119] In this experiment, the tractor's position data will be recorded during each path tracking operation. The first 373 tractor position data points from each experiment will be used for result analysis, as these 373 position records are sufficient to complete all the experimental procedures set for this experiment. Figure 18 The results show that the path tracking performance of the LPSNet algorithm based on the pre-aiming point search network is better than that of the pure path tracking algorithm at different pre-aiming distances. Table 6 shows the lateral error analysis after the tractor enters the path. The table reveals that the overall mean, maximum, and standard deviation of the lateral error of the LPSNet algorithm are all superior to those of the pure path tracking algorithm at different pre-aiming distances. The average error of the LPSNet algorithm is 0.019, a reduction of 39% compared to the best performance of the pure path tracking algorithm; the maximum error is 0.079, a reduction of 28% compared to the best performance of the pure path tracking algorithm; and the standard deviation is 0.015, a reduction of 50% compared to the best performance of the pure path tracking algorithm. The data demonstrate that the tracking performance of the LPSNet algorithm based on the pre-aiming point search network has certain advantages over other algorithms.

[0120] Table 6. Global Error Analysis Statistics of the Inverted U-Shaped Path Experiment

[0121]

[0122] This application analyzes why the LPSNet algorithm for pre-aiming point search has this advantage by examining the change in pre-aiming distance. The change in pre-aiming distance in the LPSNet algorithm is as follows: Figure 19 As shown in the figure, the aiming distance of the path tracking based on the LPSNet algorithm of aiming point search network changes continuously depending on the location of the tractor. This also indicates that the tractor adopts different tracking measures in the tracking process on straight and curved sections.

[0123] In this experiment, the tractor was positioned two meters off-center from the path direction at the starting point. Therefore, this experiment included one path-tracking onboarding process. A partial magnification of the path-tracking onboarding process at the starting point is shown below. Figure 20 As shown in Table 7, the statistics for online time, online distance, and overshoot are as follows.

[0124] Table 7. Statistics of Results During the Launch Process

[0125]

[0126]

[0127] Depend on Figure 20 As shown in Table 7, path tracking based on the LPSNet algorithm with pre-aiming point search network can enter the path faster, with an entry distance of 2.51m, which is 58% faster than the pure path tracking algorithm; and an entry time of 2.59s, which is 15% faster than the pure path tracking algorithm. The results indicate that path tracking based on the LPSNet algorithm with pre-aiming point search network has better entry performance in field tests.

Claims

1. A path tracking and preview point selection system for tractor autonomous driving based on a deep learning network, characterized in that: This deep learning-based method for selecting preview points in tractor autonomous driving path tracking includes the following steps: Step 1: Construction of the kinematic model of the tractor test platform; Step 2: Perform path tracking using the LPSNet algorithm, which involves the following steps: S1. Building the neural network; S2, Data Collection; S3, the training process of network parameters; S4. Extract the aiming point; S1 further includes constructing a pre-aiming point search network LPSNet based on a deep learning neural network DLNN. The pre-aiming point search network LPSNet is mainly composed of FC layers, and each layer is activated by the ReLU function. When the tractor is working, the information obtained at each moment is input into the pre-aiming point search network LPSNet to obtain the pre-aiming point information required to complete the adjustment of vehicle attitude control. S4 further includes, in the curved segment, drawing a straight line by connecting the predicted aiming point with the curvature center of the current road segment, the intersection of the straight line and the path is the aiming point; in the straight segment, drawing a perpendicular line from the predicted aiming point to the path, the intersection of the perpendicular line and the path is the aiming point. After determining the aiming point, and combining it with the pure path tracking algorithm, the front wheel steering angle control variable for path tracking of the vehicle can be obtained, and then the vehicle can be steering controlled.

2. The tractor automatic driving path tracking preview point selection system based on deep learning network according to claim 1, characterized in that: The specific method of step one is as follows: simplify the tractor into a two-wheeled bicycle model in the XOY inertial coordinate system, and obtain the kinematic model of the tractor based on the relationship between the tractor, the tracking path, and the turning radius in the model: in These are the tractor's x-coordinate, the vehicle's y-coordinate, the front wheel steering angle, the vehicle's heading angle, the vehicle's wheelbase, the vehicle speed, and the actual turning radius. It is the target steering curvature of the tractor.

3. The tractor automatic driving path tracking preview point selection system based on deep learning network according to claim 1, characterized in that: The S2 also includes establishing a simulation path composed of straight line segments and circular arc segments to simulate most of the situations during the tractor's travel in the field, and the path is composed of discrete points.

4. The tractor automatic driving path tracking preview point selection system based on deep learning network according to claim 1, characterized in that: S3 further includes using the MSE loss function to reflect the distance error between the predicted aiming point and the expected aiming point, specifically expressed as follows: In the formula These are the pre-aiming points obtained through the pre-aiming point search network LPSNet. Coordinate prediction values These are the expected values.

5. A tractor automatic driving path tracking preview point selection system based on deep learning networks according to claim 1, characterized in that: The pure path tracing algorithm is achieved by calculating the steering angle of the front wheels. The pure path tracing algorithm takes the center point of the rear axle of the vehicle as the tangent point and the vehicle body as the tangent line. By controlling the steering angle of the front wheels, it continuously tries to make the center point of the rear axle of the tractor smoothly pass the pre-aiming point along a curve.

Images