Mulberry twig clamping and positioning method and system

Abstract

The invention discloses a mulberry twig clamping and positioning method and system, belongs to the technical field of intelligent mulberry leaf picking equipment, and aims to solve the technical problems that an existing RGB camera is easily interfered by illumination, and picking damage is caused due to the fact that a demarcation point of a tender branch and an old branch is difficult to accurately recognize. According to the method and system, 500-600 nm visible light and 750-900 nm near-infrared band images are synchronously collected through a multispectral imaging device, HSV color space parameters are extracted, and a color gradient model is established; by analyzing the gradient change rate of the hue value H and the saturation degree S and combining the curvature change rate and the near-infrared reflectivity difference value, the transition boundary of the tender and old branches is accurately judged, and the three-dimensional coordinates of the clamping datum point are generated by dynamically shifting 5-15 mm towards the side of the tender branches. According to the scheme, the limitation of single hue detection is broken through, the xylem recognition capability is enhanced by utilizing multispectral information, the clamping accuracy is improved, the method is suitable for automatic operation of a top-remaining picking mode in large-scale silkworm and mulberry planting, the picking efficiency can be effectively improved, and branch damage can be reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent mulberry leaf picking equipment technology, and more specifically, to a method and system for clamping and positioning mulberry branches. Background Technology

[0002] Currently, large-scale mulberry cultivation mostly adopts the top-leaf harvesting method, which means that when picking mulberry leaves, only the leaves from the upper, middle and lower parts of the branches are picked, while leaving a few leaves at the top of the branches. This harvesting method will not cause excessive damage to the mulberry tree, which is conducive to the rapid recovery and continuous growth of the mulberry tree, thus enabling multiple harvests. After the mulberry tree grows to 2-3 meters, it is completely cut down for the next batch of planting, thereby ensuring the yield and quality of mulberry leaves.

[0003] Current pruning and leaf-picking machinery cannot preserve the top of mulberry leaves without damaging them, thus manual harvesting is still the primary method, resulting in very low efficiency. Chinese patent application CN107309876A discloses a control method for robotic harvesting. This method utilizes a stereo vision system and GPS module to detect and identify the positional information of mulberry leaves, branches, and obstacles. It then calculates coordinates, extracts features from the leaves, branches, and obstacles, establishes a virtual 3D model, and plans a path. The robotic arm is then controlled to move to the designated position for harvesting. While this method solves the obstacle avoidance problem and reduces collisions, it cannot accurately and quickly find the boundary between preserving the top and harvesting.

[0004] Furthermore, existing RGB cameras often suffer from color recognition deviations due to changes in ambient lighting, and are prone to misjudgment under conditions of strong reflection. For branches with significant differences in diameter (3-15mm) and gradual transitions in bark color, existing methods struggle to accurately identify the boundary between young and old branches, easily causing the clamping point to deviate from the dense vascular bundle area, leading to branch breakage or bark damage, thus hindering the widespread use of automated harvesting equipment. Summary of the Invention

[0005] To address the above technical problems, this invention provides a mulberry branch clamping and positioning method and system. It uses multispectral imaging to overcome the limitations of traditional RGB cameras, enhances the ability to identify xylem by adding the near-infrared band, effectively distinguishes branch areas with similar bark colors but different degrees of lignification, and uses dual threshold determination to jointly detect hue value H and saturation S, avoiding the inaccuracy of clamping caused by the defect of a single hue parameter being affected by changes in light.

[0006] The purpose of this invention is to provide a method for clamping and positioning mulberry branches, comprising the following steps: S1. Real-time image data of mulberry branches are acquired through a multispectral imaging device, which includes the acquisition functions of the 500-600 nm visible light band and the 750-900 nm near-infrared band. S2. Extract the HSV color space parameters of the branch area in the image and establish a color gradient model between the tender branches and the old branches. The HSV value range of the tender branches is defined as hue value H in the range of 80~160, saturation S in the range of 40~100, and lightness V in the range of 30~70. The HSV value range of the old branches is defined as hue value H in the range of 20 to 50, saturation S in the range of 10 to 40, and lightness V in the range of 10 to 30. S3. Calculate the color gradient change rate along the branch axis. When the hue value difference between adjacent pixels is greater than or equal to 15 and the saturation difference is greater than or equal to 20, it is determined to be a color transition boundary. S4. Using the color transition boundary offset 5 mm to 15 mm towards the tender branch as the clamping reference point, generate positioning data containing three-dimensional coordinates; S5. Convert the coordinates of the positioning point in the image coordinate system to the base coordinate system of the robotic arm, and drive the gripping mechanism to move to the target position for gripping. Preferably, it also includes S6. Real-time acquisition of contact pressure data during clamping, and when the deviation between the measured pressure value and the preset pressure value exceeds 20%, triggering image reacquisition and executing closed-loop correction action.

[0007] Preferably, the multispectral imaging device in step S1 adopts a time-division multiplexed three-channel imaging mode, specifically including: The visible light channel includes a first filter with a center wavelength of 560±5 nm and a second filter with a center wavelength of 590±5 nm, while the near-infrared channel is equipped with a third filter with a center wavelength of 810±10 nm. The three filters switch in rotation with a period of 30 milliseconds, synchronously triggering the CMOS sensor to acquire 1024×768 pixel image data in each channel. The imaging resolution of the visible light channel reaches 0.1 mm / pixel, and the imaging resolution of the near-infrared channel reaches 0.15 mm / pixel. Furthermore, the multispectral imaging device is hardware-level synchronized with the depth sensor, synchronously acquiring point cloud data of the corresponding band each time the filter switches.

[0008] Preferably, the process of establishing the color gradient model in step S2 specifically includes: S21. Perform illumination compensation on the original HSV parameters and use an adaptive white balance algorithm based on the gray world assumption to adjust the average brightness value of the image to the range of 50-60 luminance. S22. Establish a dual-channel principal component analysis model. The first principal component channel corresponds to a linear combination of (H+0.8S), and the second principal component channel corresponds to a linear combination of (2V-S). Construct a two-dimensional feature space for young branches and old branches by calculating the principal component scores. S23. Dynamically adjust the HSV threshold boundary. When the ambient light intensity exceeds 80,000 lux, raise the lower limit of the hue value H of young branches to 100. When the ambient humidity exceeds 75%, lower the upper limit of the saturation S of old branches to 35.

[0009] Preferably, the execution process of the adaptive white balance algorithm in step S21 specifically includes: The acquired image is divided into 5×5 equal-area sub-regions, and the average brightness value of each sub-region is calculated. For sub-regions with average brightness values ​​exceeding the range of 40-70, a compensation weight of 0.1-0.9 is assigned according to their distance from the image center. Based on the weight values, the HSV parameters of each sub-region are non-linearly adjusted so that the average brightness of the overall image converges to the range of 52-58. When the rate of change of ambient light intensity exceeds 100 lux per second, the white balance calculation cycle is automatically shortened to 0.2 seconds, and the data is fused and calibrated with the light intensity sensor data of the multispectral imaging device, with the calibration error controlled within ±3 brightness units.

[0010] Preferably, the process of determining the color transition boundary in step S3 specifically includes: When calculating the color gradient change rate, the curvature change rate along the branch axis is analyzed simultaneously. When the hue value difference and saturation difference simultaneously meet the threshold conditions and the curvature change rate is less than 0.05 mm per millimeter, it is determined to be a valid color transition boundary. For mulberry branches with a diameter greater than 8 mm, the hue value difference threshold is dynamically increased to 18, and the near-infrared band reflectance difference is introduced as an auxiliary criterion. When the reflectance difference exceeds 15%, the three-dimensional point cloud data is resampled. If the hue value difference of adjacent pixels exceeds the threshold three times in a row but the curvature change rate is abnormal, the polarization mode of the multispectral imaging device is activated for secondary verification. The polarization angle is set to 45±5 degrees, and the color gradient value is corrected by the extinction ratio change.

[0011] Preferably, the process of determining the clamping reference point in step S4 specifically includes: The offset is dynamically adjusted based on the branch diameter. When the detected mulberry branch diameter is less than 5 mm, the offset towards the tender branch is set to 8-12 mm; when the diameter is between 5-10 mm, the offset is adjusted to 5-8 mm; and when the diameter is greater than 10 mm, the offset is increased to 12-15 mm. At the same time, the offset is dynamically compensated by combining the branch elastic modulus data measured in the near-infrared band. For every 100 MPa increase in elastic modulus, the offset is reduced by 1 mm. The final positioning data includes the three-dimensional coordinates of the clamping point, the recommended clamping pressure value, and the clamping angle correction value. The clamping angle correction value is calculated based on the branch curvature radius at that point to ensure that the angle deviation between the clamping surface and the axial tangent of the branch is less than 3 degrees.

[0012] Preferably, the closed-loop correction process in step S6 specifically includes: A pressure-image dual-modal feedback mechanism is established. When the contact pressure deviation exceeds a threshold, real-time vibration compensation images from a multispectral imaging device and three-dimensional torque data from a six-axis force sensor are simultaneously acquired. The pressure fluctuation spectrum characteristics are analyzed using a machine learning model to identify the source of the deviation as either branch slippage or foreign object interference. If it is branch slippage, image reacquisition with 0.5 mm precision is triggered; if it is foreign object interference, the clamping surface vibration cleaning mode is activated. During the correction process, an incremental PID control algorithm is used to dynamically adjust the proportional coefficient based on historical deviation data. For every 5% increase in pressure deviation, the proportional coefficient increases by 0.2, and the integral time constant decays exponentially. After the correction is completed, the preset pressure reference value is automatically updated. The new reference value is the moving average of the last 10 effective clamping pressures and is stored in the branch feature database to form a self-learning curve.

[0013] The present invention provides a mulberry branch clamping and positioning system, which, when performing any of the methods described above, includes: The multispectral imaging device includes a switchable filter wheel, a CMOS sensor, and a lens, used to acquire images of mulberry branches in the 500-600 nm visible light band and the 750-900 nm near-infrared band. The image processing unit is communicatively connected to the multispectral imaging device and is configured to convert the image to the HSV color space and perform principal component analysis to establish a principal component channel containing a linear combination of hue value H and saturation value S. The color gradient analysis module is used to calculate the rate of change of hue and saturation gradients along the branch axis. When the difference in hue value between adjacent pixels is ≥15 and the difference in saturation value is ≥20, it is determined to be a color transition boundary. The clamping point calculation module generates three-dimensional coordinates of the clamping reference point by offsetting it 5-15 mm towards the tender branch based on the color transition boundary. The offset is dynamically adjusted according to the branch diameter. The coordinate transformation system converts the image coordinate system into the robot arm's base coordinate system, driving the robot arm to move to the target position for clamping.

[0014] Preferably, the mulberry branch clamping and positioning system further includes: The pressure closed-loop control module includes a six-dimensional force sensor that collects contact pressure data in real time. When the deviation is greater than 20%, it triggers image reacquisition and clamping point correction.

[0015] The present invention has at least the following beneficial effects: 1. This invention utilizes multispectral imaging technology, integrating 500-600nm visible light and 750-900nm near-infrared band data, to accurately distinguish between tender branches and old branches, effectively solving the problem that traditional RGB cameras are easily affected by light interference and have difficulty accurately identifying the boundary between tender and old branches.

[0016] 2. This invention employs a color gradient model constructed based on HSV color space parameters and principal component analysis, combined with data such as branch diameter and elastic modulus, to dynamically adjust the offset of the clamping reference point. When the branch diameter is less than 5 mm, the offset is 8-12 mm; when the diameter is between 5-10 mm, the offset is adjusted to 5-8 mm; and when the diameter is greater than 10 mm, the offset increases to 12-15 mm. Simultaneously, for every 100 MPa increase in elastic modulus, the offset decreases by 1 mm, ensuring that the clamping point avoids vulnerable areas with dense vascular bundles, and that the angular deviation between the clamping surface and the branch axial tangent is less than 3 degrees, reducing the risk of branch breakage and epidermal damage.

[0017] 3. This invention can dynamically adjust the HSV threshold boundary based on ambient light intensity and humidity. When the ambient light intensity exceeds 80,000 lux, the lower limit of the hue value H of young branches is increased to 100; when the ambient humidity exceeds 75%, the upper limit of the saturation S of old branches is reduced to 35. Furthermore, auxiliary criteria such as the rate of curvature change and the difference in near-infrared reflectivity are introduced, along with secondary verification using a polarized light mode, to ensure accurate identification of the transition boundary between young and old branches even in complex environments, reducing the system's misjudgment rate under harsh conditions such as strong reflection.

[0018] 4. This invention realizes automated top-harvesting operation, which effectively improves the efficiency of mulberry leaf harvesting.

[0019] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the mulberry branch clamping and picking device of the present invention; Figure 2 This is a schematic diagram of the positioning system of the present invention; Figure 3 This image illustrates the difference and transition between old and young branches of mulberry twigs in this invention.

[0021] Among them, there are leaf-holding mechanism 10, telescopic cylinder 101, clamping mechanism 20, actuator 201, laser displacement sensor 202, leaf-plucking mechanism 30, multispectral imaging device 40, bracket 401, depth sensor 402, CMOS sensor 403, robotic arm 50, environmental sensor 60, and six-dimensional force sensor 70. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to examples, so that those skilled in the art can implement it based on the description.

[0023] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0024] The harvesting device of this invention can use the automatic mulberry leaf harvester of publication number CN107309876B, or other existing harvesting machines, or the mulberry leaf harvesting device given in this invention with mulberry branch clamping as described below, such as... Figure 1 The diagram shows the structure of a mulberry leaf picking device, which includes: a leaf-holding mechanism 10, a telescopic cylinder 101, a clamping mechanism 20, a leaf-plucking mechanism 30, a multispectral imaging device 40, a support 401, a depth sensor 402, a CMOS sensor 403, a robotic arm 50, an environmental sensor 60, and a six-dimensional force sensor 70.

[0025] A six-dimensional force sensor 70 is connected between the robotic arm 50 and the clamping mechanism 20. A multispectral imaging device 40 is mounted on the front support 401 of the robotic arm 50 and is installed parallel to the depth sensor 402. The distance between the central axes is controlled within 50mm. An environmental sensor 60 is set above the multispectral imaging device 40.

[0026] The leaf-plucking mechanism 30 is located directly below the clamping mechanism 20. The leaf-plucking mechanism 30 has the same structure as the clamping mechanism 20, the only difference being the clamping force. Its clamping force is only 1-10% of that of the clamping mechanism 20. Both the clamping mechanism 20 and the leaf-plucking mechanism 30 insert mulberry branches. After the clamping mechanism 20 completes the positioning and clamping of the mulberry branch, the leaf-plucking mechanism 30 moves downwards to pluck the mulberry leaves from the branch. The leaf-holding mechanism 10 includes a left half-frame and a right half-frame, using an aluminum alloy profile frame. The left and right half-frames are hinged to form a ring structure. A double-acting telescopic cylinder 101 is installed in the middle to drive the opening and closing. After the clamping mechanism 20 fixes the mulberry branch, the cylinder 101 pushes the left and right half-frames to close, forming an enclosing space with a diameter of 80-200mm, encircling the mulberry branch. The leaf-receiving frame receives the mulberry leaves below. A negative pressure device can also be added to achieve rapid adsorption and collection of mulberry leaves.

[0027] like Figure 1 and 2 As shown, a mulberry branch clamping and positioning method of the present invention includes the following steps: S1. Real-time image data of mulberry branches are acquired through a multispectral imaging device, which includes the acquisition functions of the 500-600 nm visible light band and the 750-900 nm near-infrared band.

[0028] Specifically, the multispectral imaging device 40 is mounted on the front support 401 of the robotic arm 50, parallel to the depth sensor 402, with the central axis distance controlled within 50mm. The multispectral imaging device includes a CMOS sensor and its lens for image acquisition, and a filter wheel mechanism for filter switching, located 20mm in front of the CMOS sensor. The depth sensing unit can be fixed to the support by screws or other means, maintaining a 50mm horizontal distance from the filter wheel.

[0029] The visible light channel of the multispectral imaging device can be selected at center wavelengths of 560nm, 565nm, 590nm, and 595nm, while the near-infrared channel can be selected at 800nm, 810nm, and 820nm. Imaging resolution is selectable at 0.08mm / px, 0.1mm / px, and 0.12mm / px for the visible light channel, and 0.13mm / px, 0.15mm / px, and 0.18mm / px for the near-infrared channel. Filter switching cycle is selectable at 25ms, 30ms, and 35ms. A Sony IMX415 CMOS sensor can be used, equipped with Thorlabs FBH560-10, FBH590-10, and FBH810-10 filters. The depth sensor can be an Intel RealSense L515 LiDAR. The support material can be 6061 aluminum alloy.

[0030] During the commissioning of the multispectral imaging device, the USAF 1951 resolution test chart was used for calibration. The visible light channel achieved a resolution of 128 line pairs / mm at 560nm, and the near-infrared channel achieved 85 line pairs / mm at 810nm. After multispectral data fusion, the recognition accuracy of mulberry branch bark texture reached 0.08mm within a working distance of 1.2 meters, and the xylem feature extraction error was less than 5%.

[0031] During operation, the filter wheel rotates at 30ms intervals to switch the positions of the three filters. Each switch triggers the CMOS sensor to acquire the corresponding spectral band image, achieving efficient acquisition of three-band spectral data. The visible light channel provides epidermal color characteristics, while the near-infrared channel enhances the detection capability of lignification. Simultaneously, the depth sensor is triggered to acquire point cloud data. Image data is transmitted to the main control computer via Gigabit Ethernet, and point cloud data is transmitted via USB 3.0 interface.

[0032] S2. Extract the HSV color space parameters of the branch region in the image, and establish a color gradient model between young and old branches. The HSV value range for young branches is defined as hue (H) between 80 and 160, saturation (S) between 40 and 100, and lightness (V) between 30 and 70. The HSV value range for old branches is defined as hue (H) between 20 and 50, saturation (S) between 10 and 40, and lightness (V) between 10 and 30. See the table below. Figure 3 As shown: Specifically, the environmental sensor 60 uses a Bosch BME680 temperature and humidity sensor and a TI OPT3001 light intensity sensor, which are fixed to the top area of ​​the imaging device housing using screws or other means. The center point of the sensor is 50 mm away from the optical center of the lens. The light intensity sensor collects ambient light data at a frequency of 10 times per second, with a measurement range covering 0 to 100,000 lux. The temperature and humidity sensor simultaneously monitors the ambient temperature and humidity, and the data is transmitted to the main control unit via an I2C bus.

[0033] The image processing unit uses an Intel i7-12700H processor and connects to the image acquisition card via a PCIe 4.0 interface. The processing flow first inputs images in the 560nm, 590nm, and 810nm bands acquired by the multispectral imaging device into the system, and uses OpenCV library color gamut conversion functions to convert the images to the HSV color space. In the principal component analysis stage, two feature channels are constructed. The first channel uses a linear combination of hue and saturation values, with coefficients selectable within the range of H+0.7S to H+0.9S. The second channel uses a combination of lightness and saturation values, with coefficients including options such as 2V-S, 2.2VS, or 1.8VS.

[0034] During implementation, the light intensity sensor in the environmental sensor group collects ambient light intensity data in real time, while the temperature and humidity sensor simultaneously measures air temperature and humidity. The collected data is transmitted to the PCIe interface of the image processing unit via the I2C bus. The image processing unit calls the OpenCV library to perform HSV conversion on the 560nm / 590nm / 810nm three-band images acquired by the multispectral imaging device, using a feature combination of H+0.8S and 2V-S during principal component analysis. Based on the real-time data from the environmental sensor group, the dynamic threshold adjustment module raises the lower limit of the H threshold for young branches from 80 to 100 when the light intensity is >80000 lux, while simultaneously lowering the upper limit of saturation for older branches from 40 to 35. When the humidity sensor measures humidity >75%, the upper limit of the V threshold for older branches is lowered from 30 to 25, and the lower limit of saturation for young branches is raised from 40 to 45. The processed feature data is transmitted to the robotic arm control unit via Gigabit Ethernet.

[0035] S3. Calculate the color gradient change rate along the branch axis. When the hue value difference between adjacent pixels is greater than or equal to 15 and the saturation difference is greater than or equal to 20, it is determined to be a color transition boundary.

[0036] Specifically, in the color gradient analysis module, after the CMOS sensor of the multispectral imaging device acquires an image in the 560nm band, the main control computer of the image processing unit starts the color gradient analysis program. In the processing flow, OpenCV is first used to extract the branch axis path, and a detection window is set with a step size of 3 pixels along the axis. Within each window, the hue value difference ΔH and saturation value difference ΔS between adjacent pixels are calculated. When three consecutive windows meet the conditions of ΔH≥15 and ΔS≥20, they are marked as candidate transition boundaries. This step effectively overcomes the defect of single hue parameters being affected by illumination interference by using a dual threshold (ΔH≥15 and ΔS≥20) joint determination.

[0037] S4. Using the color transition boundary offset 5 mm to 15 mm towards the tender branch as the clamping reference point, generate positioning data containing three-dimensional coordinates.

[0038] Specifically, in the clamping point calculation module, the clamping reference point is determined using a dynamic offset mechanism. The offset can be adjusted based on the branch diameter within the range of 5 mm to 15 mm. The specific parameters are set as follows: 8-12 mm for a branch diameter less than 5 mm, 5-8 mm for a diameter of 5-10 mm, and 12-15 mm for a diameter greater than 10 mm. The 3D point cloud data acquired by the depth sensor and the 2D image data from the multispectral imaging device are fused using the point cloud registration algorithm in the PCL library to generate the 3D coordinates of the clamping point.

[0039] During implementation, the initial offset is first calculated along the normal direction of the color transition boundary, and then three-dimensional coordinates are generated by combining the point cloud data from the depth sensor. When the robotic arm end effector 201 moves to within 5 mm of the target point, the laser displacement sensor 202 scans the surface contour at a density of 500 points / second to correct the gripping point position in real time.

[0040] S5. Convert the coordinates of the positioning point in the image coordinate system to the base coordinate system of the robotic arm, and drive the gripping mechanism to move to the target position for gripping.

[0041] Specifically, the marking plate is first fixed to the calibration platform within the robotic arm's working area, ensuring that the parallelism error between the marking plate plane and the XY plane of the robotic arm's base coordinate system is less than 0.1 degrees. A calibration cone tool is installed on the robotic arm's end effector 201, and the positional deviation between the tip of the six-dimensional force sensor tool and the center point of the six-dimensional force sensor is controlled within 0.05 mm. The robotic arm is then moved sequentially to 20 preset calibration poses, remaining stationary for 2 seconds in each pose to collect data. The X / Y / Z axis movement range is set to ±300 mm, and the rotation angle range around the Z-axis is set to ±30 degrees. A multispectral imaging device acquires images of the marking plate in each pose, simultaneously recording the angle data from the robotic arm's joint encoders. The image processing unit uses OpenCV's Aruco module to detect the marking corners and calculates the transformation matrix from the image coordinate system to the marking plate coordinate system. The Tsai-Lenz algorithm is used for hand-eye calibration calculations, with 200 iterations and a convergence threshold of 0.01 mm. After calibration, five verification poses were randomly selected, and the actual coordinate deviations were measured using an API Radial laser tracker, with a maximum error requirement of no more than 0.1 mm. During the motion planning phase, the OMPL planner of the ROS MoveIt framework was called, setting the maximum trajectory speed to 0.5 m / s and the maximum acceleration to 0.3 m / s². The joint space trajectory was generated using 7th-order polynomial interpolation, and then the robotic arm was driven to move according to the space trajectory.

[0042] S6. During the clamping process, the contact pressure data is collected in real time. When the deviation between the measured pressure value and the preset pressure value exceeds 20%, the image is reacquired and a closed-loop correction action is executed.

[0043] Specifically, the pressure closed-loop control module monitors the clamping contact force in real time using a six-dimensional force sensor, with a sampling frequency set to 1 kHz. When the deviation between the measured pressure and the preset value exceeds 20%, the control unit triggers a three-level correction process: first, the robotic arm movement is frozen and the multispectral imaging device is activated to reacquire images; second, the S4 module is called to recalculate the clamping point coordinates; finally, the gripper position is finely adjusted at a speed of 0.1 mm / s, while the clamping force is adjusted in 5% increments. Pressure data and vibration spectrum (0-500 Hz) are analyzed using Fast Fourier Transform to distinguish between branch slippage (dominant frequency < 100 Hz) and foreign object interference (dominant frequency > 200 Hz). The PID control parameters are set to proportional coefficient Kp = 0.8, integral time Ti = 0.5 seconds, and derivative time Td = 0.1 seconds. The preset pressure reference value is dynamically adjusted according to the branch diameter: 8 N for a diameter of 5 mm, and 1.2 N for every 1 mm increase in diameter.

[0044] During clamping, the six-dimensional force sensor transmits XYZ three-axis force data to the control unit in real time. When the pressure deviation exceeds 20% for more than 0.2 seconds, the operation is paused and the multispectral imaging device is activated to reacquire a local image (area size 50×50 mm). During the correction phase, a PID algorithm is used to adjust the clamping parameters every 0.05 seconds.

[0045] In this implementation, multispectral imaging overcomes the limitations of traditional RGB cameras by adding a near-infrared band to enhance xylem recognition capabilities, effectively distinguishing branch areas with similar epidermal colors but different degrees of lignification. Dual-threshold determination employs joint detection of both H and S parameters, avoiding the drawback of a single hue parameter being affected by changes in light intensity. The offset positioning mechanism is designed based on plant physiological characteristics, avoiding transitional zones (vulnerable areas) with dense vascular bundles.

[0046] Furthermore, in another embodiment, the multispectral imaging device in step S1 employs a time-division multiplexed three-channel imaging mode, specifically including: The visible light channel includes a first filter with a center wavelength of 560±5 nm and a second filter with a center wavelength of 590±5 nm, while the near-infrared channel is equipped with a third filter with a center wavelength of 810±10 nm. The three filters switch in rotation with a period of 30 milliseconds, synchronously triggering the CMOS sensor to acquire 1024×768 pixel image data in each channel. The imaging resolution of the visible light channel reaches 0.1 mm / pixel, and the imaging resolution of the near-infrared channel reaches 0.15 mm / pixel. Furthermore, the multispectral imaging device is hardware-level synchronized with the depth sensor, synchronously acquiring point cloud data of the corresponding band each time the filter switches.

[0047] Specifically, during a single operation, the filter wheel rotates at 30ms intervals to switch the positions of the three filters. After switching, an external trigger signal synchronously initiates CMOS sensor exposure: 5ms for the visible light channel and 8ms for the near-infrared channel. Each channel acquires a 1024×768 pixel image, which is then transmitted to the main control unit via Gigabit Ethernet.

[0048] The visible light channel resolution can be set to 0.08 mm / pixel, 0.1 mm / pixel, or 0.12 mm / pixel, with 0.1 mm / pixel being preferred. The near-infrared channel resolution can be set to 0.13 mm / pixel, 0.15 mm / pixel, or 0.18 mm / pixel, with 0.15 mm / pixel being preferred. The depth sensor point cloud acquisition frequency is consistent with the filter switching cycle, preferably 30 ms.

[0049] When the filter is switched, a synchronization pulse triggers the depth sensor to acquire point cloud data. Microsecond-level timestamps are added to the visible light image and point cloud data, and registration is performed using the ICP algorithm, with the registration error controlled within ±0.1mm. The registered data includes pixel-level 3D coordinates, HSV parameters, and near-infrared reflectance.

[0050] This implementation utilizes a specific combination of the 560 / 590nm dual visible light band and the 810nm near-infrared band. The former accurately captures the chlorophyll absorption peak (around 580nm) and the carotenoid characteristic region (550-600nm), while the latter sensitively reflects differences in lignin content. The higher resolution (0.1mm / px) of the visible light channel ensures accurate color gradient detection, while the lower resolution (0.15mm / px) of the near-infrared band matches its superior penetration characteristics. This combination enhances the system's effective detection distance in dense foliage, making it suitable for automated clamping operations in environments with dense mulberry foliage.

[0051] Furthermore, in another embodiment, the process of establishing the color gradient model in step S2 specifically includes: S21. Perform illumination compensation on the original HSV parameters, using an adaptive white balance algorithm based on the gray-scale world assumption to adjust the average brightness value of the image to the range of 50-60 luminance; during white balance processing, the image can be divided into 4×4, 5×5 or 6×6 sub-regions, and the compensation weight coefficient can be set to a linear gradient distribution of 0.1-0.9; S22. Establish a dual-channel principal component analysis model. The first principal component channel corresponds to a linear combination of (H+0.8S), and the second principal component channel corresponds to a linear combination of (2V-S). Construct a two-dimensional feature space for young branches and old branches by calculating the principal component scores. The principal component analysis model can select H+0.7S, H+0.8S, or H+0.9S as the first principal component and 2V-S, 2.1VS, or 1.9VS as the second principal component. S23. Dynamically adjust the HSV threshold boundary. When the ambient light intensity exceeds 80,000 lux, raise the lower limit of the hue value H of the young branches to 100. When the ambient humidity exceeds 75%, lower the upper limit of the saturation S of the old branches to 35. When dynamically adjusting the threshold, the light intensity trigger point can be set to 75,000 lux, 80,000 lux, or 85,000 lux, and the humidity threshold can be set to 70%, 75%, or 80%.

[0052] In one workflow, after acquiring a multispectral image, the image is divided into 5×5 sub-regions, and the average brightness value of each sub-region is calculated. For sub-regions with brightness values ​​exceeding the range of 40-70, a compensation weight of 0.1-0.9 is assigned based on their distance from the image center. A non-linear adjustment algorithm is used to correct the HSV parameters of each sub-region, bringing the overall image average brightness into the range of 52-58. When the rate of change in illumination intensity exceeds 100 lux / second, the white balance calculation cycle is shortened to 0.2 seconds.

[0053] The HSV parameters are input into the principal component analysis model to calculate the linear combination score of the first principal component (H+0.8S) and the second principal component (2V-S). In the two-dimensional feature space, the data points of young branches are concentrated in the first quadrant (score ≥ 0.6), and the data points of old branches are concentrated in the third quadrant (score ≤ 0.4). The model dynamically divides the decision boundary between young and old branches using an SVM classifier.

[0054] The system monitors ambient light intensity and humidity data in real time. When the light intensity exceeds 80,000 lux, the lower limit of the H value for young branches is automatically increased from 80 to 100; when the humidity exceeds 75%, the upper limit of the S value for older branches is decreased from 40 to 35. The adjusted thresholds are transmitted to the image processing unit via the CAN bus to update the classification model parameters.

[0055] In this implementation, (H+0.8S) and (2V-S) are used as principal component parameters to improve the differentiation between young and old mulberry branches and the recognition accuracy under low visibility conditions. By establishing a two-factor compensation of light and humidity, the H value threshold is automatically corrected when strong light causes hue shift, and the S value tolerance is reduced in high humidity environments, thereby improving the system's all-weather operational stability in open mulberry orchards.

[0056] Furthermore, in another embodiment, the execution process of the adaptive white balance algorithm in step S21 specifically includes: The acquired image is divided into 5×5 equal-area sub-regions, and the average brightness value of each sub-region is calculated. For sub-regions with average brightness values ​​exceeding the range of 40-70, a compensation weight of 0.1-0.9 is assigned according to their distance from the image center. Based on the weight values, the HSV parameters of each sub-region are non-linearly adjusted so that the average brightness of the overall image converges to the range of 52-58. When the rate of change of ambient light intensity exceeds 100 lux per second, the white balance calculation cycle is automatically shortened to 0.2 seconds, and the data is fused and calibrated with the light intensity sensor data of the multispectral imaging device, with the calibration error controlled within ±3 brightness units.

[0057] In one workflow, after acquiring a multispectral image, the image is divided into 5×5 equal-area sub-regions. The average brightness value of each sub-region is calculated. If the brightness value of a sub-region exceeds the range of 40-70, a weight of 0.1-0.9 is assigned based on its distance from the image center. The central region has the highest weight (0.9), and the edge regions have the lowest (0.1). A nonlinear adjustment algorithm is used to weight and correct the HSV parameters of each sub-region, so that the average brightness of the overall image converges to the range of 52-58.

[0058] The system monitors the rate of change in ambient light intensity in real time. When the rate of change exceeds 100 lux / second, the white balance calculation cycle is automatically shortened from the default 1 second to 0.2 seconds. After shortening the cycle, real-time data from the light intensity sensor is synchronously read and fused with the image white balance parameters using a weighted average algorithm to ensure that the calibration error does not exceed ±3 luminance units. The calibrated parameters are then transmitted to the image processing unit via the CAN bus.

[0059] This implementation improves the white balance algorithm's adaptability to sudden changes in illumination through block-weighted compensation and dynamic periodic adjustment. The high-weight design in the central region reduces edge reflection interference, and the dynamic calibration mechanism mitigates the impact of ambient light fluctuations, thereby enhancing the accuracy of mulberry branch color feature extraction and providing a stable and reliable image data foundation for subsequent clamping and positioning.

[0060] Furthermore, in another embodiment, the process of determining the color transition boundary in step S3 specifically includes: When calculating the color gradient change rate, the curvature change rate along the branch axis is analyzed simultaneously. When the hue value difference and saturation difference simultaneously meet the threshold conditions and the curvature change rate is less than 0.05 mm per millimeter, it is determined to be a valid color transition boundary. For mulberry branches with a diameter greater than 8 mm, the hue value difference threshold is dynamically increased to 18, and the near-infrared band reflectance difference is introduced as an auxiliary criterion. When the reflectance difference exceeds 15%, the three-dimensional point cloud data is resampled. If the hue value difference of adjacent pixels exceeds the threshold three times in a row but the curvature change rate is abnormal, the polarization mode of the multispectral imaging device is activated for secondary verification. The polarization angle is set to 45±5 degrees, and the color gradient value is corrected by the extinction ratio change.

[0061] Specifically, the curvature change rate threshold can be set to 0.03 mm / mm, 0.05 mm / mm, or 0.07 mm / mm, and the dynamic adjustment coefficient for hue difference can be selected as 0.3 / mm, 0.4 / mm, or 0.5 / mm. The polarization component can be a Thorlabs LPVIS100-MP2 linear polarizer, mounted on a rotatable bracket 401, with an angle adjustment range of ±10 degrees. The point cloud resampling density can be set to 400 points / cm², 500 points / cm², or 600 points / cm².

[0062] During processing, the color gradient along the branch axis is first calculated. Curvature verification is triggered when the hue difference between adjacent pixels is ≥15 and the saturation difference is ≥20. Curvature calculation uses cubic spline interpolation, with sampling window widths set to 5mm, 7mm, or 9mm. For branches with a diameter >8mm, the near-infrared reflectance difference is analyzed simultaneously, and an SVM classifier is used for boundary type determination. In abnormal situations, polarized light mode is activated, and extinction images are acquired at a 45-degree angle. The extinction ratio change is calculated using OpenCV.

[0063] In one working process, color gradient data is extracted along the branch axis with a window width of 7mm, and the hue difference and saturation difference between adjacent pixels are calculated. If the hue difference is ≥15 and the saturation difference is ≥20, the curvature change rate of the region is calculated simultaneously. When the curvature change rate is <0.05mm / mm, it is determined to be a valid color transition boundary.

[0064] When a mulberry branch diameter > 8mm is detected, the hue difference threshold is automatically increased from 15 to 18. Simultaneously, the near-infrared reflectance difference is analyzed; if the difference exceeds 15%, 3D point cloud resampling (density 500 points / cm²) is triggered. After resampling, the clamping point coordinates are updated to avoid misjudgments caused by interference from branch surface texture.

[0065] When the hue difference between adjacent pixels exceeds the threshold three times consecutively but the rate of curvature change is abnormal, the polarized light mode is activated. The rotating bracket is used to adjust the polarizer to 45±5 degrees, and an extinction image is acquired. The change in extinction ratio is calculated using OpenCV. If the change is ≥15%, the color gradient value is corrected and the boundary is re-evaluated.

[0066] This implementation improves the accuracy of identifying mulberry branch transition boundaries by combining color gradient, curvature change, and near-infrared reflectivity parameters. Dynamic threshold adjustment adapts to the characteristic differences of branches with different diameters, and a polarized light secondary verification mechanism reduces the risk of misjudgment in complex environments, providing a reliable basis for automated clamping.

[0067] Furthermore, in another embodiment, the process of determining the clamping reference point in step S4 specifically includes: The offset is dynamically adjusted based on the branch diameter. When the detected mulberry branch diameter is less than 5 mm, the offset towards the tender branch is set to 8-12 mm; when the diameter is between 5-10 mm, the offset is adjusted to 5-8 mm; and when the diameter is greater than 10 mm, the offset is increased to 12-15 mm. At the same time, the offset is dynamically compensated by combining the branch elastic modulus data measured in the near-infrared band. For every 100 MPa increase in elastic modulus, the offset is reduced by 1 mm. The final positioning data includes the three-dimensional coordinates of the clamping point, the recommended clamping pressure value, and the clamping angle correction value. The clamping angle correction value is calculated based on the branch curvature radius at that point to ensure that the angle deviation between the clamping surface and the axial tangent of the branch is less than 3 degrees.

[0068] Specifically, the diameter dividing points can be set to 4mm, 5mm, 6mm, 10mm, 12mm, and 14mm. Mulberry branch segments with diameters of 3-15mm are selected, impurities are removed from the surface, and test points are marked to prepare samples. Near-infrared spectral data uses the absorbance at 850nm. Three spectral data points are collected for each sample, and the average is taken. Simultaneously, the elastic modulus is measured using a universal testing machine to establish a correspondence model between the near-infrared spectrum and the elastic modulus of mulberry branches. Based on this model, the elastic modulus data of the branches can be obtained from the near-infrared band. The clamping angle correction can also be set with allowable deviations of ±3 degrees, ±5 degrees, or ±7 degrees.

[0069] During the operation, the diameter data of the mulberry branch is acquired by a laser rangefinder. If the diameter is <5mm, the clamping reference point is offset 8-12mm towards the tender branch; if the diameter is 5-10mm, the offset is 5-8mm; if the diameter is >10mm, the offset is 12-15mm. The offset is converted into three-dimensional coordinates by the robotic arm control unit, which drives the clamping mechanism to move.

[0070] Near-infrared absorbance data in the 850nm band were collected from mulberry branch samples with diameters ranging from 3 to 15 mm. Each sample was measured three times, and the average value was taken. Simultaneously, the elastic modulus was measured using a universal testing machine, and a linear regression model between absorbance and elastic modulus was established. In actual operation, the elastic modulus was inferred from the near-infrared data; for every 100 MPa increase, the clamping point offset was reduced by 1 mm.

[0071] The angle between the clamping surface and the axial tangent is calculated based on the radius of curvature of the branch. If the angle deviation exceeds 3 degrees, the pitch angle of the clamping mechanism is adjusted. The final positioning data includes three-dimensional coordinates, recommended clamping pressure (e.g., 9.6N for a diameter of 8mm), and angle correction values, which are transmitted to the robotic arm for execution via the EtherCAT bus.

[0072] This implementation reduces the risk of damage to the mulberry branch bark during clamping through a dynamic offset mechanism involving diameter segmentation and elastic modulus compensation. Near-infrared modeling enables non-destructive testing of mechanical properties, and angle correction ensures the clamping surface adheres to the branch surface, comprehensively improving the positioning accuracy and reliability of automated harvesting.

[0073] Furthermore, in another embodiment, the closed-loop correction process in step S6 specifically includes: A pressure-image dual-modal feedback mechanism is established. When the contact pressure deviation exceeds a threshold, real-time vibration compensation images from a multispectral imaging device and three-dimensional torque data from a six-axis force sensor are simultaneously acquired. The pressure fluctuation spectrum characteristics are analyzed using a machine learning model to identify the source of the deviation as either branch slippage or foreign object interference. If it is branch slippage, image reacquisition with 0.5 mm precision is triggered; if it is foreign object interference, the clamping mechanism vibration is activated. During the correction process, an incremental PID control algorithm is used to dynamically adjust the proportional coefficient based on historical deviation data. For every 5% increase in pressure deviation, the proportional coefficient increases by 0.2, and the integral time constant decays exponentially. After the correction is completed, the preset pressure reference value is automatically updated. The new reference value is the moving average of the last 10 effective clamping pressures and is stored in the branch feature database to form a self-learning curve.

[0074] Specifically: Pressure-image dual-modal feedback mechanism The pressure deviation threshold can be set to 15%, 20%, or 25%, preferably 20%. The vibration compensation image acquisition accuracy can be set to 0.3mm, 0.5mm, or 0.7mm, preferably 0.5mm.

[0075] The device can be equipped with a six-axis force sensor with a range of ±50N and an accuracy of ±0.5%. It can also be equipped with a vibration-compensated camera with a frame rate of ≥120fps and support for global shutter.

[0076] During one operation, when the six-axis force sensor detects a deviation of more than 20% between the contact pressure and the preset value, it simultaneously triggers a vibration-compensated camera to acquire real-time images and read three-dimensional torque data. The images and torque data are transmitted to the main control unit via Ethernet for deviation analysis.

[0077] Deviation source identification and correction action triggering The threshold for dividing the pressure spectrum main frequency can be set to 100Hz, 150Hz, or 200Hz, with 100Hz being preferred. The size of the image reacquisition area can be set to 40×40mm, 50×50mm, or 60×60mm, with 50×50mm being preferred.

[0078] The device can utilize an embedded AI acceleration module, supporting the TensorFlow Lite framework. A miniature vibration motor with an amplitude of 0.1-0.5mm and a frequency of 10-100Hz can be selected. The AI ​​acceleration module is integrated into the main control circuit board and connected via a PCIe slot. The vibration motor is embedded inside the clamping mechanism, ≤20mm from the clamping surface.

[0079] In one operation, a machine learning model analyzes the spectral characteristics of pressure fluctuations. If the dominant frequency is less than 100 Hz, it is determined to be branch slippage, triggering a local image re-acquisition with a precision of 0.5 mm; if the dominant frequency is greater than 200 Hz, it is determined to be foreign object interference, and the clamping mechanism is activated to vibrate (amplitude 0.3 mm, frequency 50 Hz) to remove the foreign object.

[0080] Incremental PID control algorithm The scaling factor adjustment step can be set to 0.1, 0.2, or 0.3, preferably 0.2. The integration time decay factor can be set to 0.8, 0.85, or 0.9, preferably 0.85. The device can be a real-time controller supporting μs-level timer interrupts. Non-volatile memory with a write / erase cycle of ≥100,000 times can be used.

[0081] During operation, the PID parameters are dynamically adjusted based on historical pressure deviation data. For every 5% increase in pressure deviation, the proportional gain increases by 0.2; the integral time constant is calculated using the formula: Attenuation (n is the number of corrections). Correction commands are sent to the robotic arm drive unit via the EtherCAT bus.

[0082] Stress benchmark update and self-learning mechanism The moving average sampling frequency can be set to 8, 10, or 12 times, with 10 times being preferred. The database storage entries can be set to 1000, 2000, or 5000 groups, with 2000 groups being preferred. A relational database module supporting the SQLite protocol can be used. A data acquisition card with a sampling rate of 1kHz and a resolution of 16bit can be selected.

[0083] During a certain operation, after correction, the moving average of the 10 most recent effective clamping pressures is taken as the new baseline value. The new value, along with the branch diameter and elastic modulus data, is stored in the database for adaptive adjustment in subsequent clamping operations.

[0084] This implementation achieves real-time closed-loop correction of the clamping process through pressure-image dual-modal feedback and adaptive PID control. The machine learning model accurately distinguishes the sources of deviation, and the self-learning mechanism accumulates historical data to optimize baseline parameters, thereby improving the system's stability and fault tolerance under complex operating conditions.

[0085] The present invention provides a mulberry branch clamping and positioning system, which performs any of the mulberry branch clamping and positioning methods described in any one of the claims, including: The multispectral imaging device includes a switchable filter wheel, a CMOS sensor, and a lens, used to acquire images of mulberry branches in the 500-600 nm visible light band and the 750-900 nm near-infrared band. The image processing unit is communicatively connected to the multispectral imaging device and is configured to convert the image to the HSV color space and perform principal component analysis to establish a principal component channel containing a linear combination of hue value H and saturation value S. The color gradient analysis module is used to calculate the rate of change of hue and saturation gradients along the branch axis. When the difference in hue value between adjacent pixels is ≥15 and the difference in saturation value is ≥20, it is determined to be a color transition boundary. The clamping point calculation module generates three-dimensional coordinates of the clamping reference point by offsetting it 5-15 mm towards the tender branch based on the color transition boundary. The offset is dynamically adjusted according to the branch diameter. The coordinate transformation system converts the image coordinate system into the robot arm's base coordinate system, driving the robot arm to move to the target position for clamping.

[0086] It also includes a pressure closed-loop control module, which contains a six-dimensional force sensor to collect contact pressure data in real time. When the deviation is greater than 20%, it triggers image reacquisition and clamping point correction.

[0087] In one operation, the multispectral imaging device switches filters at a 30ms cycle to acquire images at 560nm, 590nm visible light, and 810nm near-infrared light. The CMOS sensor outputs 1024×768 pixel data, which is transmitted to the image processing unit via gigabit Ethernet. The processing unit converts the image to the HSV color space, performs principal component analysis (H+0.8S, 2V-S), and constructs a two-dimensional feature space for young and old branches.

[0088] The color gradient analysis module calculates the hue and saturation differences between adjacent pixels along the branch axis. When the hue difference is ≥15 and the saturation difference is ≥20, it is determined to be a color transition boundary. The gripping point calculation module dynamically adjusts the offset based on the branch diameter to generate three-dimensional coordinate data. The coordinate transformation system maps the image coordinates to the robotic arm's base coordinate system through a hand-eye calibration matrix, driving the robotic arm to move to the target position.

[0089] A six-dimensional force sensor monitors the clamping pressure in real time. When the pressure deviation exceeds 20%, a vibration-compensated camera is triggered to re-acquire a local image (50×50mm area), and the clamping point coordinates are updated synchronously. Correction commands dynamically adjust the robotic arm's motion parameters through an incremental PID algorithm.

[0090] This system achieves accurate identification of the boundary between young and old mulberry branches through multispectral imaging and dynamic positioning calculations. A pressure closed-loop control module corrects clamping errors in real time, and combined with coordinate transformation and robotic arm motion control, it improves the stability and adaptability of clamping operations, reducing mechanical damage to the mulberry branches.

[0091] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Further modifications can be readily implemented by those skilled in the art.

Claims

1. A method for clamping and positioning mulberry branches, characterized in that, Includes the following steps: S1. Real-time image data of mulberry branches are acquired through a multispectral imaging device, which includes the acquisition functions of the 500-600 nm visible light band and the 750-900 nm near-infrared band. S2. Extract the HSV color space parameters of the branch area in the image and establish a color gradient model between the tender branches and the old branches. The HSV value range of the tender branches is defined as hue value H in the range of 80~160, saturation S in the range of 40~100, and lightness V in the range of 30~70. The HSV value range of the old branches is defined as hue value H in the range of 20 to 50, saturation S in the range of 10 to 40, and lightness V in the range of 10 to 30. S3. Calculate the color gradient change rate along the branch axis. When the hue value difference between adjacent pixels is greater than or equal to 15 and the saturation difference is greater than or equal to 20, it is determined to be a color transition boundary. S4. Using the color transition boundary offset 5 mm to 15 mm towards the tender branch as the clamping reference point, generate positioning data containing three-dimensional coordinates; S5. Convert the coordinates of the positioning point in the image coordinate system to the base coordinate system of the robotic arm, drive the gripping mechanism to move to the target position, and perform gripping.

2. The mulberry branch clamping and positioning method as described in claim 1, characterized in that, In step S1, the multispectral imaging device adopts a time-division multiplexing three-channel imaging mode, specifically including: The visible light channel includes a first filter with a center wavelength of 560±5 nm and a second filter with a center wavelength of 590±5 nm, while the near-infrared channel is equipped with a third filter with a center wavelength of 810±10 nm. The three filters switch in rotation with a period of 30 milliseconds, synchronously triggering the CMOS sensor to acquire 1024×768 pixel image data in each channel. The imaging resolution of the visible light channel reaches 0.1 mm / pixel, and the imaging resolution of the near-infrared channel reaches 0.15 mm / pixel. The multispectral imaging device is synchronized with the depth sensor, and point cloud data of the corresponding band is acquired synchronously each time the filter switches.

3. The mulberry branch clamping and positioning method as described in claim 1, characterized in that, The process of establishing the color gradient model in step S2 specifically includes: S21. Perform illumination compensation on the original HSV parameters and use an adaptive white balance algorithm based on the gray world assumption to adjust the average brightness value of the image to the range of 50-60 luminance. S22. Establish a dual-channel principal component analysis model, with linear combinations corresponding to the first principal component channel and the second principal component channel. Construct a two-dimensional feature space for young branches and old branches by calculating the principal component scores. S23. Dynamically adjust the HSV threshold boundary. When the ambient light intensity exceeds 80,000 lux, raise the lower limit of the hue value H of young branches to 100. When the ambient humidity exceeds 75%, lower the upper limit of the saturation S of old branches to 35.

4. The mulberry branch clamping and positioning method as described in claim 3, characterized in that, The execution process of the adaptive white balance algorithm in step S21 specifically includes: The acquired image is divided into 5×5 equal-area sub-regions, and the average brightness value of each sub-region is calculated. For sub-regions with average brightness values ​​exceeding the range of 40-70, a compensation weight of 0.1-0.9 is assigned according to their distance from the image center. Based on the weight values, the HSV parameters of each sub-region are non-linearly adjusted so that the average brightness of the overall image converges to the range of 52-58. When the rate of change of ambient light intensity exceeds 100 lux per second, the white balance calculation cycle is automatically shortened to 0.2 seconds, and the data is fused and calibrated with the light intensity sensor data of the multispectral imaging device, with the calibration error controlled within ±3 brightness units.

5. The mulberry branch clamping and positioning method as described in claim 1, characterized in that, The process of determining the color transition boundary in step S3 specifically includes: When calculating the color gradient change rate, the curvature change rate along the branch axis is analyzed simultaneously. When the hue value difference and saturation difference simultaneously meet the threshold conditions and the curvature change rate is less than 0.05 mm per millimeter, it is determined to be a valid color transition boundary. For mulberry branches with a diameter greater than 8 mm, the hue value difference threshold is dynamically increased to 18, and the near-infrared band reflectance difference is introduced as an auxiliary criterion. When the reflectance difference exceeds 15%, the three-dimensional point cloud data is resampled. If the hue value difference of adjacent pixels exceeds the threshold three times in a row but the curvature change rate is abnormal, the polarization mode of the multispectral imaging device is activated for secondary verification. The polarization angle is set to 45±5 degrees, and the color gradient value is corrected by the extinction ratio change.

6. The mulberry branch clamping and positioning method as described in claim 1, characterized in that, The process of determining the clamping reference point in step S4 specifically includes: The offset is dynamically adjusted based on the branch diameter. When the detected mulberry branch diameter is less than 5 mm, the offset towards the tender branch is set to 8-12 mm; when the diameter is between 5-10 mm, the offset is adjusted to 5-8 mm; and when the diameter is greater than 10 mm, the offset is increased to 12-15 mm. At the same time, the offset is dynamically compensated by combining the branch elastic modulus data measured in the near-infrared band. For every 100 MPa increase in elastic modulus, the offset is reduced by 1 mm. The final positioning data includes the three-dimensional coordinates of the clamping point, the recommended clamping pressure value, and the clamping angle correction value. The clamping angle correction value is calculated based on the branch curvature radius at that point to ensure that the angle deviation between the clamping surface and the axial tangent of the branch is less than 3 degrees.

7. The mulberry branch clamping and positioning method as described in claim 1, characterized in that, Also includes: S6. During the clamping process, the contact pressure data is collected in real time. When the deviation between the measured pressure value and the preset pressure value exceeds 20%, the image is reacquired and a closed-loop correction action is executed.

8. The mulberry branch clamping and positioning method as described in claim 7, characterized in that, The closed-loop correction process described in step S6 specifically includes: A pressure-image dual-modal feedback mechanism is established. When the contact pressure deviation exceeds a threshold, real-time vibration compensation images from a multispectral imaging device and three-dimensional torque data from a six-axis force sensor are simultaneously acquired. The pressure fluctuation spectrum characteristics are analyzed using a machine learning model to identify the source of the deviation as either branch slippage or foreign object interference. If it is branch slippage, image reacquisition with 0.5 mm precision is triggered; if it is foreign object interference, the clamping surface vibration cleaning mode is activated. During the correction process, an incremental PID control algorithm is used to dynamically adjust the proportional coefficient based on historical deviation data. For every 5% increase in pressure deviation, the proportional coefficient increases by 0.2, and the integral time constant decays exponentially. After the correction is completed, the preset pressure reference value is automatically updated. The new reference value is the moving average of the last 10 effective clamping pressures and is stored in the branch feature database to form a self-learning curve.

9. A mulberry branch clamping and positioning system, comprising the method described in any one of claims 1 to 8, characterized in that, include: The multispectral imaging device includes a switchable filter wheel, a CMOS sensor, and a lens, used to acquire images of mulberry branches in the 500-600 nm visible light band and the 750-900 nm near-infrared band. The image processing unit is communicatively connected to the multispectral imaging device and is configured to convert the image to the HSV color space and perform principal component analysis to establish a principal component channel containing a linear combination of hue value H and saturation value S. The color gradient analysis module is used to calculate the rate of change of hue and saturation gradients along the branch axis. When the difference in hue value between adjacent pixels is ≥15 and the difference in saturation value is ≥20, it is determined to be a color transition boundary. The clamping point calculation module generates three-dimensional coordinates of the clamping reference point by offsetting it 5-15 mm towards the tender branch based on the color transition boundary. The offset is dynamically adjusted according to the branch diameter. The coordinate transformation system converts the image coordinate system into the robot arm's base coordinate system, driving the robot arm to move to the target position for clamping.

10. The mulberry branch clamping and positioning system as described in claim 9, characterized in that, Also includes: The pressure closed-loop control module includes a six-dimensional force sensor that collects contact pressure data in real time. When the deviation is greater than 20%, it triggers image reacquisition and clamping point correction.

Images