Intelligent green belt self-adaptive precise pruning system

Abstract

The application relates to the technical field of image recognition, in particular to an intelligent green belt self-adaptive precise pruning system, which comprises an image time sequence acquisition module, a flexible micro-motion recognition module, a space contour aggregation module, a space contour aggregation module and a pruning boundary recognition module.In the application, continuous pictures are acquired to establish a fixed depth cache and consistency check, eliminate acquisition jitter, ensure stable space-time reference, capture flexible micro-motion by using gray scale symbol difference and switching count, eliminate rigid background interference, overcome light and shadow misjudgment through three-dimensional color vector direction calculation, realize discrete pixel space merging and path smoothing by combining grid density iteration and boundary constraint adjustment, avoid frequent mechanical correction, and ensure pruning precision.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to an intelligent adaptive precision pruning system for green belts. Background Technology

[0002] Image recognition technology is an interdisciplinary field encompassing computer vision, pattern recognition, artificial intelligence, and digital image processing. Its core lies in the digital capture, mathematical modeling, and semantic understanding of information from the physical world. The technical process mainly covers information acquisition, preprocessing, feature extraction and selection, classifier design, and classification decision-making. Specifically, it uses image sensors to convert light signals into electrical signals to acquire basic information; preprocessing operations such as denoising and transformation are used to enhance features; and then scale-invariant feature transformation, histogram of oriented gradients, or convolutional neural networks (CNNs) are employed for feature extraction and dimensionality reduction. CNNs, by simulating the biological visual cortex, utilize convolutional layers, pooling layers, and fully connected layers to achieve layer-by-layer perception of complex targets. Pattern recognition, combining probabilistic statistics and fuzzy mathematics, has evolved from initial text recognition to intelligent recognition of multidimensional objects.

[0003] The intelligent green belt adaptive precision pruning system, as an automated garden maintenance device, integrates omnidirectional motion, multi-degree-of-freedom cutting, and environmental perception modules. The system uses an STM32 microcontroller as its logic control core, coordinating all execution components. For motion, a Mecanum wheel drive enables omnidirectional movement and rotation of the machine body. For pruning operations, it features upper and side pruning systems, using screw drives, hydraulic drives, and servo motor control to flexibly adjust the blade height, pitch angle, and entry angle. For perception and monitoring, it utilizes lidar to detect obstacles, combines this with cameras to capture green belt image features, and sensors to monitor the status of the leaf collection box. Furthermore, the system is equipped with a negative pressure leaf recovery device, achieving integrated intelligent operation of green belt edge recognition, dynamic pruning, obstacle detection, and residue collection.

[0004] Existing technologies rely on a transient static acquisition mode using LiDAR combined with cameras. However, the laser beam easily penetrates sparse branches and leaves in green belts, resulting in discontinuous edge depth feedback and an inability to accurately construct the surface morphology of flexible vegetation. Traditional vision solutions tend to extract texture features from single-frame images, ignoring the temporal dynamic characteristics of vegetation swaying in the wind. This makes it difficult to distinguish between stationary green interference objects in the background and the actual vegetation to be pruned. Furthermore, they are highly susceptible to changes in ambient lighting and shadow occlusion, causing distortion in color feature extraction. The control logic directly drives the actuator based on the original recognition points, lacking spatial aggregation and path smoothing processing of discrete pixels. This leads to the pruning equipment frequently correcting the cutting angle during operation, resulting in jagged cut surfaces. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an intelligent green belt adaptive precision pruning system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent green belt adaptive precision pruning system comprising:

[0007] The image time sequence acquisition module acquires the image sequence obtained by the front-end camera component of the trimming device, writes it into the buffer in chronological order and performs consistency verification to establish a continuous frame pixel time sequence image set;

[0008] The flexible micro-motion recognition module calculates the gray-level difference between adjacent frames based on the continuous frame pixel time-series image set, constructs the oscillation discrimination criteria, and marks the pixel positions that meet the conditions through sign switching counting and amplitude equalization to obtain a flexible pixel candidate distribution map.

[0009] The vegetation color orientation discrimination module calls the flexible pixel candidate distribution map to obtain pixel color orientation information and performs consistency discrimination with the reference color orientation. The pixel coordinates that pass the screening are retained to obtain the vegetation effective pixel recognition map.

[0010] The spatial contour aggregation module obtains the spatial distribution information of pixels in the effective pixel recognition map of vegetation and performs density statistics and neighborhood aggregation processing to remove isolated pixels and generate a continuous contour pixel map of vegetation.

[0011] The trimming boundary recognition module determines the connectivity of contour pixels and adjusts the edge sequence based on the continuous contour pixel map of the vegetation, constructs an image boundary that can be recognized by the trimming device, and establishes an adaptive and precise trimming boundary image for the green belt.

[0012] As a further aspect of the present invention, the continuous frame pixel temporal image set includes a pixel matrix sequence, a frame index list, and a pixel coordinate mapping table;

[0013] The flexible pixel candidate distribution map specifically includes zero-crossing rate features, mean absolute difference, and channel index.

[0014] The effective pixel recognition map of vegetation specifically refers to RGB color vector, spectral angle, and vector projection modulus;

[0015] The continuous vegetation contour pixel map includes spatial point density, mean shift vector, and cluster centroid.

[0016] The adaptive and precise pruning boundary image of the green belt specifically includes boundary chain code, curvature features, and motion trajectory.

[0017] As a further aspect of the present invention, the pixel coordinates are data indexes that identify the two-dimensional positions of feature points in a matrix and are transformed and mapped to the physical operating positions of the equipment;

[0018] The pixel color orientation information is a data feature that is extracted from the red, green and blue channel values, converted into a three-dimensional vector, and compared with the vegetation reference vector to perform a spatial angle consistency comparison to eliminate the interference of light and shadow.

[0019] The outline pixels are effective points that are grouped into continuous vegetation morphology after spatial density statistics and neighborhood aggregation processing.

[0020] As a further aspect of the present invention, the isolated pixel is an invalid point that is regarded as noise and removed during the aggregation process because it is not merged or the number of points in its region is scarce;

[0021] The adjusted edge sequence is to calculate the abrupt change position of the vector angle of the connected linked list, and to perform smoothing correction and sequential recombination to optimize the pruning trajectory based on mechanical motion constraints;

[0022] The image boundary is a visual boundary line that is converted into device control commands based on a continuous contour map after connectivity analysis and path optimization.

[0023] As a further aspect of the present invention, the image timing acquisition module includes:

[0024] The data acquisition and conversion submodule acquires images continuously obtained by the front-end camera component of the pruning equipment in the green belt operation scenario, obtains the pixel matrix, frame order index and pixel two-dimensional coordinate information corresponding to each frame, and establishes a frame-level pixel coordinate matrix.

[0025] The cache verification and analysis submodule calls the frame-level pixel coordinate matrix to write the continuous frame images into a fixed-depth cache area in chronological order. After detecting that the number of frames in the cache area reaches the preset frame depth condition, it calculates the position difference between the pixel position of each frame in the cache area and the position reference benchmark. It then performs consistency verification in combination with the preset deviation tolerance range to generate a pixel-by-pixel temporal consistency verification coefficient.

[0026] The temporal sequence construction submodule, based on the pixel-by-pixel temporal consistency verification coefficient, filters frame data that meets the consistency requirements, performs temporal reorganization on the pixel positions of each frame in the buffer, forms an image sequence structure that can be used for pixel-by-pixel temporal analysis, and obtains a set of continuous frame pixel temporal images.

[0027] As a further aspect of the present invention, the flexible micro-motion recognition module includes:

[0028] The symbol sequence statistics submodule calls the continuous frame pixel time-series image set, calculates the gray-level difference value at the same coordinate position in adjacent frames, constructs a symbol change sequence based on the positive and negative polarity of the difference value, counts the cumulative number of positive and negative state switching in the sequence, and obtains the pixel oscillation frequency statistics value.

[0029] The amplitude averaging operation submodule calls the continuous frame pixel time-series image set, extracts the gray-level difference values ​​between all adjacent frames in the sequence, calculates the sum of the absolute values ​​of the difference values, calculates the mean of the sum of absolute values ​​based on the time span of the sequence, and generates the gray-level micro-motion amplitude mean.

[0030] The dual threshold mapping submodule compares the frequency values ​​and amplitude averages of the pixel oscillation frequency with the preset frequency lower limit and amplitude constraint parameters, respectively, based on the statistical values ​​of the pixel oscillation frequency and the average amplitude of the gray-scale micro-motion. It then filters the pixel positions and channel information that simultaneously meet the dual threshold conditions and establishes a flexible pixel candidate distribution map.

[0031] As a further aspect of the present invention, the vegetation color orientation discrimination module includes:

[0032] The three-dimensional vector construction submodule detects the red, green and blue channel values ​​of the marked pixels in the flexible pixel candidate distribution map, constructs a three-dimensional color direction vector at the pixel end, obtains preset vegetation reference color direction data and constructs a reference vector, and generates a three-dimensional color direction reference vector set.

[0033] The consistency operation submodule calls the three-dimensional color direction reference vector set, calculates the ratio of the dot product value between the three-dimensional color direction vector at the pixel end and the reference reference vector to the product of the magnitudes of the two vectors, and obtains the color direction consistency discrimination coefficient;

[0034] The illumination projection verification submodule filters pixels whose coefficients meet the consistency threshold requirements based on the color direction consistency discrimination coefficient, calculates the projection length of the pixel vector in the reference vector direction, compares the projection length with the illumination intensity interval threshold, retains the pixel coordinates that meet the dual verification conditions, and establishes a vegetation effective pixel recognition map.

[0035] As a further aspect of the present invention, the spatial contour aggregation module includes:

[0036] The grid density mapping submodule obtains preset image space grid division parameters for the two-dimensional coordinate data of each pixel in the effective pixel recognition image of the vegetation, maps the pixel coordinates to the corresponding spatial grid unit system, counts the number of pixels contained in each grid unit, and generates a grid pixel density distribution table.

[0037] The density gravity displacement submodule calls the grid pixel density distribution table, identifies the current grid and its neighboring grid cells, calculates the density difference between the neighboring grid and the central grid, determines the aggregation vector for pixels to move toward the high-density region by combining the spatial distance span between grids, iteratively updates the pixel coordinates within the cell based on the aggregation vector, and generates a set of pixel aggregation displacement coordinates.

[0038] The connected region merging submodule calculates the updated Euclidean space distance between pixels based on the pixel aggregation displacement coordinate set, determines whether the distance is less than a preset similar merging threshold, merges pixels that meet the condition into connected regions, removes isolated pixels that have not been aggregated, and establishes a continuous vegetation contour pixel map.

[0039] As a further aspect of the present invention, the trimming boundary recognition module includes:

[0040] The edge sequence construction submodule, based on the continuous contour pixel map of the vegetation, traverses the pixel coordinate data in the image plane, calculates the Euclidean distance between adjacent pixels, determines spatial connectivity, sorts the pixels according to the connectivity relationship, constructs a contour linked list structure with the head and tail connected, and obtains the edge pixel connectivity sequence.

[0041] The mutation constraint optimization submodule calls the edge pixel connected sequence, calculates the vector direction angle between adjacent line segments in the sequence path, identifies the location of the angle mutation, and combines the mechanical motion constraint parameters of the pruning device to perform smoothing correction and order rearrangement on the path mutation points to generate an optimized pruning path coordinate set.

[0042] The physical command conversion submodule, based on the optimized pruning path coordinate set, maps the pixel coordinates to the physical operation coordinate system of the pruning equipment through coordinate transformation and scaling, and converts them into boundary command data that matches the equipment control protocol, thereby establishing an adaptive and precise pruning boundary image for the green belt.

[0043] As a further aspect of the present invention, the process of identifying the location of the abrupt change in angle specifically involves extracting the coordinates of three consecutive pixels from the connected sequence of edge pixels and calculating the change in angle between the first vector formed by the first two pixels and the second vector formed by the last two pixels.

[0044] The process of combining the mechanical motion constraint parameters of the pruning equipment specifically involves obtaining the maximum angular acceleration value and the working linear velocity value of the servo motor of the pruning robot arm, calculating the ratio of the maximum angular acceleration value to the working linear velocity value to obtain the maximum single turning angle value, and rounding it down to set as the maximum allowable deflection angle threshold.

[0045] The value of the angle change is compared with the maximum allowable deflection angle threshold. When the value of the angle change is greater than the maximum allowable deflection angle threshold, the current position is determined to be a path change point. The process of performing smoothing correction and order rearrangement on the path change point is specifically to calculate the midpoint coordinates of the last two pixels in the coordinates of the three consecutive pixels.

[0046] The midpoint coordinates are inserted into the corresponding position of the edge pixel connected sequence, and the original angle change value is decomposed into two angle values ​​that are smaller than the maximum allowable deflection angle threshold.

[0047] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0048] In this invention, a fixed depth buffer and position consistency verification are established by acquiring continuous images to eliminate image jitter caused by camera component movement, ensuring data source stability. By utilizing grayscale sign difference and positive / negative switching count, the flexible micro-motion characteristics of vegetation are accurately captured, effectively distinguishing between rigid backgrounds and dynamic branches and leaves. This allows for the construction of a three-dimensional color vector and a reference vector for directional consistency calculation, transforming absolute color judgment into spatial angle discrimination, overcoming misjudgments caused by uneven lighting and shadow occlusion. Based on grid density comparison and iterative displacement, discrete leaf pixel spatial merging and noise removal are achieved, constructing a continuous and complete vegetation outline. Constraints are adjusted for positions with abrupt directional changes to generate smooth pruning paths, reducing ineffective adjustments and mechanical wear of the actuator. Attached Figure Description

[0049] Figure 1 This is a system flowchart of the present invention;

[0050] Figure 2 This is a schematic diagram comparing the dynamic decay model of the present invention with the prior art;

[0051] Figure 3 This is a schematic diagram of the flexible pixel screening distribution of the present invention;

[0052] Figure 4 This is a schematic diagram of pixel aggregation displacement and connected region merging in this invention;

[0053] Figure 5 This is a graph comparing the recognition accuracy of the system of the present invention with that of existing systems under different wind speed conditions. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0055] Please see Figures 1-5 An intelligent green belt adaptive precision pruning system includes:

[0056] The image time sequence acquisition module acquires images continuously obtained by the front-end camera component of the pruning equipment in the green belt operation scenario. It obtains the pixel matrix, frame order index and pixel two-dimensional coordinate information corresponding to each frame. The continuous frame images are written into a fixed-depth buffer in time order. After the number of frames in the buffer reaches the preset frame depth condition, the consistency of the pixel position of each frame in the buffer is checked and an image sequence structure that can be used for pixel-by-pixel time sequence analysis is formed, and a continuous frame pixel time sequence image set is established.

[0057] The flexible micro-motion recognition module calls the continuous frame pixel time sequence image set, calculates the gray level symbol difference of the same coordinate in adjacent frames to construct a symbol sequence, counts the number of positive and negative switching as the basis for oscillation, accumulates gray level difference and averages it simultaneously, compares the switching count and amplitude mean with the frequency lower limit and amplitude constraint conditions respectively, marks the pixels that simultaneously meet the dual thresholds, records their position and channel information, and obtains the flexible pixel candidate distribution map.

[0058] The vegetation color orientation discrimination module obtains the values ​​of the red, green and blue channels corresponding to the marked pixels based on the flexible pixel candidate distribution map and constructs a three-dimensional color vector. At the same time, it obtains the vegetation reference color orientation and constructs a reference vector. It calculates the vector direction consistency to complete the angle discrimination. For the pixels that pass the angle discrimination, it calculates the color orientation projection length and verifies the illumination interval constraint. It retains the pixel coordinates that pass the double verification and obtains the effective vegetation pixel recognition map.

[0059] The spatial contour aggregation module identifies pixel coordinate information in the effective pixel recognition image of vegetation, obtains image spatial grid division parameters and completes pixel spatial mapping, counts pixel density within grid cells, compares neighborhood density to determine aggregation direction, iterates pixel coordinates within displacement cells according to aggregation direction, judges spatial distance to complete merging of similar pixels, removes isolated pixels that have not been aggregated, and generates a continuous contour pixel map of vegetation.

[0060] The trimming boundary recognition module is based on the continuous contour pixel map of vegetation. It obtains the spatial distribution order of contour pixels in the image plane, judges the connectivity of adjacent pixels and constructs an edge sequence. It performs constraint adjustment and order rearrangement for positions with abrupt changes in direction, maps the optimized sequence into a boundary expression that can be recognized by the device, and establishes an adaptive and accurate trimming boundary image of the green belt.

[0061] A continuous frame pixel temporal image set includes a pixel matrix sequence, a frame index list, and a pixel coordinate mapping table;

[0062] The flexible pixel candidate distribution map specifically includes zero-crossing rate features, mean absolute difference, and channel index.

[0063] The effective pixel identification image of vegetation specifically refers to the RGB color vector, spectral angle, and vector projection modulus.

[0064] The continuous vegetation outline pixel map includes spatial point density, mean shift vector, and cluster centroids;

[0065] The adaptive and precise pruning boundary image of the green belt specifically includes boundary chain code, curvature features, and motion trajectory.

[0066] Please see Figures 1-2 , Figure 5 The image timing acquisition module includes:

[0067] The data acquisition and conversion submodule acquires images continuously obtained by the front-end camera component of the pruning equipment in the green belt operation scenario, obtains the pixel matrix, frame order index and pixel two-dimensional coordinate information corresponding to each frame, and establishes a frame-level pixel coordinate matrix.

[0068] Images continuously acquired by the front-end camera component of the pruning equipment in the green belt operation scenario are collected. Based on the physical characteristic that the average width of the leaves of the green belt plants is 5 mm, and combined with the Nyquist sampling theorem requiring at least 2.5 pixels to cover the minimum feature, the effective field of view in the horizontal direction is set to 4800 mm. The horizontal resolution parameter is calculated as 4800 divided by 2.5, which gives 1920 pixels. The vertical direction is set to 1080 pixels according to the 16:9 aspect ratio. For the 120th frame of image data in the pruning operation, the horizontal coordinates from 0 to 1919 and the vertical coordinates from 0 to 1079 are traversed. The coordinates are located at pixel coordinates (960, 540). The brightness values ​​of this point in the red, green, and blue channels are read as 145, 160, and 130, respectively. The system clock is called to obtain the timestamp of the current sampling time, 16890023, as the basis for generating the frame sequence index. A feature vector containing five data items is constructed. The above numerical extraction and vector construction operations are performed on all 2,073,600 pixels of the image. The feature vectors of all pixels are filled into a two-dimensional array structure in row-major order. The pixel matrix, frame sequence index, and pixel two-dimensional coordinate information corresponding to each frame are obtained, and a frame-level pixel coordinate matrix is ​​established.

[0069] The cache verification and analysis submodule calls the frame-level pixel coordinate matrix to write consecutive frame images into a fixed-depth cache area in chronological order. After the number of frames in the cache area reaches the preset frame depth condition, it calculates the position difference between the pixel position of each frame in the cache area and the position reference benchmark. It then performs consistency verification in combination with the preset deviation tolerance range and generates a pixel-by-pixel temporal consistency verification coefficient.

[0070] The frame-level pixel coordinate matrix is ​​invoked. Based on the vibration test data of 100 previous pruning operations, the average pixel drift caused by equipment vibration is statistically analyzed to be 3.2 pixels. 3.2 plus twice the standard deviation of 0.4 is calculated to get 4.0. Based on this, the deviation tolerance range is set as a circular area with a radius of 4 pixels. At the same time, the maximum drift is analyzed to be 9 pixels. The linear penalty interval for exceeding the tolerance range is set to 5 pixels. 1.0 divided by 5 is calculated to get 0.2. This is used as the preset deviation attenuation step size. The continuous frame images are written to a fixed-depth buffer in chronological order. After the number of frames in the buffer reaches the preset frame depth condition, the feature point at coordinates (500, 600) in the reference frame is selected. The position difference value between the feature point and the corresponding point (503, 604) in the subsequent frame is calculated to be 5. The comparison shows that 5 is greater than 4, and it is determined that it exceeds the tolerance range. For the 1 pixel unit of the excess part, the operation of 1.0 minus 1 multiplied by 0.2 is performed. The consistency coefficient of the point is calculated to be 0.8. Combined with the preset deviation tolerance range, consistency verification is performed to generate a pixel-by-pixel temporal consistency verification coefficient.

[0071] The temporal sequence construction submodule uses pixel-by-pixel temporal consistency verification coefficients to filter frame data that meet consistency requirements, and performs temporal reorganization on the pixel positions of each frame in the buffer to form an image sequence structure that can be used for pixel-by-pixel temporal analysis, thus obtaining a set of continuous frame pixel temporal images.

[0072] Based on the pixel-by-pixel temporal consistency verification coefficient, the average consistency coefficient of effective vegetation pixels under wind conditions is 0.92, and the average coefficient of background noise pixels is 0.70. The difference between the two is calculated to be 0.22. Two-thirds of the difference is taken as the dividing point. 0.70 is added to 0.22 and multiplied by 0.66 to get 0.845. The consistency screening threshold is set to 0.85. The verification coefficient matrix of 10 frames of data in the buffer is traversed. For the pixel time series at coordinates (500, 600), the coefficient 0.8 calculated in frame 305 is read. The numerical comparison shows that 0.8 is less than 0.85. The data of this pixel in frame 305 is marked as invalid and removed. The coefficient 0.9 in frame 306 is read. The numerical comparison shows that 0.9 is greater than 0.85. The data of this pixel in frame 306 is retained. The remaining effective frame data after screening is statistically analyzed. The pixel positions of each frame in the buffer are temporally reassembled to form an image sequence structure that can be used for pixel-by-pixel temporal analysis, resulting in a continuous frame pixel temporal image set.

[0073] Please see Figure 1 , Figure 3 , Figure 5 The flexible micro-motion recognition module includes:

[0074] The symbol sequence statistics submodule calls the continuous frame pixel time-series image set, calculates the gray-level difference value at the same coordinate position in adjacent frames, constructs a symbol change sequence based on the positive and negative polarity of the difference value, counts the cumulative number of positive and negative state switching in the sequence, and obtains the pixel oscillation frequency statistics value.

[0075] The system calls a continuous frame pixel time-series image set, selects the pixel at coordinates (500, 600) in the image matrix, and reads the grayscale value sequence of this pixel in 8 consecutive frames of time-series images [120, 125, 122, 128, 126, 130, 120, 125]. It then performs a difference operation between adjacent frames, subtracting the preceding term from the following term, to calculate a sequence containing 7 differences [5, -3, 6, -2, 4, -10, 5]. The system assigns a symbol 1 to differences greater than 0 and a symbol 0 to differences less than or equal to 0, converting the numerical sequence into a binary symbol sequence [1, 0, 1, 0, 1, 0, 1]. The initial state switching counter is set to 0. Starting from the first position of the sequence, adjacent symbol values ​​are compared bit by bit. If the first bit 1 is different from the second bit 0, the counter is incremented by 1. If the second bit 0 is different from the third bit 1, the counter is incremented by 1. A total of 6 symbol flips are detected in the entire sequence. The statistical value 6 is temporarily stored as the oscillation frequency of the coordinate point. The above difference and flip statistical operations are performed on all coordinate points in the image set. The gray level difference value of the same coordinate position in adjacent frames is calculated. The symbol change sequence is constructed based on the positive and negative polarity of the difference value. The cumulative number of positive and negative state switching in the sequence is counted to obtain the pixel oscillation frequency statistical value.

[0076] The amplitude averaging operation submodule calls the continuous frame pixel time-series image set, extracts the gray-level difference values ​​between all adjacent frames in the sequence, calculates the sum of the absolute values ​​of the difference values, calculates the mean of the sum of absolute values ​​based on the time span of the sequence, and generates the gray-level micro-motion amplitude mean.

[0077] Call the continuous frame pixel time-series image set, and continue to target the difference value sequence [5,-3,6,-2,4,-10,5] at coordinates (500,600). Call the absolute value operation function to convert the negative values ​​in the sequence into positive values, obtain the absolute difference sequence [5,3,6,2,4,10,5], construct the cumulative summation formula to calculate the absolute sum of 35, count the number of inter-frame intervals involved in the calculation as 7, use it as the sequence time span parameter, perform the numerical division operation 35 divided by 7, calculate the quotient value 5, for another coordinate point (510,600) affected by strong wind, calculate its absolute difference sum as 140, divide by the span 7 to get the mean 20, map the calculated values ​​5 and 20 to the storage units of the corresponding coordinates respectively, extract the gray level difference values ​​between all adjacent frames in the sequence, calculate the absolute sum of the difference values, calculate the mean of the absolute sum of the difference values ​​based on the sequence time span, and generate the mean of gray level micro-motion amplitude;

[0078] The dual threshold mapping submodule compares the frequency values ​​and the average grayscale micro-motion amplitude with the pixel oscillation frequency statistics and the average amplitude values ​​against the preset frequency lower limit and amplitude constraint parameters to filter the pixel positions and channel information that simultaneously meet the dual threshold conditions and establish a flexible pixel candidate distribution map.

[0079] Based on the statistical values ​​of pixel oscillation frequency and the average amplitude of grayscale micro-motion, and considering the minimum flutter frequency of 3.5 Hz of flexible branches and leaves under natural wind measured in wind tunnel experiments, the lower limit parameter of frequency is set to 4 Hz to account for measurement error. Simultaneously, the range of grayscale changes caused by leaf micro-motion is measured, excluding sensor noise floor of 1.5 and grayscale jumps of 9.0 caused by large branch swings. The effective range of the amplitude constraint parameter is set to [2.0, 8.0]. The frequency value 6 and the average amplitude 5 at coordinates (500, 600) are read. A logical judgment operation is performed to confirm that 6 is greater than 4 and 5 is within the range of [2.0, 8.0]. Within the interval [2.0, 8.0], the point is determined to meet the dual conditions. The coordinates (500, 600) of the point and the corresponding green channel data are extracted and written into the candidate list. The frequency value 2 and the amplitude mean 20 at the coordinates (510, 600) are read. It is confirmed that 2 is less than 4 and 20 exceeds the upper limit of the amplitude interval. The point is determined to not meet the flexible micro-motion characteristics and is discarded. The frequency value and amplitude mean are compared with the preset lower limit of frequency and amplitude constraint parameters in both directions. The pixel positions and channel information that meet the dual threshold conditions are screened to establish a flexible pixel candidate distribution map.

[0080] Please see Figure 1 , Figure 5 The vegetation color orientation discrimination module includes:

[0081] The 3D vector construction submodule detects the red, green and blue channel values ​​of marked pixels in the flexible pixel candidate distribution map, constructs a 3D color orientation vector at the pixel end, obtains preset vegetation reference color orientation data and constructs a reference vector, and generates a 3D color orientation reference vector set.

[0082] The red, green, and blue channel values ​​of the marked pixels in the flexible pixel candidate distribution map are detected. For the pixel to be identified at coordinates (500, 600) in the distribution map, the red channel value of 45, the green channel value of 155, and the blue channel value of 65 in its storage unit are read. The pixel-end feature vector [45, 155, 65] is constructed according to the color channel order. Based on the target object of the pruning operation, Buxus macrocarpa, 50 sets of healthy leaf sample images are collected under a standard diffuse light source environment. The red, green, and blue channel values ​​of the 50 sets of samples are summed and divided by 50 to calculate the average color values ​​of red 30, green 160, and blue 40. This is used as the standard reference data to construct the baseline reference vector [30, 160, 40]. At the same time, considering the vegetation characteristics under the shade environment, the average values ​​of red 20, green 80, and blue 25 of the leaves on the backlight side are collected to construct the auxiliary reference vector. The above two vectors are encapsulated and stored in the read-only memory. The preset vegetation reference color direction data is obtained and the baseline reference vector is constructed to generate a three-dimensional color direction baseline vector set.

[0083] The consistency operation submodule calls the three-dimensional color orientation reference vector set, calculates the ratio of the dot product between the pixel-end three-dimensional color orientation vector and the reference reference vector to the product of the magnitudes of the two vectors, and obtains the color orientation consistency discrimination coefficient.

[0084] The three-dimensional color orientation reference vector set is called. For the pixel vector [45,155,65] and the reference main vector [30,160,40], the component multiplication and accumulation operation is performed respectively to calculate the dot product value of 28750. The sum of squares of each component of the pixel vector and the square root are calculated to obtain the pixel vector magnitude of approximately 174.0. Similarly, the magnitude of the reference vector is calculated to be approximately 167.6. The multiplication operation 174.0 multiplied by 167.6 is performed to obtain the magnitude product 29162.4. The division operation 28750 divided by 29162.4 is performed to obtain the ratio 0.986. The above dot product and magnitude operation process is repeated for the auxiliary reference vector. The maximum value of the two calculation results is taken as the final judgment criterion. The ratio of the dot product value between the three-dimensional color orientation vector at the pixel end and the reference reference vector to the product of the magnitudes of the two vectors is calculated to obtain the color orientation consistency discrimination coefficient.

[0085] The illumination projection verification submodule uses the color direction consistency discrimination coefficient to filter pixels whose coefficients meet the consistency threshold requirements, calculates the projection length of the pixel vector in the reference vector direction, compares the projection length with the illumination intensity interval threshold, retains the pixel coordinates that meet the dual verification conditions, and establishes a vegetation effective pixel recognition map.

[0086] Based on the color orientation consistency discrimination coefficient, and according to the color orientation similarity test data between non-vegetation interference items (such as dead branches and soil) and the reference vector, the maximum similarity coefficient of the interference items was measured to be 0.82. To ensure discriminability, a consistency screening threshold of 0.90 was set. The calculated coefficients 0.986 and 0.90 were compared to determine their magnitudes. It was confirmed that 0.986 was greater than 0.90, and the color orientation of the pixel was determined to meet the requirements. The dot product value of 28750 calculated in the previous steps and the reference vector magnitude of 167.6 were then used to perform a division operation, 28750 divided by 1. 67.6, the calculated projection length is 171.5. Based on the histogram statistics of light intensity under daytime working conditions, the gray value 50 corresponding to the cumulative probability of 5% and the gray value 220 corresponding to 95% are extracted. The pixel projection interval corresponding to the effective light intensity is set as [50, 220]. The value 171.5 is checked and found to be within this closed interval. It is determined that the point meets the light effectiveness requirement. The pixel coordinates (500, 600) are recorded in the final output list. The pixel coordinates that meet the double verification conditions are retained to establish a vegetation effective pixel recognition map.

[0087] Please see Figure 1 , Figure 4 , Figure 5 The spatial contour aggregation module includes:

[0088] The grid density mapping submodule identifies the two-dimensional coordinate data of each pixel in the effective pixel recognition image of vegetation, obtains the preset image space grid division parameters, maps the pixel coordinates to the corresponding space grid unit system, counts the number of pixels contained in each grid unit, and generates a grid pixel density distribution table.

[0089] Based on the two-dimensional coordinate data of each pixel in the effective pixel identification image of vegetation, and considering that the average pixel area occupied by a typical leaf in the image at a resolution of 1920x1080 is approximately 320 square pixels, the square root is calculated to give its side length as approximately 18 pixels. To ensure that a single grid can fully accommodate the main leaf features, the grid size parameter for image space division is set to 20x20 pixels by rounding up. A grid index coordinate system with 96 horizontal and 54 vertical units is constructed. For the effective pixel with coordinates (500, 600) in the identification image, the rounding operation is performed: 500 divided by 20 to obtain the horizontal index 25, and 600 divided by 20 to obtain the vertical index 30. These are mapped to the grid unit numbered (25, 30). All effective pixels in the entire image are traversed. The number of pixels contained in the grid (25, 30) is counted to be 35, while the number of pixels in the adjacent grid (25, 31) is 5. A grid pixel density distribution table is generated.

[0090] The density gravity displacement submodule calls the grid pixel density distribution table, identifies the current grid and its neighboring grid cells, calculates the density difference between the neighboring grid and the central grid, determines the aggregation vector for pixels to move toward the high-density region by combining the spatial distance span between grids, iteratively updates the pixel coordinates within the cell based on the aggregation vector, and generates a set of pixel aggregation displacement coordinates.

[0091] The density-gravitational displacement submodule calls the grid pixel density distribution table to identify the current grid and its neighboring grid cells. To quantify the physical trend of pixels aggregating towards high-density regions, the following pixel position update formula is constructed:

[0092] ,

[0093] in, Indicates the first The pixel to be updated in the next iteration Two-dimensional coordinate vector ;

[0094] Represents the updated coordinate vector; This is the gravity aggregation step size coefficient, used to control the displacement amplitude of a single iteration and prevent oscillations;

[0095] Represents the neighborhood set All grids inside The resulting gravitational components are vector summed.

[0096] and Representing the neighborhood grid respectively With the center grid of the current pixel The density value;

[0097] Aimed at constructing nonlinear gravitational mass, square operation This is to amplify the attraction / repulsion weights in high-density difference regions. The sign function, with a positive difference taking 1 and a negative difference taking -1, determines the direction of the force. When the neighborhood density... Above the center When the force is at a certain angle, it manifests as attraction; conversely, when the force is at a certain angle, it manifests

[0098] Fraction The normalized direction vector divided by the distance follows the distance decay law in field theory physics, meaning that the closer the grid distance, the more significant the impact on the pixel. This represents the Euclidean norm, and the square of the denominator represents the effect that decreases inversely with the square of the distance.

[0099] Grid center coordinates Acquisition: Based on the previously set image resolution of 1920×1080 and grid size of 20×20 pixels, the geometric center coordinates corresponding to grid index (u,v) are calculated as (u×20+10,v×20+10).

[0100] Gravitational aggregation step size coefficient Settings: To prevent pixels from jumping out of the current grid range in a single iteration (i.e., the displacement must be less than 10 pixels), a typical maximum density difference is set based on the local texture features of the image. That is, 2.5 times the standard deviation, minimum center distance At this point, the maximum single gravitational force is 502 / 20=125. If the maximum single displacement is expected to be limited to 2 to 3 pixels to maintain topological stability, then η≈3 / 125=0.024. Considering the multi-neighborhood superposition effect and the convergence speed requirement, after 50 iterations on the validation set, η is finally set to 0.18.

[0101] Based on the aforementioned steps, the coordinates are selected as follows: The pixel is located within the center grid c at index (25, 30);

[0102] Calculate the center grid parameters:

[0103] center coordinates ;

[0104] Central density The value is set to 35 according to the steps outlined above.

[0105] To calculate the resultant force, the 4-neighborhood (up, down, left, right) of this grid is selected as the set. The calculation is performed, where the directly below neighborhood is the index (25, 31) explicitly mentioned in the previous steps, and the density of the remaining neighborhoods is sampled based on the local image texture continuity, with specific parameters shown in Table 1:

[0106] Table 1 Neighborhood Grid Parameter Monitoring Table

[0107] Neighborhood location Grid index ( ) Geometric center coordinates density value Relative position features Directly above (N) (25,29) (510,590) 0 Extremely low density (strong repulsive force) Directly below (S) (25,31) (510,630) 5 Low density (weak repulsive force) Left (W) (24,30) (490,610) 20 Medium density (medium repulsion) Directly to the right (E) (26,30) (530,610) 35 Isodensity (no force applied)

[0108] Referring to Table 1, the following is a step-by-step calculation of the gravitational components generated by each neighborhood on the center pixel:

[0109] For the directly above neighboring area ( ):

[0110] Density difference term: ;

[0111] ;

[0112] The squared difference term = 1225;

[0113] Distance vector term: ;

[0114] Squared distance norm: ;

[0115] Component calculation: ;

[0116] Note: The density at the top is extremely low, generating a downward thrust;

[0117] For the directly below neighborhood (k=S):

[0118] Density difference term: ;

[0119] ;

[0120] The squared difference term is 900;

[0121] Distance vector term: ;

[0122] The distance-square norm is 400;

[0123] Component calculation: ;

[0124] Note: The lower density generates an upward thrust, which partially offsets the repulsive force from above;

[0125] For the leftmost neighborhood (k=W):

[0126] Density difference term: ;

[0127] ;

[0128] The squared difference term is 225;

[0129] Distance vector term: ;

[0130] The squared distance norm is 400.

[0131] Component calculation: ;

[0132] Note: The lower density on the left generates a thrust to the right;

[0133] For the right-hand neighbor (k=E):

[0134] Density difference term: ;

[0135] The component is (0,0);

[0136] Summary and Update:

[0137] Perform vector summation on the above components:

[0138] ;

[0139] ;

[0140] Introducing step size coefficient Substitute into the formula to calculate:

[0141] ;

[0142] The calculation results show that when the surrounding density distribution is uneven, that is, the density is high at the top and low at the bottom and left, the pixels drift significantly to the relatively denser area to the lower right. This is consistent with the physical expectation that the pixels are attracted to the center of the denser vegetation area and away from the sparse edge area.

[0143] The advantage of the formula is that by introducing the square term of the density difference and the inverse distance term to construct a gravitational field model, discrete and fragmented edge pixels can automatically converge to the high-density core region with better connectivity, effectively filling the visual holes caused by wind blowing leaves, and providing a high-quality topological foundation for the subsequent generation of continuous and smooth trimmed boundaries.

[0144] The connected region merging submodule calculates the updated Euclidean space distance between pixels based on the pixel aggregation displacement coordinate set, determines whether the distance is less than the preset similar merging threshold, merges the pixels that meet the condition into connected regions, removes isolated pixels that have not been aggregated, and establishes a continuous vegetation contour pixel map.

[0145] Based on the pixel aggregation displacement coordinate set, and considering the average physical spacing of leaf lobes of 2 mm, the pixel distance is converted to approximately 4.8 pixels. Rounding up, the Euclidean distance threshold for merging similar pixels is set to 5.0 pixels. Pixels A (502, 603) and B (504, 605) after aggregation and updating are selected. Using the Pythagorean theorem, their spatial distance is calculated to be approximately 2.828. The value 2.828 is compared with the threshold 5.0. Since 2.828 is less than 5.0, the two points are determined to belong to the same connected region and assigned the same region label ID. For another pixel C (515, 620), its distance to point A is calculated to be 21.4, which is greater than the threshold 5.0, so it is not merged. The number of pixels under each connected region label is counted. Based on the number of pixels corresponding to the minimum effective leaf area, the isolated noise removal threshold is set to 10. If a region ID contains only 3 pixels, it is determined to be an invalid isolated point and removed from the list. Regions containing more than 10 pixels are retained, and a continuous vegetation contour pixel map is established.

[0146] Please see Figure 1 , Figure 5 The trimming boundary recognition module includes:

[0147] The edge sequence construction submodule is based on the continuous contour pixel map of vegetation. It traverses the pixel coordinate data in the image plane, calculates the Euclidean distance between adjacent pixels, determines the spatial connectivity, sorts the pixels according to the connectivity relationship, constructs the contour linked list structure with the head and tail connected, and obtains the edge pixel connectivity sequence.

[0148] Based on the continuous contour pixel map of vegetation, the entire image is scanned to locate the starting edge point coordinates (500, 600). According to the image rasterization principle, the maximum geometric distance between adjacent pixels in the diagonal direction is √2, which is approximately 1.414 pixels. To ensure the connectivity coverage of the 8-neighborhood search, the neighborhood search radius is set to 1.5 pixels. The existence of pixels in the eight-neighborhood of this point is checked. It is found that there are valid pixels at coordinates (501, 601). The horizontal difference 1 and the vertical difference 1 between the two points are calculated, and the Euclidean distance is approximately 1.414. It is confirmed that 1.414 is less than 1.5, so the two points are spatially connected. (501, 601) is linked to the linked list structure as the next node. The search continues with (501, 601) as the center for the subsequent point (502, 602). The pixels that meet the connectivity criteria are arranged clockwise according to their spatial distribution position. The pixels are spatially distributed and sorted according to the connectivity relationship. A contour linked list structure with the head and tail connected is constructed to obtain the edge pixel connectivity sequence.

[0149] The mutation constraint optimization submodule calls the edge pixel connectivity sequence, calculates the vector direction angle between adjacent line segments in the sequence path, identifies the location of the angle mutation, and combines the mechanical motion constraint parameters of the pruning equipment to perform smoothing correction and order rearrangement on the path mutation points to generate an optimized pruning path coordinate set.

[0150] The edge pixel connectivity sequence is called, and three consecutive points A(500,600), B(505,605), and C(505,610) are extracted from the sequence. The path angle change value is calculated to be 45 degrees. Based on the ratio of the maximum angular acceleration of the servo motor of the trimming robot arm to the working line speed, the maximum single turning angle to maintain trajectory smoothness is calculated to be 28.5 degrees. The maximum allowable deflection angle threshold is set to 30 degrees. The numerical comparison shows that 45 degrees is greater than 30 degrees, and it is determined that there is a path change at this point. The midpoint coordinates (505,607.5) of points B and C are calculated and inserted into the sequence. The original 45-degree turning angle is decomposed into two smaller turning angles. Combined with the mechanical motion constraint parameters of the trimming equipment, smoothing correction and order rearrangement are performed on the path change point to generate an optimized trimming path coordinate set.

[0151] The physical command conversion submodule, based on the optimized pruning path coordinate set, maps pixel coordinates to the physical operation coordinate system of the pruning equipment through coordinate transformation and scaling, and converts them into boundary command data that matches the equipment control protocol, thus establishing an adaptive and precise pruning boundary image for the green belt.

[0152] Based on the optimized pruning path coordinate set, the relative offsets between the origin of the image coordinate system and the origin of the robotic arm base coordinate system are read: 150 mm on the X-axis and 200 mm on the Y-axis. According to the physical relationship of 1920 horizontal pixels corresponding to a field of view width of 4800 mm for the camera component, 4800 divided by 1920 is calculated to obtain a calibration scale coefficient of 2.5 mm per pixel. For the path point coordinates (500, 600), the horizontal physical coordinate is calculated as 500 multiplied by 2.5 plus 150 to obtain 1400 mm, and the vertical physical coordinate is calculated as 600 multiplied by 2.5 plus 200 to obtain 1700 mm. The calculated physical coordinate values ​​are encapsulated into hexadecimal instruction codes according to the industrial robot control protocol, and start characters and check bits are added. The pixel coordinates are mapped to the physical operation coordinate system of the pruning equipment through coordinate transformation and scaling, and converted into boundary instruction data matching the equipment control protocol to establish an adaptive and precise pruning boundary image for the green belt.

[0153] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An intelligent adaptive precision pruning system for green belts, characterized in that, The system includes: The image time sequence acquisition module acquires the image sequence obtained by the front-end camera component of the trimming device, writes it into the buffer in chronological order and performs consistency verification to establish a continuous frame pixel time sequence image set; The flexible micro-motion recognition module calculates the gray-level difference between adjacent frames based on the continuous frame pixel time-series image set, constructs the oscillation discrimination criteria, and marks the pixel positions that meet the conditions through sign switching counting and amplitude equalization to obtain a flexible pixel candidate distribution map. The vegetation color orientation discrimination module calls the flexible pixel candidate distribution map to obtain pixel color orientation information and performs consistency discrimination with the reference color orientation. The pixel coordinates that pass the screening are retained to obtain the vegetation effective pixel recognition map. The spatial contour aggregation module obtains the spatial distribution information of pixels in the effective pixel recognition map of vegetation and performs density statistics and neighborhood aggregation processing to remove isolated pixels and generate a continuous contour pixel map of vegetation. The trimming boundary recognition module determines the connectivity of contour pixels and adjusts the edge sequence based on the continuous contour pixel map of the vegetation, constructs an image boundary that can be recognized by the trimming device, and establishes an adaptive and precise trimming boundary image for the green belt.

2. The intelligent green belt adaptive precision pruning system according to claim 1, characterized in that, The continuous frame pixel temporal image set includes a pixel matrix sequence, a frame index list, and a pixel coordinate mapping table; the flexible pixel candidate distribution map specifically includes zero-crossing rate features, mean absolute difference, and channel index; the vegetation effective pixel recognition map specifically refers to RGB color vector, spectral angle, and vector projection modulus; the vegetation continuous contour pixel map includes spatial point density, mean drift vector, and cluster centroid; the green belt adaptive precise pruning boundary image specifically includes boundary chain code, curvature features, and motion trajectory.

3. The intelligent green belt adaptive precision pruning system according to claim 1, characterized in that, The pixel coordinates are data indexes that identify the two-dimensional positions of feature points in the matrix and are transformed and mapped to the physical operating positions of the equipment. The pixel color orientation information is a data feature that is extracted from the red, green and blue channel values, converted into a three-dimensional vector, and compared with the vegetation reference vector to perform a spatial angle consistency comparison to eliminate the interference of light and shadow. The outline pixels are effective points that are grouped into continuous vegetation morphology after spatial density statistics and neighborhood aggregation processing.

4. The intelligent green belt adaptive precision pruning system according to claim 1, characterized in that, The isolated pixels are invalid points that are considered noise and removed during the aggregation process because they are not merged or the number of points in their respective regions is scarce; The adjusted edge sequence is to calculate the abrupt change position of the vector angle of the connected linked list, and to perform smoothing correction and sequential recombination to optimize the pruning trajectory based on mechanical motion constraints; The image boundary is a visual boundary line that is converted into device control commands based on a continuous contour map after connectivity analysis and path optimization.

5. The intelligent green belt adaptive precision pruning system according to claim 1, characterized in that, The image timing acquisition module includes: The data acquisition and conversion submodule acquires images continuously obtained by the front-end camera component of the pruning equipment in the green belt operation scenario, obtains the pixel matrix, frame order index and pixel two-dimensional coordinate information corresponding to each frame, and establishes a frame-level pixel coordinate matrix. The cache verification and analysis submodule calls the frame-level pixel coordinate matrix to write the continuous frame images into a fixed-depth cache area in chronological order. After detecting that the number of frames in the cache area reaches the preset frame depth condition, it calculates the position difference between the pixel position of each frame in the cache area and the position reference benchmark. It then performs consistency verification in combination with the preset deviation tolerance range to generate a pixel-by-pixel temporal consistency verification coefficient. The temporal sequence construction submodule, based on the pixel-by-pixel temporal consistency verification coefficient, filters frame data that meets the consistency requirements, performs temporal reorganization on the pixel positions of each frame in the buffer, forms an image sequence structure that can be used for pixel-by-pixel temporal analysis, and obtains a set of continuous frame pixel temporal images.

6. The intelligent green belt adaptive precision pruning system according to claim 1, characterized in that, The flexible micro-motion recognition module includes: The symbol sequence statistics submodule calls the continuous frame pixel time-series image set, calculates the gray-level difference value at the same coordinate position in adjacent frames, constructs a symbol change sequence based on the positive and negative polarity of the difference value, counts the cumulative number of positive and negative state switching in the sequence, and obtains the pixel oscillation frequency statistics value. The amplitude averaging operation submodule calls the continuous frame pixel time-series image set, extracts the gray-level difference values ​​between all adjacent frames in the sequence, calculates the sum of the absolute values ​​of the difference values, calculates the mean of the sum of absolute values ​​based on the time span of the sequence, and generates the gray-level micro-motion amplitude mean. The dual threshold mapping submodule compares the frequency values ​​and amplitude averages of the pixel oscillation frequency with the preset frequency lower limit and amplitude constraint parameters, respectively, based on the statistical values ​​of the pixel oscillation frequency and the average amplitude of the gray-scale micro-motion. It then filters the pixel positions and channel information that simultaneously meet the dual threshold conditions and establishes a flexible pixel candidate distribution map.

7. The intelligent green belt adaptive precision pruning system according to claim 1, characterized in that, The vegetation color orientation discrimination module includes: The three-dimensional vector construction submodule detects the red, green and blue channel values ​​of the marked pixels in the flexible pixel candidate distribution map, constructs a three-dimensional color direction vector at the pixel end, obtains preset vegetation reference color direction data and constructs a reference vector, and generates a three-dimensional color direction reference vector set. The consistency operation submodule calls the three-dimensional color direction reference vector set, calculates the ratio of the dot product value between the three-dimensional color direction vector at the pixel end and the reference reference vector to the product of the magnitudes of the two vectors, and obtains the color direction consistency discrimination coefficient; The illumination projection verification submodule filters pixels whose coefficients meet the consistency threshold requirements based on the color direction consistency discrimination coefficient, calculates the projection length of the pixel vector in the reference vector direction, compares the projection length with the illumination intensity interval threshold, retains the pixel coordinates that meet the dual verification conditions, and establishes a vegetation effective pixel recognition map.

8. The intelligent green belt adaptive precision pruning system according to claim 1, characterized in that, The spatial contour aggregation module includes: The grid density mapping submodule obtains preset image space grid division parameters for the two-dimensional coordinate data of each pixel in the effective pixel recognition image of the vegetation, maps the pixel coordinates to the corresponding spatial grid unit system, counts the number of pixels contained in each grid unit, and generates a grid pixel density distribution table. The density gravity displacement submodule calls the grid pixel density distribution table, identifies the current grid and its neighboring grid cells, calculates the density difference between the neighboring grid and the central grid, determines the aggregation vector for pixels to move toward the high-density region by combining the spatial distance span between grids, iteratively updates the pixel coordinates within the cell based on the aggregation vector, and generates a set of pixel aggregation displacement coordinates. The connected region merging submodule calculates the updated Euclidean space distance between pixels based on the pixel aggregation displacement coordinate set, determines whether the distance is less than a preset similar merging threshold, merges pixels that meet the condition into connected regions, removes isolated pixels that have not been aggregated, and establishes a continuous vegetation contour pixel map.

9. The intelligent green belt adaptive precision pruning system according to claim 1, characterized in that, The trimming boundary recognition module includes: The edge sequence construction submodule, based on the continuous contour pixel map of the vegetation, traverses the pixel coordinate data in the image plane, calculates the Euclidean distance between adjacent pixels, determines spatial connectivity, sorts the pixels according to the connectivity relationship, constructs a contour linked list structure with the head and tail connected, and obtains the edge pixel connectivity sequence. The mutation constraint optimization submodule calls the edge pixel connected sequence, calculates the vector direction angle between adjacent line segments in the sequence path, identifies the location of the angle mutation, and combines the mechanical motion constraint parameters of the pruning device to perform smoothing correction and order rearrangement on the path mutation points to generate an optimized pruning path coordinate set. The physical command conversion submodule, based on the optimized pruning path coordinate set, maps the pixel coordinates to the physical operation coordinate system of the pruning equipment through coordinate transformation and scaling, and converts them into boundary command data that matches the equipment control protocol, thereby establishing an adaptive and precise pruning boundary image for the green belt.

10. The intelligent green belt adaptive precision pruning system according to claim 9, characterized in that, The process of identifying the location of the abrupt change in angle specifically involves extracting the coordinates of three consecutive pixels from the connected sequence of edge pixels and calculating the change in angle between the first vector formed by the first two pixels and the second vector formed by the last two pixels. The process of combining the mechanical motion constraint parameters of the pruning equipment specifically involves obtaining the maximum angular acceleration value and the working linear velocity value of the servo motor of the pruning robot arm, calculating the ratio of the maximum angular acceleration value to the working linear velocity value to obtain the maximum single turning angle value, and rounding it down to set as the maximum allowable deflection angle threshold. The value of the angle change is compared with the maximum allowable deflection angle threshold. When the value of the angle change is greater than the maximum allowable deflection angle threshold, the current position is determined to be a path change point. The process of performing smoothing correction and order rearrangement on the path change point is specifically to calculate the midpoint coordinates of the last two pixels in the coordinates of the three consecutive pixels. The midpoint coordinates are inserted into the corresponding position of the edge pixel connected sequence, and the original angle change value is decomposed into two angle values ​​that are smaller than the maximum allowable deflection angle threshold.

Images