Intelligent driving system and method applied to orchard

Abstract

The invention relates to the technical field of agricultural intelligent equipment, and discloses an intelligent driving system and method applied to an orchard, and the system comprises a four-wheel independent hub motor drive chassis, a terrain sensing module, a dynamic torque distribution module, a multi-sensor fusion positioning module, a path rapid detection module, a transverse tracking control module, and an emergency obstacle avoidance module. According to the intelligent driving system and method applied to the orchard, a four-wheel independent hub motor driving chassis, a terrain sensing module, a dynamic torque distribution module, a multi-sensor fusion positioning module, a rapid path detection module, a transverse tracking control module and an emergency obstacle avoidance module are fused; all-terrain self-adaptive driving and obstacle avoidance of the unmanned transport vehicle, high-precision and anti-interference positioning in a complex environment and rapid detection and navigation of an inter-ridge road advancing path can be realized, so that mechanical intelligent transportation is realized, the labor intensity can be effectively reduced, and the operation efficiency and safety of an orchard can be improved.

Description

An intelligent driving system and method for use in orchards Technical Field

[0001] This invention relates to the field of agricultural intelligent equipment technology, specifically to an intelligent driving system and method applied to orchards. Background Technology

[0002] Intelligent agricultural machinery is a crucial support for the development of modern agriculture. Guangdong Province has approximately 16.1351 million mu (1.16 million hectares) of fruit planting area, ranking third in the country. The fruit industry has developed into an important sector for increasing farmers' income, improving agricultural efficiency, and promoting rural development in the province. However, 70% of the province's total land area is hilly and mountainous, with varied terrain, steep slopes, and numerous bends. Crop transportation in orchards mainly relies on backpacking, stretcher-mounted, and tractor-drawn methods, resulting in high risks, high labor intensity, and low work efficiency for fruit farmers. There is an urgent need for efficient transportation equipment suitable for mountainous orchard environments.

[0003] In recent years, with the implementation of national policies, agricultural tractors, tricycles, and other transport vehicles have gradually entered the orchard operation field, improving orchard transportation to some extent. However, the hilly and mountainous terrain of Lingnan is quite complex, with slopes reaching up to 20°. Traditional transport vehicles lack the ability to climb slopes and overcome obstacles, are too large to turn easily, cannot operate in forests, and are manually driven, making them unsuitable for the modernization of orchards. Furthermore, the large scale of orchards, the repetitive geometric features of the environment, and the non-uniform coverage of BeiDou signals make positioning difficult, posing a significant challenge to the development of intelligent driving transport vehicles. Summary of the Invention

[0004] To address the problems mentioned in the background section, this invention provides the following technical solution: an intelligent driving system for orchards, comprising: a four-wheel independent hub motor driven chassis, wherein the chassis has a ground clearance ≥150mm, a single-wheel peak torque ≥100N·m, and is equipped with an electronic differential lock; a terrain perception module for acquiring three-dimensional terrain information ahead and outputting a passability score S∈[0,1]; a dynamic torque distribution module connected to the terrain perception module and the four motors for adjusting the torque of each motor in real time according to the score S; and a multi-sensor fusion positioning module, comprising at least... RTK-GNSS, LiDAR, IMU, and binocular depth camera are used to switch to LiDAR, vision, or IMU tightly coupled mode when GNSS signal is lost; a path fast detection module is used to perform HSV adaptive threshold segmentation, probabilistic Hough line detection, and multi-frame historical weighted fusion on the road images between ridges to output a drivable centerline; a lateral tracking control module is used to calculate the front wheel steering angle based on the error between the centerline and the vehicle's current position to achieve line following; an emergency obstacle avoidance module is used to perform millisecond-level electric braking when an obstacle is detected at a distance ≤0.3m ahead.

[0005] Furthermore, the terrain perception module includes: a binocular depth camera for acquiring point clouds in front; a convolutional neural network that takes the point clouds or disparity map as input and outputs slope, gully, and mud probability maps; and a passability scoring unit for weighted summation of the probability maps to obtain a score S.

[0006] Furthermore, a binocular depth camera is used to acquire 3D information of the terrain ahead, and ultrasonic sensors are combined to detect nearby obstacles in real time, thereby enhancing the ability to avoid obstacles at close range in complex terrain. A Kalman filter fusion algorithm is used to fuse multi-source information from the binocular depth camera and ultrasonic sensor data to achieve real-time 3D reconstruction of the terrain ahead. A convolutional neural network is used to train the terrain images acquired by the binocular depth camera. Based on the output of the convolutional neural network, a passability scoring model is constructed to generate a quantified terrain passability score S, with the score S ranging from 0 to 1, where 0 indicates impassable and 1 indicates fully passable, providing data support for path planning.

[0007] Furthermore, the dynamic torque distribution module adopts a PD controller, which automatically locks the differential when S < 0.4. Specifically, the PD control algorithm is used to dynamically adjust the torque and speed of the four independent motors, and the traction distribution is optimized through real-time feedback and closed-loop control to ensure the stability and passability of the vehicle in complex terrain. The system integrates a differential lock control strategy, automatically switches the differential lock mode, and optimizes the torque distribution between wheels.

[0008] Furthermore, the multi-sensor fusion positioning module employs extended Kalman filtering to dynamically weight multi-sensor data and performs global BA optimization upon successful loop closure detection to eliminate accumulated errors. Specifically, in areas with stable GNSS signals, RTK-GPS is used to provide global centimeter-level positioning, ensuring high-precision navigation; when entering a weak signal environment, it switches to a tightly coupled lidar-inertial IMU mode for local positioning. The use of extended Kalman filtering to dynamically weight multi-sensor data suppresses single-sensor failures or noise interference, and dynamically adjusts the fusion weights based on the confidence level of each sensor to ensure the robustness of the positioning system.

[0009] Furthermore, the multi-sensor fusion localization module also includes: using a feature-point-based binocular visual SLAM algorithm for SLAM mapping; firstly, offline calibration of the intrinsic and extrinsic parameters of the binocular depth camera, including the principal points, focal lengths, distortion coefficients, relative poses, and baseline lengths of the left and right cameras; then, distortion correction and stereo correction are performed on the left and right images based on the distortion coefficients to obtain the left and right camera remapping matrix with coplanar row alignment; next, initialization of the feature-point-based binocular visual SLAM algorithm is performed to construct an initial 3D environment map. The initialization includes feature extraction and feature matching of the left and right images to find corresponding points in the left and right images, and then calculating the feature point depth based on disparity to reconstruct the 3D coordinates of all feature points; finally, a direct method is used to assist in improving localization in scenes with missing textures, and IMU data is fused to compensate for localization loss caused by camera motion; for areas where visual SLAM has been run, if the camera moves along a non-closed-loop path, the localization result of the current image in the historical map is used as the current localization result, and BA optimization is performed on the data within the current trajectory.

[0010] Furthermore, the path fast detection module includes: an HSV color space conversion unit for dynamically adjusting the threshold based on the prior interval of the yellow component of the orchard ground; a Gaussian filtering unit for suppressing high-frequency noise; a probabilistic Hough line detection unit for detecting road edges in the image; a DBSCAN clustering unit for merging similar lines; and a historical frame weighted fusion unit for caching the line parameters of the previous n frames and outputting a weighted average centerline, where n=5.

[0011] Furthermore, the lateral tracking control module adopts PD control with a proportional coefficient Kp=0.8, a derivative coefficient Kd=0.15, and a line-following error ≤5mm.

[0012] A smart driving method for orchards includes a fast path detection method, specifically: acquiring RGB images of inter-row roads; converting the images to the HSV color space; adaptive threshold segmentation to extract a binary mask of the drivable area; performing edge detection and probabilistic Hough line detection on the mask to obtain a candidate line set; using DBSCAN to cluster the candidate lines and merge lines in the same lane; introducing historical line parameters from the previous n frames and weighted averaging to output the current centerline; and sending the centerline to a lateral tracking control module to achieve line-following driving.

[0013] Furthermore, it also includes an all-terrain adaptive control method, specifically: acquiring real-time three-dimensional terrain information ahead; obtaining a passability score S through a convolutional neural network; dynamically allocating the torque of the four motors according to S, locking the differential when S < 0.4; using RTK positioning when GNSS signal is good, switching to lidar or IMU or visual tightly coupled positioning when lock is lost; paralleling the aforementioned path rapid detection method to obtain the centerline; calculating the front wheel steering angle based on the centerline error and controlling the vehicle to travel along the road; and immediately executing electric braking when an obstacle is detected at a distance ≤ 0.3m.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The intelligent driving system and method applied to orchards, through the integration of a four-wheel independent hub motor drive chassis, a terrain perception module, a dynamic torque distribution module, a multi-sensor fusion positioning module, a path rapid detection module, a lateral tracking control module, and an emergency obstacle avoidance module, can realize all-terrain adaptive driving and obstacle avoidance of unmanned transport vehicles, high-precision and anti-interference positioning in complex environments, and rapid detection and navigation of the travel path between rows, thereby realizing mechanized intelligent transportation, which can effectively reduce labor intensity and improve the operational efficiency and safety of orchards. Attached Figure Description

[0015] Figure 1 is a system framework diagram of the present invention; Figure 2 is a flowchart of the multimodal sensor fusion localization process of the present invention; Figure 3 is a flowchart of the SLAM technology path adapted to multiple operating environments of the present invention; Figure 4 is a Hough transform detection effect diagram of the present invention; Figure 5 is a detection effect diagram after image processing of the present invention; Figure 6 is a detection effect diagram of road combined with historical information of the present invention; Figure 7 is a road centerline output effect diagram of the present invention; Figure 8 is a flowchart of road centerline calculation of the present invention; Figure 9 is a flowchart of the PD control algorithm of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] An embodiment of the intelligent driving system and method applied to orchards is as follows: Please refer to Figures 1-9. An intelligent driving system applied to orchards includes: a four-wheel independent hub motor driven chassis, i.e., a four-wheel four-wheel drive chassis, wherein the chassis has a ground clearance ≥150mm, a single-wheel peak torque ≥100N·m, and is equipped with an electronic differential lock; a terrain perception module, used to acquire three-dimensional information of the terrain ahead and output a passability score S∈[0,1], wherein the terrain perception module includes: a binocular depth camera, used to acquire a point cloud ahead; a convolutional neural network, wherein the convolutional neural network takes the point cloud or disparity map as input and outputs a probability map of slope, gully, and mud; and a passability scoring unit, used to obtain a score S by weighted summation of the probability map, specifically: using a binocular depth camera to acquire three-dimensional information of the terrain ahead, and combining it with an ultrasonic sensor in real time. This system detects nearby obstacles to enhance obstacle avoidance capabilities in complex terrain. A Kalman filter fusion algorithm is used to fuse data from a binocular depth camera and an ultrasonic sensor, enabling real-time 3D reconstruction of the terrain ahead with an accuracy of ±5cm. Simultaneously, the ultrasonic sensor's obstacle detection accuracy within a 0.5m range is improved to 99%, ensuring high reliability for near-field obstacle avoidance. A convolutional neural network (CNN) is used to train deep learning on terrain images acquired by the binocular depth camera, achieving accurate identification of various complex terrain features such as slope, gullies, and mud. Based on the output of the CNN, a passability scoring model is constructed, generating a quantified terrain passability score S, ranging from 0 to 1, where 0 indicates impassable and 1 indicates fully passable, providing data support for path planning. The CNN model achieves over 95% accuracy in identifying terrain features such as slope and gullies, and the passability scoring system effectively distinguishes the difficulty of traversing different terrains, significantly improving the reliability and decision-making efficiency of path planning.

[0018] A dynamic torque distribution module, connected to the terrain sensing module and the four motors, adjusts the torque of each motor in real time based on a score S. This module employs a PD controller, automatically locking the differential when S < 0.4. Specifically, it uses a PD control algorithm to dynamically adjust the torque and speed of the four independent motors, optimizing traction distribution through real-time feedback and closed-loop control to ensure vehicle stability and passability in complex terrains. For challenging terrains such as steep slopes and mud, the system integrates a differential lock control strategy, automatically switching differential lock modes to optimize wheel torque distribution, effectively suppressing slippage and improving traction performance under extreme conditions. The system improves climbing efficiency by 35% on a 30° steep slope, with more reasonable traction distribution. Simultaneously, the differential lock mode switching response time in muddy terrain is reduced to ≤0.2s, significantly enhancing the vehicle's terrain adaptability and passability, providing reliable technical support for all-terrain operations.

[0019] The multi-sensor fusion positioning module includes at least RTK-GNSS, lidar, IMU, and a binocular depth camera. It is used to switch to lidar, vision, or IMU tightly coupled mode when GNSS signal is lost. Beidou navigation + IMU combined navigation is suitable for outdoor scenarios with satellite signals, while lidar + IMU fusion positioning is suitable for indoor scenarios, greenhouses, canyons, and other scenarios where satellite signals are blocked or even completely absent.

[0020] The multi-sensor fusion positioning module uses extended Kalman filtering to dynamically weight multi-sensor data and performs global BA optimization upon successful loop closure detection to eliminate accumulated errors. Specifically, in areas with stable GNSS signals (HDOP≤2.5), RTK-GPS is preferentially used to provide global centimeter-level positioning to ensure high-precision navigation. When entering weak signal environments such as greenhouses or dense fruit tree areas (HDOP>2.5), it switches to a tightly coupled lidar-inertial IMU mode for local positioning. It adaptively achieves seamless positioning switching between indoor and outdoor environments, or dense forests and open areas, to achieve the goal of automatic positioning and navigation across the entire domain.

[0021] The system employs Extended Kalman Filter (EKF) to dynamically weight multi-sensor data, suppressing single-sensor failures or noise interference. The fusion weights are dynamically adjusted based on the confidence level of each sensor, ensuring the robustness of the positioning system. The EKF effectively suppresses single-sensor noise, improving the stability of the positioning system. Through multi-sensor data fusion, the positioning error is significantly reduced, with the lateral error controlled within ≤3cm, significantly improving the accuracy and anti-interference capability of the positioning system in complex environments.

[0022] As shown in Figure 2, the specific process for achieving high-precision positioning by fusing IMU (Inertial Measurement Unit) and LiDAR (Light Detection and Ranging) data is as follows: IMU data processing: pre-integration processing is performed on the raw IMU measurement data to generate IMU pre-integration factors; using the pre-integration results, the system pose change and initial global position value are estimated; the BeiDou IMU coordinate system initialization and BeiDou (GNSS) global position constraint factors are introduced to construct IMU factors and BeiDou factors to suppress cumulative errors.

[0023] LiDAR data processing: Extract geometric features (such as edges and planes) from the raw point cloud of the LiDAR; perform point cloud registration between the current frame features and the constructed global map using the Scan-to-Map strategy; calculate the LiDAR odometry factor to provide relative pose constraints.

[0024] Multimodal fusion optimization: Under the factor graph optimization framework, the IMU pre-integration factor, BeiDou global factor and laser odometry factor are jointly optimized; the optimal pose estimation of the system state variables is solved by nonlinear optimization methods (such as GTSAM, Ceres Solver); and the fused high-precision global pose is output to achieve robust, low-drift long-term positioning.

[0025] The system uses an IMU to provide high-frequency motion estimation, a lidar to provide low-drift geometric constraints, and a BeiDou system to provide global absolute reference. The three work together to achieve multimodal sensor fusion positioning within a factor graph optimization framework, balancing real-time performance and accuracy.

[0026] The multi-sensor fusion localization module also includes: using a feature-point-based binocular visual SLAM algorithm for SLAM mapping. First, the intrinsic and extrinsic parameters of the binocular depth camera are calibrated offline, including the principal points, focal lengths, distortion coefficients, relative poses, and baseline lengths of the left and right cameras. Distortion correction and stereo correction are performed on the left and right images based on the distortion coefficients to obtain the left and right camera remapping matrix with coplanar row alignment. Then, the feature-point-based binocular visual SLAM algorithm is initialized to construct an initial 3D environment map. The initialization process includes feature extraction and feature matching of the left and right images to find corresponding points in the left and right images. Then, the depth of the feature points is calculated based on the disparity, and the 3D coordinates of all feature points are reconstructed. Finally, a direct method is used to assist in improving localization in scenes with missing textures, and IMU data is fused to compensate for localization loss caused by camera motion.

[0027] For areas where visual SLAM has already been run, if the camera moves along a non-closed-loop path, the localization result of the current image on the historical map is used as the current localization result, and data within the current trajectory is optimized by BA to eliminate accumulated errors and improve localization accuracy.

[0028] As shown in Figure 3, the SLAM mapping adopts a feature point-based binocular visual SLAM algorithm. The specific technical path is as follows: sensor data input receives IMU data and new image data (visual information) as the basis for system perception.

[0029] IMU data is used for high-frequency motion estimation, and image data is used for environmental feature extraction and geometric constraint construction.

[0030] IMU pre-integration and pose prediction: IMU data is pre-integrated to generate short-term motion constraints; combined with the localization results of the previous moment, the initial pose of the current frame is predicted, providing prior information for subsequent visual tracking.

[0031] Visual-IMU fusion localization tightly couples and fuses visual features with IMU pre-integration results (such as VINS and MSCKF methods) to output the localization results of the current frame, including pose, velocity, and IMU bias estimation.

[0032] Tracking status detection and failure recovery determine whether visual tracking is successful: if successful → continue the normal process; if it fails (e.g., feature loss, motion blur, drastic changes in lighting) → activate the failure recovery mechanism: rely on IMU short-term dead reckoning to maintain positioning; or trigger the relocalization module (e.g., global localization based on bag-of-words model) to reinitialize tracking.

[0033] Local feature extraction and map localization: Extract local feature points (such as ORB, SIFT) of the current frame and match them with existing map data to achieve map localization (i.e., confirm the pose of the current frame in the map).

[0034] If the SLAM system status update is normal and the keyframe conditions are met: insert the current frame as a keyframe into the map; update the map (add map points, adjust the common view structure); trigger local BA (Bundle Adjustment) to optimize neighboring keyframes and map points, and improve local accuracy.

[0035] Loop closure detection and global optimization are based on the bag-of-words model to detect loop closures and identify whether the robot has returned to its previous position. If a loop closure is detected, loop closure constraints are added and global graph optimization (such as Pose Graph Optimization) and global BA are initiated to eliminate accumulated errors, correct the overall pose and path, and achieve long-term consistency.

[0036] The map expansion and multi-environment adaptation system continuously expands the map to support large-scale environment operation; through modular design (such as sensor switching, dynamic parameter adjustment, and multi-map management), it adapts to different operating environments, such as indoor / outdoor, lighting changes, dynamic scenes, and sensor degradation.

[0037] This technical approach constitutes a robust vision-IMU SLAM system that adapts to multiple operating environments and is suitable for complex scenarios such as robotics, AR / VR, and autonomous driving.

[0038] The path rapid detection module performs HSV adaptive threshold segmentation, probabilistic Hough line detection, and multi-frame historical weighted fusion on the inter-row road image to output a drivable centerline. It should be noted that when the satellite signal is good (HDOP≤2.5), the unmanned transport vehicle uses teach-in navigation. When the satellite signal is obstructed (HDOP>2.5), it automatically switches to SLAM positioning mode, uses an RGB camera to collect road information, applies Gaussian blur to the image, performs preprocessing operations such as setting a demonstration mask based on the road surface, and then uses the Hough algorithm to quickly identify the road driving line. Combined with fuzzy PID control, this enables rapid identification and following of the transport equipment's road. Specific technical details are as follows: an HSV color space conversion unit dynamically adjusts the threshold based on the prior interval of the yellow component of the orchard ground; a Gaussian filtering unit suppresses high-frequency noise; a probabilistic Hough line detection unit detects road edges in the image; a DBSCAN clustering unit merges similar lines; and a historical frame weighted fusion unit caches the line parameters of the previous n frames and outputs a weighted average centerline (n=5). Specifically, during the rapid road detection and recognition process, an RGBD camera (depth camera) (parameters: structured light; depth range: 0.8-3.5m; RGB resolution: 640*480; frame rate: 30fps; latency: 1.5 frames; field of view: 58°H, 45°V; physical size: 180x40x25 mm (head); interface: USB 2.0) is used to collect detailed road data. After that, the video data is segmented, preprocessed, cleaned, and classified.

[0039] The image was converted from the BGR color space to the HSV color space (as shown in Figure 4). HSV thresholds were set based on the colors of the plants and the ground (green and yellow, respectively), and thresholding segmentation was used. A binary mask image was created, where the yellow portion was preserved and other portions were filtered out. Edge detection was then performed on the mask image to obtain the edge image. Finally, a Hough transform was applied to the edge image to detect road edges in the image. Gaussian blur was further added to filter out unimportant high-frequency noise while preserving low-frequency details, thereby improving the accuracy of line recognition. The transformation effect is shown in Figure 5.

[0040] The DBSCAN algorithm is used to merge clusters of straight lines with similar positions and angles into a single straight line. As shown in Figure 6, the n lanes detected in the previous few frames are stored as historical information (here, n is set to 5). If a lane is detected in the current frame, the historical information is refreshed, averaged, and then output; otherwise, the historical information is directly averaged and output. Finally, the centerline needs to be output, as shown in Figure 7, as the travel path of the autonomous navigation transport vehicle.

[0041] As shown in Figure 8, the specific calculation process for the road centerline is as follows: 1. Input two line parameters to obtain the detected two road boundary lines: Line 1: starting point x1, ending point x1_; Line 2: starting point x2, ending point x2_; 2. Calculate the x-coordinates of the third points. For each line, calculate the third point closest to the vehicle end (i.e., 1 / 3 of the distance from the starting point): Line 1 third point x: If X1>X1_, X=X1+(X1-X1_) / 3; If X1≤X1_, X=X1-(X1_-X 1) / 3; The x-coordinate of the third point of the straight line X_: If X2>X2_, X_=X2+(X2-X2_) / 3; If X2≤X2_, X_=X2-(X2_-X2) / 3; 3. Calculate the x-coordinate of the centerline endpoint. Take the average of the x-coordinates of the third points of the two straight lines as the x-coordinate of the centerline endpoint X_center: X_center=(X_+X) / 2; 4. Output the centerline. Starting from the current position of the vehicle, use X_center as the x-coordinate of the centerline to generate the road centerline for lateral tracking control.

[0042] The lateral tracking control module calculates the front wheel steering angle based on the error between the centerline and the vehicle's current position, enabling lane-following navigation. This module employs PD control with a proportional-derivative (PD) coefficient of 0.8, a derivative coefficient of 0.15, and a lane-following error ≤ 5mm. Specifically, the module detects lane lines on both sides of the road using the steps described above, calculates the road's centerline for the transport vehicle to follow. The centerline's position coordinates are compared with the vehicle's current position to obtain an error signal. This error signal is then input into a PD control algorithm to calculate an angular velocity control signal, which adjusts the transport vehicle's direction of motion, enabling it to move more accurately along the road's centerline.

[0043] Among them, as shown in Figure 9, the specific process of the PD control algorithm is as follows: 1. Input parameter acquisition: Obtain the current horizontal coordinate position X_car of the trolley; obtain the horizontal coordinate X_center of the end point of the road center line (calculated by the path rapid detection module); 2. Error calculation: Calculate the position error: error = X_center - X_car; 3. PD control law: Proportional term: P = Kp * error (Kp = 0.8); Differential term: D = Kd * (error - prev_error) / dt (here Kd = 0.15, prev_error is the error of the previous cycle, and dt is the control cycle); Output the front wheel steering angle: steering_angle = P + D; 4. Steering logic: If error > 0 (i.e., X_car < X_center, the trolley is to the left), then steering_angle is a positive value, controlling the trolley to turn right; if error < 0 (i.e., X_car > X_center, the trolley is to the right), then steering_angle is a negative value, controlling the trolley to turn left; if error = 0, keep the current direction.

[0044] 5. Constraints and optimization: Line following error threshold: ≤ 5 mm (i.e., |error| ≤ 5 mm at steady state); Update the historical error: prev_error = error (used for the differential term calculation in the next cycle); Note: In implementation, the steering_angle needs to be limited (such as ±30°), and the control cycle dt needs to be ensured to be stable (such as 10 ms).

[0045] Through the follow-up navigation of tractor roads and inter-row roadways, effective tracking of the road can be achieved while keeping the trolley moving on the center line of the road, thus improving the accuracy and stability of intelligent driving.

[0046] The emergency obstacle avoidance module is used to perform millisecond-level electric braking when the detected distance to the obstacle ahead ≤ 0.3 m. The ultrasonic sensor is used to continuously sense the environmental information. When the distance to the obstacle is less than 0.3 m, the emergency braking mechanism is triggered to achieve millisecond-level response, ensuring safe operation to the greatest extent. The braking response time of the emergency braking system ≤ 0.1 s, and the braking distance does not exceed 0.15 m.

[0047] An intelligent driving method applied to an orchard includes a path rapid detection method, which is specifically as follows: Obtain the RGB image of the inter-row road; convert the image to the HSV color space; perform adaptive threshold segmentation to extract the binary mask of the drivable area; perform edge detection and probabilistic Hough line detection on the mask to obtain a candidate line set; use DBSCAN to cluster the candidate lines and merge the lines in the same lane; introduce the historical line parameters of the previous n frames and output the current center line by weighted average; send the center line to the lateral tracking control module to achieve line following driving.

[0048] Furthermore, it also includes an all-terrain adaptive control method, specifically: acquiring real-time three-dimensional terrain information ahead; obtaining a passability score S through a convolutional neural network; dynamically allocating the torque of the four motors according to S, locking the differential when S < 0.4; using RTK positioning when GNSS signal is good, switching to lidar or IMU or visual tightly coupled positioning when lock is lost; paralleling the aforementioned path rapid detection method to obtain the centerline; calculating the front wheel steering angle based on the centerline error and controlling the vehicle to travel along the road; and immediately executing electric braking when an obstacle is detected at a distance ≤ 0.3m.

[0049] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent driving system for use in orchards, characterized in that, include: The chassis is driven by four independent hub motors, with a ground clearance of ≥150mm, a peak torque of ≥100N·m per wheel, and is equipped with an electronic differential lock. The terrain perception module is used to acquire three-dimensional information of the terrain ahead and output a passability score S∈[0,1]; the dynamic torque distribution module is connected to the terrain perception module and the four motors and is used to adjust the torque of each motor in real time according to the score S; the multi-sensor fusion positioning module includes at least RTK-GNSS, lidar, IMU and binocular depth camera, and is used to switch to lidar or vision or IMU tight coupling mode when the GNSS signal is lost; the path fast detection module is used to perform HSV adaptive threshold segmentation, probabilistic Hough line detection and multi-frame historical weighted fusion on the inter-row road image and output a drivable centerline; the lateral tracking control module is used to calculate the front wheel steering angle according to the error between the centerline and the current position of the vehicle to realize line following. The emergency obstacle avoidance module is used to perform millisecond-level electric braking when an obstacle is detected at a distance of ≤0.3m.

2. The intelligent driving system applied to orchards according to claim 1, characterized in that: The terrain perception module includes: a binocular depth camera for acquiring point clouds in front; a convolutional neural network that takes the point clouds or disparity map as input and outputs slope, gully, and mud probability maps; and a passability scoring unit for weighted summation of the probability maps to obtain a score S.

3. The intelligent driving system applied to orchards according to claim 2, characterized in that: A binocular depth camera is used to acquire 3D information of the terrain ahead, and an ultrasonic sensor is used to detect nearby obstacles in real time to enhance the obstacle avoidance capability in complex terrain. A Kalman filter fusion algorithm is used to fuse multi-source information from the binocular depth camera and the ultrasonic sensor to achieve real-time 3D reconstruction of the terrain ahead. A convolutional neural network is used to train the terrain images acquired by the binocular depth camera. Based on the output of the convolutional neural network, a passability scoring model is constructed to generate a quantified terrain passability score S, which ranges from 0 to 1, where 0 is impassable and 1 is fully passable, providing data support for path planning.

4. The intelligent driving system applied to orchards according to claim 1, characterized in that: The dynamic torque distribution module uses a PD controller, which automatically locks the differential when S < 0.

4. Specifically, it uses a PD control algorithm to dynamically adjust the torque and speed of the four independent motors, and optimizes the traction distribution through real-time feedback and closed-loop control to ensure the stability and passability of the vehicle in complex terrain. The system integrates a differential lock control strategy, automatically switches the differential lock mode, and optimizes the torque distribution between wheels.

5. The intelligent driving system applied to orchards according to claim 1, characterized in that: The multi-sensor fusion positioning module employs extended Kalman filtering to dynamically weight multi-sensor data and performs global BA optimization upon successful loop closure detection to eliminate accumulated errors. Specifically, in areas with stable GNSS signals, RTK-GPS is used to provide global centimeter-level positioning, ensuring high-precision navigation. When entering a weak signal environment, it switches to a tightly coupled lidar-inertial IMU mode for local positioning. The use of extended Kalman filtering to dynamically weight multi-sensor data suppresses single-sensor failures or noise interference, and dynamically adjusts the fusion weights based on the confidence level of each sensor to ensure the robustness of the positioning system.

6. The intelligent driving system applied to orchards according to claim 5, characterized in that: The multi-sensor fusion localization module also includes: using a feature-point-based binocular visual SLAM algorithm for SLAM mapping. First, the intrinsic and extrinsic parameters of the binocular depth camera are calibrated offline, including the principal points, focal lengths, distortion coefficients, relative poses, and baseline lengths of the left and right cameras. Distortion correction and stereo correction are performed on the left and right images based on the distortion coefficients to obtain the left and right camera remapping matrix with coplanar row alignment. Then, the feature-point-based binocular visual SLAM algorithm is initialized to construct an initial 3D environment map. The initialization process includes feature extraction and feature matching of the left and right images to find corresponding points in the left and right images. Then, the depth of the feature points is calculated based on the disparity, and the 3D coordinates of all feature points are reconstructed. Finally, a direct method is used to assist in improving localization in scenes with missing textures, and IMU data is fused to compensate for localization loss caused by camera motion. For areas where visual SLAM has been run, if the camera moves along a non-closed-loop path, the localization result of the current image in the historical map is used as the current localization result, and BA optimization is performed on the data within the current trajectory.

7. The intelligent driving system applied to orchards according to claim 1, characterized in that: The path fast detection module includes: an HSV color space conversion unit for dynamically adjusting the threshold based on the prior interval of the yellow component of the orchard ground; a Gaussian filtering unit for suppressing high-frequency noise; a probabilistic Hough line detection unit for detecting road edges in the image; a DBSCAN clustering unit for merging similar lines; and a historical frame weighted fusion unit for caching the line parameters of the previous n frames and outputting a weighted average centerline, where n=5.

8. The intelligent driving system applied to orchards according to claim 1, characterized in that: The lateral tracking control module adopts PD control with a proportional coefficient Kp=0.8, a differential coefficient Kd=0.15, and a line-following error ≤5mm.

9. An intelligent driving method for orchards, employing an intelligent driving system for orchards as described in claim 1, characterized in that, The method includes a fast path detection approach, specifically: acquiring an RGB image of the road between ridges; converting the image to the HSV color space; adaptive threshold segmentation to extract a binary mask of the drivable area; performing edge detection and probabilistic Hough line detection on the mask to obtain a candidate line set; using DBSCAN to cluster the candidate lines and merge lines in the same lane; introducing historical line parameters from the previous n frames and weighted averaging to output the current centerline; and sending the centerline to the lateral tracking control module to achieve line-following driving.

10. The intelligent driving method applied to orchards according to claim 9, characterized in that: It also includes an all-terrain adaptive control method, specifically: real-time acquisition of three-dimensional terrain information ahead; obtaining a passability score S through a convolutional neural network; dynamically allocating the torque of the four motors according to S, locking the differential when S < 0.4; using RTK positioning when GNSS signal is good, and switching to lidar, IMU, or visual tightly coupled positioning when lock is lost; paralleling the aforementioned path fast detection method to obtain the centerline; calculating the front wheel steering angle based on the centerline error and controlling the vehicle to travel along the road; and immediately executing electric braking when an obstacle is detected at a distance ≤ 0.3m.

Images