A litchi winter shoot detection and pruning control system based on digital signal processing

Abstract

The present application provides a kind of based on digital signal processing's litchi winter shoot detection and pruning control system, it is related to the field of automatic control, and the present system includes image acquisition terminal and data processing center.Image acquisition terminal synchronous acquisition litchi crown layer visible light and near infrared band multispectral image digital signal.Data processing center includes signal preprocessing unit, feature extraction unit, winter shoot discrimination unit and drive signal unit.Preprocessing unit is by radiation calibration, gain compensation constructs time series vegetation index, and the data field is output after filtering standardization.Feature extraction unit extracts branch area multidimensional feature vector.Discrimination unit determines winter shoot and positioning according to feature vector and cluster center distance and preset bias.Drive signal unit is by coordinate mapping to obtain three-dimensional pruning coordinates, solve joint rotation angle and convert into pulse signal output.The present application realizes litchi winter shoot automatic precision detection and pruning control.

Description

Technical Field

[0001] This invention relates to the field of automation control, specifically to a litchi winter shoot detection and pruning control system and system based on digital signal processing. Background Technology

[0002] Timely management of litchi winter shoots is a key agronomic measure determining the quality of flowering and fruit yield in the following year. In the field of pattern recognition for agricultural phenological monitoring, existing classification systems mostly construct feature spaces based on single-phase sensor data and rely on large-scale labeled samples to train discriminative models. For the task of identifying the growth status of fruit tree shoots, related methods typically extract independent feature dimensions of color and geometry from multispectral data, simply concatenate them, and then input them into the classifier, failing to adequately explore the dynamic evolution patterns inherent in the time-series signals. Simultaneously, the lighting conditions in the agricultural canopy environment are complex and variable, and sensor data from a single moment are easily affected by instantaneous irradiance disturbances, resulting in systematic biases. Existing systems lack effective adaptive filtering mechanisms to address this. A more prominent problem is that temporal dynamic features and spatial structural features are often directly superimposed as orthogonal dimensions in existing feature engineering, failing to establish an intrinsic connection between them at the phenological physical mechanism level. This leads to blurred category boundaries in the feature space and insufficient discriminative confidence. Especially under conditions of scarce sample labels, these defects significantly reduce the system's ability to distinguish subtle differences in growth states, making it difficult to meet the high-reliability classification requirements of precision agriculture. Summary of the Invention

[0003] To achieve the above objectives, the present invention is implemented through the following technical solution: a litchi winter shoot detection and pruning control system based on digital signal processing, comprising an image acquisition terminal and a data processing center;

[0004] The image acquisition terminal is configured to simultaneously acquire raw multispectral image digital signals of the litchi tree canopy in the visible light and near-infrared bands and output them to the data processing center.

[0005] The data processing center includes a signal preprocessing unit, a feature extraction unit, a winter shoot discrimination unit, and a driving signal unit connected in sequence.

[0006] The signal preprocessing unit is configured to perform radiometric calibration on the original multispectral image digital signal to map it into a reflectance value, construct a time-series vegetation index based on the reflectance of the visible light band and the reflectance of the near-infrared band, perform time-domain low-pass filtering on the time-series vegetation index to suppress high-frequency noise of illumination, and output a standardized vegetation index spatiotemporal data field.

[0007] The feature extraction unit is configured to extract a multidimensional feature vector of the branch tip region based on the standardized vegetation index spatiotemporal data field and visible light image signal. The multidimensional feature vector includes at least relative growth activity features and geometric and spectral coupling features.

[0008] The winter shoot discrimination unit is configured to calculate the distance between the multidimensional feature vector and the pre-stored winter shoot cluster center and non-winter shoot cluster center in the feature space. If the distance to the winter shoot cluster center is less than the distance to the non-winter shoot cluster center and the difference exceeds the preset offset, it is determined to be a winter shoot and its position information is output.

[0009] The drive signal unit is configured to obtain the actual three-dimensional trimming coordinates based on the position information through a preset coordinate mapping relationship, and use the actual three-dimensional trimming coordinates as the input of the inverse kinematics solution to solve the target rotation angle of each joint of the trimming actuator, and convert each target rotation angle into a time-sequential digital drive pulse signal and output it to the trimming actuator.

[0010] Preferably, before constructing the time-series vegetation index, the signal preprocessing unit generates a pixel-level band ratio map based on the reflectance of the visible red band and the reflectance of the near-infrared band in the current frame, wherein the value of each pixel in the ratio map is R. local(x,y) =ρ NIR(x,y) / ρ red(x,y) The gain compensation for the reflectivity of the visible light red band is performed using the pixel-level band ratio map as weights. The compensation formula is as follows: , where R global The median of pixel-level band ratios within the canopy area is used to suppress pseudo-variations in reflectance caused by spatial inhomogeneity of illumination through gain compensation. After gain compensation, the signal preprocessing unit constructs a time series of the Normalized Difference Vegetation Index (NDVI) based on the corrected red band reflectance and near-infrared band reflectance. The calculation formula is NDVI=(ρ NIR -ρ red ) / (ρ NIR +ρ red The signal preprocessing unit uses a time-domain low-pass filtering method with sliding polynomial fitting to smooth the temporal vegetation index. For the vegetation index value of frame t, the system uses the frame data of several frames before and after it, with a preset filtering window width, as the window, and fits the polynomial g(z) = a0 + a1·z + a2·z² + ... + a n ·z n The value of the polynomial at the center of the window, g(0) = a0, is used as the vegetation index value of the filtered frame. After temporal filtering, the signal preprocessing unit performs Z-score normalization on the vegetation index values ​​within the canopy area of ​​a single frame. The normalization formula is NDVI. std(x,y,t) =[NDVI(x,y,t)-μ(t)] / σ(t), where μ(t) is the mean NDVI of the canopy region in frame t, and σ(t) is the standard deviation of NDVI of the canopy region in frame t, forming a standardized temporal vegetation index spatiotemporal data field with pixel coordinates as the spatial dimension and frame number as the time dimension.

[0011] Preferably, before extracting various features, the feature extraction unit first uses the temporal variance of the near-infrared reflectance of each pixel within a preset time window as a stability criterion to calculate the variance of the near-infrared reflectance of each pixel within a consecutive N-frame time window. Pixels with variance less than a preset fluctuation threshold are marked as stable candidate pixels, and all subsequent feature extraction operations are limited to the range of stable candidate pixels.

[0012] Preferably, the relative growth activity feature is extracted through inter-frame spatial registration and global baseline drift subtraction. The feature extraction unit first uses a feature point detection and matching algorithm to complete the spatial registration between adjacent frames, eliminating the inter-frame displacement deviation caused by the flight attitude change of the image acquisition terminal. After spatial registration, taking the registered branch area of ​​the current frame as the center, pixels whose vegetation index temporal change rate ΔNDVI(x,y,t)=|NDVI(x,y,t)-NDVI(x,y,t-1)| is less than the preset stability threshold and do not belong to the winter shoot candidate area marked in the previous frame are selected as background reference pixels. The mean of the first-order temporal difference of all background reference pixels between adjacent frames is calculated as the whole tree baseline drift. The relative growth activity value of each pixel is obtained by subtracting the whole tree baseline drift from the first-order temporal difference value of each pixel in the branch area. The mean of the relative growth activity values ​​of all stable candidate pixels in the branch area is taken as the relative growth activity feature of the branch area.

[0013] Preferably, the geometric and spectral coupling features are extracted by combining spectral thickness weighting with geometric direction analysis; the feature extraction unit first extracts the branch region from the visible light image, and performs edge detection on the branch region in the visible light image using an edge detection operator to obtain the gradient magnitude map G(x,y); based on the time-series mean of near-infrared reflectance μ of stable candidate pixels... NIR(x,y) Compared with the preset reference reflectivity R base Construct the steady-state canopy thickness index H(x,y)=[μ NIR(x,y) -R base ] / R base Using the steady-state canopy thickness index as the spatial weight, the gradient magnitude within the stable candidate pixel region is spatially weighted to obtain the weighted gradient field G. w(x,y) =G(x,y)×[1+H(x,y)]; Perform principal component analysis on the two-dimensional coordinates of each pixel in the branch region, determine the direction of the first principal component as the branch principal axis direction, and the direction of the second principal component as the branch secondary axis direction, and calculate the gradient energy of the weighted gradient field in the branch principal axis direction respectively. With gradient energy in the secondary axis direction Find the ratio R between the two. g=E axis1 / E axis2 Simultaneously calculate the aspect ratio R of the minimum bounding rectangle of the branch tip region. rect And the regional mean CV of the temporal variation coefficient of near-infrared reflectance of stable candidate pixels. avg Where the coefficient of variation CV = σ NIR / μ NIR Finally, the geometric and spectral coupling characteristics F are obtained. geom =R g ×R rect ×(1+CV avg ).

[0014] Preferably, the multidimensional feature vector further includes chromaticity distribution features; within the stable candidate pixel set, the feature extraction unit divides the stable candidate pixel set into a highlight stable subset, a shadow stable subset, and a transitional stable subset based on the temporal mean of the near-infrared reflectance of each pixel. The division threshold is obtained by estimating the kernel density of the temporal mean of the near-infrared reflectance of the stable candidate pixels, using the local minimum valley of the kernel density curve as the division boundary; the hue mean of the three stable subsets in the HSV color space is extracted respectively. During the calculation, the circular mean algorithm is used to handle the angular periodicity of the hue. The hue mean is obtained by calculating the arctangent of the mean of the sine and cosine values ​​of the hue of each pixel; the hue mean of the transitional stable subset is H transition Based on this, the ratio of the hue mean of the highlight-stable subset to the shadow-stable subset is normalized to obtain H. high_norm =H high / H transition H shadow_norm =H shadow / H transition H high H is the circular mean of the hue of the highlight-stable subset. shadow H is the circular mean of the hue of the stable subset of shadows. transition The hue circular mean of the transitional stable subset; normalized reference values ​​1 and H are adjusted according to preset weights w1, w2, and w3. high_norm H shadow_norm We perform a weighted summation to obtain the basic value C of the chromaticity distribution characteristics. base =w1×1+w2×H high_norm +w3×H shadow_norm Then, the inverse of the time-series variation coefficient of near-infrared reflectance of each stable candidate pixel is used as the chromaticity confidence weight to weight the base values, thus obtaining the final chromaticity distribution feature C. final =C base ×[1 / (1+CV avg The chromaticity distribution features are incorporated into the initial multidimensional feature vector to form a complete multidimensional feature vector.

[0015] Preferably, the winter shoot discrimination unit pre-stores winter shoot cluster centers C. bud Non-winter shoot cluster center C non The clustering is pre-trained using historical sample data. Specifically, complete multidimensional feature vectors are extracted from manually labeled samples, and the mean of each dimension is calculated as the cluster center. The initial bias value is determined statistically based on the sample data. If the multidimensional feature vector X of a certain branch region satisfies dist(X,C)... bud ) <dist(X,C non And dist(X,C) non )-dist(X,C bud If a preset bias is set, the branch is determined to be a winter shoot and its location information is output. After the winter shoot identification is completed in the current frame, the winter shoot identification unit uses the multi-dimensional feature vectors corresponding to the regions identified as winter shoots and the regions identified as non-winter shoots as incremental samples, and updates the pre-stored cluster centers using the exponentially weighted moving average method. The update formula is C. new =(1-λ)×C old +λ×X sample Where λ is the forgetting factor, which uses an exponential decay function λ(Δn) = λ0·exp(-k·Δn), λ0 is the initial forgetting factor, k is the decay coefficient, and Δn is the difference in frame number between the current frame and the initial frame. new For the updated cluster centers, C old X is the cluster center before the update. sample This is the feature vector of the new sample in the current frame.

[0016] Preferably, the winter shoot discrimination unit calculates the inter-class scatter matrix trace tr(S) of the two types of regions in the feature space based on the multi-dimensional feature vectors of the winter shoot region and the non-winter shoot region identified in the previous frame. B )=||μ bud -μ non ||², where μ bud Let μ be the mean vector of the winter shoot region. non The mean vector of the non-winter shoot region is used; when the trace of the inter-class scatter matrix is ​​less than the preset stability threshold, the preset bias is increased by a preset first fixed step size; when the trace of the inter-class scatter matrix is ​​greater than or equal to the preset stability threshold, the preset bias is decreased by a preset second fixed step size, and the upper and lower limits of the preset bias value are set at the same time.

[0017] Preferably, the coordinate mapping relationship of the drive signal unit is constructed jointly by camera calibration and on-site hand-eye calibration. The camera intrinsic parameters are obtained by planar calibration, and the camera extrinsic parameters are obtained by the positioning module mounted on the image acquisition terminal. The transformation relationship between the base coordinate system and the world coordinate system of the trimming actuator is obtained by on-site hand-eye calibration. Through inverse perspective projection transformation combined with coordinate system transformation, the two-dimensional coordinates of the image are mapped to the three-dimensional actual trimming coordinates under the base coordinate system of the trimming actuator. The mapping relationship is the homogeneous coordinate transformation equation λ·[u,v,1]. T =K·[R|t]·[X,Y,Z,1] T Where λ is the depth factor, K is the camera intrinsic parameter matrix, and R and t are the rotation matrix and translation vector from the camera coordinate system to the robot's base coordinate system. Solving this equation yields the actual 3D trimming coordinates (X,Y,Z). The trimming actuator uses a multi-degree-of-freedom articulated robot. For example, a kinematic model is established using standard DH parameters. With the actual 3D trimming coordinates as input, the inverse kinematics is solved using an algebraic analytical method, and the target rotation angle of each joint is obtained sequentially.

[0018] Preferably, after the drive signal unit obtains the target rotation angle of each joint, it extracts the branch edge contour in the neighborhood according to the corresponding coordinates of the winter shoot position information in the visible light image. Using the branch edge contour as the collision constraint boundary, the end effector of the pruning actuator is restricted from entering the internal area enclosed by the contour. Taking the current joint angle of the pruning actuator as the starting point and the target rotation angle as the ending point, a fast search random tree algorithm is used to plan a collision-free trajectory in the joint space. The search step size of the fast search random tree algorithm is set according to the pixel ratio of the branch edge contour in the neighborhood according to a preset segmented mapping rule. The pixel ratio is defined as the ratio of the number of edge pixels in the neighborhood window to the total number of pixels in the window. The pixel ratio is inversely proportional to the search step size, that is, the larger the pixel ratio, the smaller the search step size.

[0019] Preferably, after the pruning execution mechanism completes a single pruning, the data processing center controls the image acquisition terminal to acquire visible light and near-infrared images of the pruning section. The spatial distribution R(λ) of the near-infrared reflectance of the pruning section is then correlated with the pre-stored spectral feature template T(λ) of the xylem of winter shoots. The correlation matching uses the Pearson correlation coefficient. Where r is the Pearson correlation coefficient, R i To trim the reflectance value of the cross-sectional image at the i-th wavelength sampling point, T i μ is the reflectance value of the xylem spectral characteristic template of winter shoots at the i-th wavelength sampling point. R μ is the mean of the cross-sectional reflectance of all sampling points. TThe mean of the template reflectance of all sampling points; when the correlation coefficient is greater than the preset matching threshold, it is determined that the winter shoots have not been completely removed, and a supplementary cutting signal is output to the drive signal unit to drive the pruning execution mechanism to perform secondary pruning on the remaining winter shoots.

[0020] This invention provides a litchi winter shoot detection and pruning control system based on digital signal processing. It has the following beneficial effects:

[0021] 1. This invention solves the problem of unstable feature space distribution caused by the coupling of uneven spatial distribution of light and temporal fluctuations in traditional methods, and improves the anti-interference ability of pattern recognition system against the fluctuation of sensor data quality in complex agricultural canopy environments.

[0022] 2. This invention physically correlates the temporal dynamic features of vegetation index with the spatial structural features of canopy thickness, uses the steady-state canopy thickness index as the global weight to spatially modulate the gradient field, and uses the temporal variation coefficient of near-infrared reflectance as a spatial distribution adjustment factor to correct the geometric extension features. This solves the problem of blurred category boundaries caused by the simple superposition of temporal dynamic features and spatial structural features as orthogonal dimensions in the existing system, and improves the distinguishability and discrimination confidence of feature vectors for subtle differences in growth states.

[0023] 3. This invention solves the problems of category distribution drift and discrimination threshold rigidity caused by individual differences in the tree structure and temporal evolution under small sample conditions, and improves the system's online learning ability and long-term discrimination stability for unlabeled samples. Attached Figure Description

[0024] Figure 1 This is a system framework diagram as described in this invention. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] refer to Figure 1The litchi winter shoot detection and pruning control system based on digital signal processing according to the present invention is applied to a litchi plantation in a certain area. The plantation covers an area of 52 mu, with the cultivated variety being Feizixiao, the tree age ranging from 8 to 12 years, and the canopy height ranging from 2.6 meters to 3.4 meters. The image acquisition terminal adopts a multispectral drone equipped with a MicaSense RedEdge-MX multispectral camera, which synchronously collects image data of 5 bands, including the blue band with a central wavelength of 475 nm, the green band with a central wavelength of 560 nm, the red band with a central wavelength of 668 nm, the red edge band with a central wavelength of 717 nm, and the near-infrared band with a central wavelength of 840 nm. The spatial resolution of the images is 2.5 centimeters of actual ground size per pixel. The image acquisition terminal conducts aerial photography on the litchi canopy layer at an acquisition rate of 1 frame per second according to a preset route, synchronously collects the original multispectral image digital signals of the visible light band and the near-infrared band, and outputs them to the data processing center deployed on the ground workstation through a wireless data transmission link. The ground workstation is equipped with an Intel Xeon E5-2690 central processing unit, an NVIDIA Quadro P4000 graphics processing unit, and 32 gigabytes of memory, which can complete the full-process signal processing and motion control operations.

[0027] After receiving the original multispectral image digital signal, the signal preprocessing unit first performs radiometric calibration on the original multispectral image digital signal to map the original digital quantization value to the surface reflectance value. The radiometric calibration adopts the empirical linear method. Before each flight operation, a standard reflectance white board is laid on the ground, and the reflectance of the white board in the visible light to near-infrared band is calibrated to 99% in the laboratory. At the same time, an artificial lawn is selected as a dark target in the orchard area to be tested, and the measured reflectance of the dark target in the corresponding band is 4%. The radiometric calibration equation is ρ i =Gain i ×DN i +Bias i , where ρ i is the reflectance of the i-th band, DN i is the original digital quantization value of this band, Gain i and Bias i are the gain coefficient and offset coefficient of this band, respectively. The gain coefficient and the offset coefficient are obtained by substituting the known reflectance of the standard white board and its image DN value, as well as the known reflectance of the dark target and its image DN value into the above equation for simultaneous solution.

[0028] After the radiometric calibration is completed, the signal preprocessing unit generates a pixel-level band ratio map based on the reflectance of the visible red band and the reflectance of the near-infrared band of the current frame, and the value of each pixel in the ratio map is R local(x,y) =ρ NIR(x,y) / ρ red(x,y)The gain compensation for the reflectivity of the visible light red band is performed using the pixel-level band ratio map as weights. The compensation formula is as follows: , where R global R is the median of the pixel-level band ratios within the tree canopy region of the entire image. local(x,y) The band ratio at pixel (x,y) is used to suppress pseudo-changes in reflectivity caused by spatial non-uniformity of illumination through gain compensation. The canopy region is initially extracted through initial threshold segmentation, with the initial NDVI value set to 0.2.

[0029] After gain compensation, the signal preprocessing unit constructs a time-series vegetation index based on the reflectance of the visible red band and the near-infrared band. In this embodiment, the Normalized Difference Vegetation Index (NDVI) is used, and its calculation formula is NDVI=(ρ NIR -ρ red ) / (ρ NIR +ρ red ), where ρ NIR ρ is the reflectance in the near-infrared band. red To compensate for the red light band reflectance, the time-series sequence is accumulated frame by frame at a rate of 1 frame per second to form the NDVI time-series data for each pixel. After the construction of the time-series vegetation index is completed, the signal preprocessing unit performs time-domain low-pass filtering on the time-series vegetation index to suppress high-frequency noise from illumination. The filtering method adopts the Savitzky-Golay filter with a filtering window width of 7 frames and a smoothing polynomial order of 2. For the NDVI value of the t-th frame, the data of the 7 frames before and after it are used as the window, and a second-order polynomial g(z)=a0+a1·z+a2·z² is fitted within the window. The value of the polynomial at the center of the window, i.e., z=0, g(0)=a0, is used as the filtered NDVI value of that frame. The selection of the window width and polynomial order is based on the following: Power spectral density analysis of 120 consecutive frames of raw NDVI time-series data revealed that high-frequency noise from illumination is mainly concentrated in the frequency band with a normalized frequency greater than 0.25 Hz, while the effective signal of vegetation phenological changes is concentrated in the frequency band with a normalized frequency less than 0.08 Hz. A window width of 7 frames corresponds to an equivalent low-pass cutoff frequency of 0.18 Hz, which can filter out most high-frequency noise while retaining the effective signal. The mean square error (RMSE) of the 2nd-order polynomial fitting within the 7-frame window is 0.007, while the mean RMSE of the 3rd-order polynomial fitting is 0.0065. The difference is less than 8%, but the 3rd-order polynomial is more susceptible to overfitting due to outliers within the window; therefore, the 2nd-order polynomial was chosen. After temporal filtering, the signal preprocessing unit performs Z-score normalization on the NDVI values ​​of all canopy regions within a single frame. The normalization formula is: NDVI std(x,y,t)=[NDVI(x,y,t)-μ(t)] / σ(t), where μ(t) is the mean NDVI of the canopy region in frame t, and σ(t) is the standard deviation of NDVI of the canopy region in frame t; after standardization, a standardized vegetation index spatiotemporal data field is formed. This data field is stored with pixel coordinates as the spatial dimension and frame number as the time dimension, and is output to the feature extraction unit.

[0030] After receiving the standardized vegetation index spatiotemporal data field and visible light image signal, the feature extraction unit first extracts the branch and shoot regions from the visible light red band image. In this embodiment, the Otsu's maximum inter-class variance method is used to obtain the segmentation threshold. Combined with morphological opening operations to remove minor noise, the candidate regions of branches and shoots within the canopy are initially segmented. Then, the leaf regions are removed by combining the near-infrared reflectance threshold, finally obtaining an accurate mask of the branch and shoot regions. Subsequent feature extraction is performed within this mask. Subsequently, the feature extraction unit uses the temporal stability of the near-infrared reflectance of each pixel in the current frame as a screening index. Within a continuous 15-frame time window, the variance of the near-infrared reflectance of each pixel is calculated. Pixels with a variance less than the preset fluctuation threshold of 0.005 are marked as stable candidate pixels, and subsequent features are extracted only for stable candidate pixels. The fluctuation threshold of 0.005 was determined as follows: a time-series stability statistical analysis was performed on 200 randomly selected pixels in the test area. Among them, 146 pixels were manually judged to be time-series stable. The mean of the variance of their near-infrared reflectance was 0.0021 and the standard deviation was 0.0017. The mean plus twice the standard deviation, i.e., 0.0021 plus 0.0034, equals 0.0055. After rounding, 0.005 was used as the threshold.

[0031] When extracting relative growth activity features, the feature extraction unit uses a scale-invariant feature transformation algorithm to detect and match feature points in adjacent frames. A 16×16 pixel neighborhood centered on the feature point is taken as the descriptor sampling window, generating a 128-dimensional feature vector. The Euclidean distance minimum matching criterion is used to complete spatial registration between adjacent frames, eliminating inter-frame displacement deviations caused by changes in UAV flight attitude. Centered on the registered branch region of the current frame, pixels within a 15×15 pixel spatial neighborhood window whose NDVI temporal change rate is less than the preset stability threshold of 0.005 and do not belong to the winter shoot candidate region marked in the previous frame are selected as background reference pixels. The NDVI temporal change rate is defined as the absolute value of the difference between the NDVI value of the current frame and the NDVI value of the previous frame. The mean of the first-order temporal differences of all background reference pixels between adjacent frames is calculated as the whole-tree baseline drift. The formula for calculating this first-order temporal difference is ΔNDVI. bg(x,y,t) =NDVI bg(x,y,t) -NDVI bg(x,y,t-1) whole tree baseline drift The relative growth activity value of each pixel is obtained by subtracting the tree-wide baseline drift from the first-order temporal difference value of each pixel in the branch tip region. The mean of the relative growth activity values ​​of all stable candidate pixels in the branch tip region is taken as the relative growth activity feature.

[0032] When extracting geometric and spectral coupling features, the feature extraction unit performs edge detection on the branch region in the visible light image signal, for example, using the Canny operator to obtain a gradient magnitude map. The high threshold of the Canny operator is set to 20% of the maximum gradient magnitude, and the low threshold is set to 10% of the maximum gradient magnitude, with a Gaussian filter standard deviation of 1.4. The selection process for the high and low thresholds is as follows: an edge detection parameter optimization experiment was conducted on 50 visible light images of branches under different lighting conditions in the test area. Using manually labeled edges as the benchmark, the average F1-score was 0.87 when the high threshold was 20% of the maximum gradient, 0.76 when the high threshold was 30% of the maximum gradient, and 0.79 when the high threshold was 10% of the maximum gradient. Therefore, 20% was selected. The average edge connectivity integrity was 92% when the low threshold was 10% of the maximum gradient, 88% when it was 15% of the maximum gradient, and 40% when it was 5% of the maximum gradient. Therefore, 10% was selected.

[0033] The feature extraction unit constructs a steady-state canopy thickness index based on the difference between the temporal mean of near-infrared reflectance of each stable candidate pixel and a preset benchmark reflectance of 0.42. The preset benchmark reflectance of 0.42 was obtained through field measurement and calibration of the near-infrared reflectance of healthy, mature leaves of the same litchi variety at the 840 nm wavelength. The sample size was 100 leaves, with a mean of 0.418 and a standard deviation of 0.035. The mean of 0.42 was taken as the benchmark reflectance. The steady-state canopy thickness index is calculated using the formula H(x,y)=[μ NIR(x,y) -0.42] / 0.42, where μ NIR(x,y) The near-infrared reflectance of pixel (x, y) is the temporal mean over a 15-frame time window. Using the steady-state canopy thickness index as the spatial weight, the visible light gradient magnitudes of each pixel within the stable candidate pixel region are spatially weighted to obtain the weighted gradient field G. w(x,y) =G(x,y)×[1+H(x,y)]. The feature extraction unit further performs principal component analysis on the two-dimensional coordinates of each pixel in the branch region to determine the principal axis and secondary axis directions of the branch, where the first principal component direction is the principal axis direction and the second principal component direction is the secondary axis direction. The gradient energy of the weighted gradient field in the principal axis direction and the gradient energy in the secondary axis direction are calculated respectively. Gradient energy along the secondary axis The ratio R between the two g =E axis1 / E axis2 The feature extraction unit calculates the temporal coefficient of variation of the near-infrared reflectance of each stable candidate pixel, where the coefficient of variation CV = σ. NIR / μ NIR , where σ NIR The standard deviation of the near-infrared reflectance of this pixel within 15 frames is taken as the ratio of the mean of the coefficients of variation of all pixels in the region to the aforementioned gradient energy value R. g The product is used as a spatial distribution adjustment factor for the geometric and spectral coupling features. Simultaneously, the feature extraction unit calculates the aspect ratio of the minimum bounding rectangle of the branch region, and uses this spatial distribution adjustment factor to correct the product of the gradient energy ratio and the aspect ratio, ultimately obtaining the geometric and spectral coupling feature F. geom =R g ×R rect ×(1+CV avg ), where R rect CV is the aspect ratio of the minimum bounding rectangle of the branching region. avg The mean of the coefficients of variation within the region is used to form the initial multidimensional feature vector.

[0034] The multidimensional feature vector in this embodiment also includes chromaticity distribution features. Within the stable candidate pixel set, the feature extraction unit divides the set into three subsets based on the temporal mean of the near-infrared reflectance of each pixel: the specular stable subset is defined as μ. NIR The set of pixels with a value ≥ 0.45, the shadow-stable subset is defined as μ. NIR The set of pixels with a value < 0.38 is defined as the transitionally stable subset, where 0.38 ≤ μ. NIR The set of pixels with a value <0.45. The process for determining the above three partitioning thresholds is as follows: for a total of 12400 pixels in the stable candidate pixel set, μ... NIR Kernel density estimation was performed on the values. The kernel density curve showed three local maxima peaks located at 0.32, 0.41, and 0.48, and two local minima valleys located at 0.38 and 0.45, naturally forming three clustering intervals. Therefore, the valley values ​​of 0.38 and 0.45 were used as the dividing boundaries. The hue mean of the three stable subsets in the HSV color space was extracted. Since hue is an angular quantity, it has a periodic boundary between 0 degrees and 360 degrees. Direct arithmetic averaging is prone to boundary errors. Therefore, a circular mean algorithm was used to calculate the hue mean. The calculation formula is that the hue mean is equal to the arctangent of the mean of the sine and cosine values ​​of the hue of each pixel. Finally, the hue mean of the three subsets was obtained, denoted as H. high H shadow and H transition The hue mean H of the transitionally stable subset. transition Based on this, the ratio of the hue mean of the highlight stable subset and the hue mean of the shadow stable subset is normalized. The normalized ratio yields the highlight normalized hue H.high_norm =H high / H transition Shadow normalized hue H shadow_norm =H shadow / H transition The normalized baseline value 1, the normalized hue of highlights, and the normalized hue of shadows of the transitional stable subset are weighted and summed with weights of 0.6, 0.2, and 0.2, respectively, to obtain the basic value C of the chromaticity distribution characteristics. base Then, using the reciprocal of the time-series variation coefficient of near-infrared reflectance of each stable candidate pixel as the chromaticity confidence weight, the confidence weights of all pixels within the region are normalized and multiplied by the basic chromaticity value to complete the weighted calculation, thus obtaining the final chromaticity distribution feature C. final =C base ×[1 / (1+CV avg )], where CV avg To stabilize the candidate pixel region, the arithmetic mean of the coefficients of variation of all pixels is calculated. The values ​​of each component of the chromaticity distribution feature are incorporated into the initial multidimensional feature vector to form a complete multidimensional feature vector. This complete multidimensional feature vector contains, in turn, 1 dimension of relative growth activity feature, 3 dimensions of geometric and spectral coupling feature, and 4 dimensions of chromaticity distribution feature, for a total of 8 dimensions.

[0035] After receiving the complete multidimensional feature vector, the winter shoot discrimination unit calculates the Euclidean distance in the feature space between the multidimensional feature vector of each shoot region and the pre-stored winter shoot cluster centers and non-winter shoot cluster centers. The pre-stored winter shoot cluster centers and non-winter shoot cluster centers are pre-trained from historical sample data. Specifically, 300 samples manually labeled as winter shoots and 300 samples manually labeled as non-winter shoots are selected, and the 8-dimensional complete multidimensional feature vectors of each sample are extracted. The mean values ​​of each dimension for winter shoot samples are [0.23, 1.45, 0.67, 0.52, 0.34, 0.71, 0.43, 0.56], and the mean values ​​of each dimension for non-winter shoot samples are [0.08, 2.10, 0.32, 0.28, 0.51, 0.39, 0.62, 0.48]. These are used as the winter shoot cluster centers C. bud Non-winter shoot cluster center C non The initial bias value is set to 0.15. This initial value is determined as follows: Based on the 300 winter shoot samples and 300 non-winter shoot samples, the Euclidean distance from each sample to its respective cluster center is calculated. The average distance from the winter shoot samples to the winter shoot cluster center is 0.32 with a standard deviation of 0.13, and the average distance from the non-winter shoot samples to the non-winter shoot cluster center is 0.28 with a standard deviation of 0.12. The lower limit of the 95% confidence interval for the difference in distances between the two types of samples to their respective cluster centers is approximately 0.14. Therefore, 0.15 is used as the initial bias. If the multidimensional feature vector X of a certain branch region satisfies dist(X,C)... bud ) <dist(X,C nonAnd dist(X,C) non )-dist(X,C bud If a preset bias is applied, the branch is determined to be a winter shoot and its location information is output. After the winter shoot identification is completed in the current frame, the winter shoot identification unit uses the multi-dimensional feature vectors corresponding to the regions identified as winter shoots and the regions identified as non-winter shoots as incremental samples. These are then updated using an exponentially weighted moving average method to update the pre-stored winter shoot cluster centers and non-winter shoot cluster centers, respectively. The update formula is C. new =(1-λ)×C old +λ×X sample C old X is the cluster center before the update. sample Let λ be the feature vector of the new sample in the current frame, and λ be the forgetting factor. The forgetting factor adopts an exponential decay function λ(Δn) = λ0·exp(−k·Δn), where λ0 is the initial forgetting factor with a value of 0.1, k is the decay coefficient with a value of 0.01, and Δn is the difference in frame number between the current frame and the initial frame. The parameters of this decay function are determined as follows: when Δn = 0, λ = 0.1, and the weight of the new sample is 0.1; when Δn = 10, λ = 0.1 × exp(-0.1) = 0.0905, and the weight of the new sample increases slightly; when Δn = 50, λ = 0.1 × exp(-0.5) = 0.0607, and the weight of the new sample gradually decays. After 100 frames, the influence of historical data gradually decreases, allowing the cluster centers to adjust slowly with the seasons and growth status, ensuring the long-term stability of the cluster centers.

[0036] The winter shoot discrimination unit also calculates the trace of the inter-class divergence matrix S in the feature space based on the multi-dimensional feature vectors of the winter shoot region and the non-winter shoot region identified in the previous frame. B =(μ bud -μ non )(μ bud -μ non ) T , its trace tr(S B )=||μ bud -μ non ||², which is the squared L2 norm of the difference between the means of two classes. When tr(S B When tr(S) is less than the preset stable threshold of 0.5, the preset bias is increased by a preset first fixed step size of 0.02; when tr(S) is less than the preset stable threshold of 0.5, the preset bias is increased by a preset first fixed step size of 0.02; B When the value is greater than or equal to the preset stability threshold of 0.5, the preset bias is reduced by a preset second fixed step size of 0.01. The preset stability threshold of 0.5 is determined by calculating tr(S) over 100 consecutive frames during runtime. B ) value, when tr(S BWhen the bias is below 0.5, the Mahalanobis distance between the two classes of samples in the feature space is determined to be less than 2.0, indicating insufficient inter-class discriminability. Increasing the bias can reduce the misclassification rate caused by feature aliasing. B When the bias is above 0.5, the inter-class discrimination is sufficient, and reducing the bias can improve the detection rate of boundary samples. At the same time, a preset upper and lower limit for the bias value is set, with a minimum value of 0.05 and a maximum value of 0.3, to avoid the bias exceeding the reasonable range.

[0037] After receiving the position information output by the winter shoot discrimination unit, the drive signal unit calculates the actual three-dimensional pruning coordinates based on the position information and a preset coordinate mapping relationship. The coordinate mapping relationship is constructed jointly through camera calibration and on-site hand-eye calibration. The camera intrinsic parameters are obtained using the Zhang Zhengyou calibration method, and the focal length f... x =852 pixels, f y =850 pixels, principal point coordinates (c x ,c y (640 pixels, 512 pixels), radial distortion coefficients k1=-0.28, k2=0.12, tangential distortion coefficients p1=-0.001, p2=0.0015; extrinsic parameters are obtained by the real-time dynamic differential positioning module on the UAV to obtain the camera's position and attitude in the world coordinate system, with position coordinates (x... c ,y c ,z c The coordinates are (12.5 m, 8.3 m, 4.2 m), with attitude angles of yaw 15 degrees, pitch -45 degrees, and roll 2 degrees. The transformation relationship between the base coordinate system and the world coordinate system of the trimming actuator is obtained through on-site hand-eye calibration, yielding the rotation matrix and translation vector of the robot arm's base coordinate system relative to the world coordinate system. Through inverse perspective projection transformation combined with coordinate system transformation, the two-dimensional coordinates of the image are mapped to the three-dimensional coordinates in the robot arm's base coordinate system. The mapping relationship is the homogeneous coordinate transformation equation λ·[u,v,1]. T =K·[R|t]·[X,Y,Z,1] T, where λ is the depth factor, K is the camera intrinsic parameter matrix, and R and t are the rotation matrix and translation vector from the camera coordinate system to the robot arm base coordinate system. Solving this equation yields the actual 3D trimming coordinates (X,Y,Z). The pruning actuator is a 6-DOF articulated robotic arm, with each joint being a rotary joint. Standard DH parameters are used for modeling, as follows: Link 1: Length a1 = 0 mm, Torsion angle α1 = 90 degrees, Joint offset d1 = 250 mm, Joint angle θ1 is a variable; Link 2: Length a2 = 450 mm, Torsion angle α2 = 0 degrees, Joint offset d2 = 0 mm, Joint angle θ2 is a variable; Link 3: Length a3 = 300 mm, Torsion angle α3 = 0 degrees, Joint offset d3 = 0 mm, Joint angle θ3 is a variable; Link 4: Length a4 = 0 mm, Torsion angle α4 = 90 degrees, Joint offset d4 = 280 mm, Joint angle θ4 is a variable; Link 5: Length a5 = 0 mm, Torsion angle α5 = -90 degrees, Joint offset d5 = 0 mm, Joint angle θ5 is a variable; Link 6: Length a6 = 0 mm, Torsion angle α6 = 0 degrees, Joint offset d6 = 100 mm, Joint angle θ6 is a variable. The inverse kinematics solution employs an algebraic analytical method, solving for θ1 to θ6 sequentially by simultaneously eliminating variables based on the trigonometric relationships between the elements of the end-effector pose matrix and the joint angles. After obtaining the target rotation angles of each joint, the drive signal unit extracts the branch edge contours within the neighborhood based on the corresponding coordinates of the position information in the visible light image. The neighborhood window size is 31×31 pixels centered at these coordinates, and the same Canny operator as described above is used to extract edges within this window. Using the extracted branch edge contours as collision constraint boundaries, the end effector of the pruning actuator must not enter the internal region enclosed by these contours during its movement. Starting from the current joint angle of the pruning actuator and ending at the target rotation angle, a fast search random tree algorithm is used to plan a collision-free trajectory in the joint space.

[0038] The search step size of the fast search random tree algorithm is set according to a preset segmented mapping table based on the pixel percentage of the branch edge contours within the neighborhood. The pixel percentage is defined as the number of edge pixels within the neighborhood window divided by the total number of pixels in the window (961). The preset segmented mapping table is as follows: 0.10 radians for a pixel percentage less than or equal to 10%; 0.07 radians for a pixel percentage greater than 10% and less than or equal to 30%; 0.05 radians for a pixel percentage greater than 30% and less than or equal to 50%; and 0.03 radians for a pixel percentage greater than 50%. The determination of the above segmented step size is based on the following: In open areas where the pixel ratio is less than or equal to 10%, a single step of 0.10 radians corresponds to an end-effector linear displacement of approximately 45 millimeters, which allows for relatively fast convergence when there is no risk of collision; in high-density twig areas where the pixel ratio is greater than 50%, a single step of 0.03 radians corresponds to an end-effector linear displacement of approximately 13.5 millimeters, which is sufficient for precise obstacle avoidance around complex contours. The pixel ratio is inversely proportional to the search step size. After trajectory planning is completed, the drive signal unit converts the angular trajectory of each joint into a time-sequential digital drive pulse signal and outputs it to the pruning actuator. The reference frequency of the pulse signal is 1 kHz, and the electronic gear ratio of each joint servo driver is 0.001 radians per pulse.

[0039] After the pruning execution mechanism completes a single pruning, the data processing center controls the image acquisition terminal to acquire visible light and near-infrared images of the pruned section at a preset close-range imaging focal length. The preset close-range imaging focal length is equivalent to 50 mm in 35 mm focal length. At this time, the drone hovering height is adjusted to 0.6 meters above the pruned section, and the image spatial resolution is approximately 0.5 mm per pixel. The spatial distribution of near-infrared reflectance of the pruned section is correlated and matched with a pre-stored winter shoot xylem spectral feature template. The method for constructing the winter shoot xylem spectral feature template is as follows: 50 manually confirmed winter shoot xylem samples are collected. Under standard light source conditions (color temperature 5500 Kelvin, illuminance 8000 lux), the reflectance curves of each sample in the 800 nm to 900 nm band are measured using an ASD FieldSpec 4 ground object spectrometer. Sampling is performed at wavelength intervals of 1 nm. The arithmetic mean of the 50 reflectance curves is taken wavelength by wavelength to form the template vector T(λ), where the nominal reflectance value at 840 nm wavelength is 0.28. The correlation matching uses the Pearson correlation coefficient. , where R i T represents the reflectance value of the cross-sectional image in the corresponding band. i The template reflectance value, μ R and μ TThese are the mean values ​​of cross-sectional reflectance and template reflectance, respectively, calculated over a range of 800 nm to 900 nm across 101 wavelength sampling points. When the calculated Pearson correlation coefficient exceeds the preset matching threshold of 0.75, it is determined that the winter shoots have not been completely removed, and a supplementary cutting signal is output to the drive signal unit to drive the pruning actuator to perform secondary pruning on the remaining winter shoots. The preset matching threshold of 0.75 was determined as follows: Template matching experiments were conducted on 50 known completely removed sections and 50 known sections with remaining winter shoots. The mean correlation coefficient between the completely removed sections and the template was 0.48 with a standard deviation of 0.11, while the mean correlation coefficient between the sections with remaining winter shoots and the template was 0.85 with a standard deviation of 0.07. The median of the mean correlation coefficients of the two classes, (0.48 + 0.85) / 2, equaled 0.665. This was then shifted by one standard deviation (0.07) towards the remaining winter shoot side, resulting in 0.735. 0.75 was chosen as the threshold, at which the classification sensitivity was 92% and the specificity was 96%. The feed depth for the secondary pruning was set to 5 mm to ensure complete removal of the xylem tissue from the winter shoots. After the secondary pruning was completed, the system continued to execute the above processing steps on the next frame image until all winter shoots were pruned.

[0040] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A litchi winter shoot detection and pruning control system based on digital signal processing, comprising an image acquisition terminal and a data processing center, characterized in that: The image acquisition terminal is configured to simultaneously acquire raw multispectral image digital signals of the litchi tree canopy in the visible light and near-infrared bands and output them to the data processing center. The data processing center includes a signal preprocessing unit, a feature extraction unit, a winter shoot discrimination unit, and a driving signal unit connected in sequence. The signal preprocessing unit is configured to perform radiometric calibration on the original multispectral image digital signal to map it into a reflectance value, construct a time-series vegetation index based on the reflectance of the visible light band and the reflectance of the near-infrared band, perform time-domain low-pass filtering on the time-series vegetation index to suppress high-frequency noise of illumination, and output a standardized vegetation index spatiotemporal data field. The feature extraction unit is configured to extract a multidimensional feature vector of the branch tip region based on the standardized vegetation index spatiotemporal data field and visible light image signal. The multidimensional feature vector includes at least relative growth activity features and geometric and spectral coupling features. The winter shoot discrimination unit is configured to calculate the distance between the multidimensional feature vector and the pre-stored winter shoot cluster center and non-winter shoot cluster center in the feature space. If the distance to the winter shoot cluster center is less than the distance to the non-winter shoot cluster center and the difference exceeds the preset offset, it is determined to be a winter shoot and its position information is output. The drive signal unit is configured to obtain the actual three-dimensional trimming coordinates based on the position information through a preset coordinate mapping relationship, and use the actual three-dimensional trimming coordinates as the input of the inverse kinematics solution to solve the target rotation angle of each joint of the trimming actuator, and convert each target rotation angle into a time-sequential digital drive pulse signal and output it to the trimming actuator.

2. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 1, characterized in that: The relative growth activity feature is extracted as follows: feature point registration is performed on adjacent frames; with the branch area of ​​the current frame after registration as the center, pixels whose vegetation index temporal change rate is less than a preset stability threshold and do not belong to the winter shoot candidate area are selected as background reference pixels within a preset spatial neighborhood window; the mean of the first-order temporal difference of the background reference pixels between adjacent frames is calculated as the whole tree baseline drift; the whole tree baseline drift is subtracted from the first-order temporal difference of the branch area to obtain the relative growth activity feature.

3. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 1, characterized in that: The geometric and spectral coupling features are extracted as follows: edge detection is performed on the branch area in the visible light image signal to obtain a gradient amplitude map; based on the near-infrared reflectance of the branch area in the standardized vegetation index spatiotemporal data field, the difference between the near-infrared reflectance of each pixel and the preset reference reflectance is calculated as a relative canopy thickness index, and the gradient amplitude map is weighted with the relative canopy thickness index as a spatial weight; the ratio of the gradient energy of the weighted gradient amplitude in the main axis direction and the secondary axis direction of the branch is calculated, and combined with the aspect ratio of the minimum bounding rectangle of the branch area, the geometric and spectral coupling features are constituted.

4. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 1, characterized in that: The signal preprocessing unit is further configured to: before constructing the temporal vegetation index, generate a pixel-level band ratio map based on the reflectance of the visible light band and the reflectance of the near-infrared band in the current frame, and use the pixel-level band ratio map as a weight to perform gain compensation on the reflectance of the visible light band in order to suppress pseudo-changes in reflectance caused by spatial non-uniformity of illumination; the temporal low-pass filter is applied to the temporal vegetation index after gain compensation.

5. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 1, characterized in that: The multidimensional feature vector is obtained as follows: The feature extraction unit, based on the standardized vegetation index spatiotemporal data field, uses the temporal stability of near-infrared reflectance of each pixel in the current frame as a screening criterion. Within a consecutive N-frame time window, it calculates the variance of the near-infrared reflectance of each pixel, marks pixels with variances less than a preset fluctuation threshold as stable candidate pixels, and extracts the relative growth activity feature and geometric-spectral coupling feature only for the stable candidate pixels. During the extraction of the geometric-spectral coupling feature, the feature extraction unit compares the temporal mean of the near-infrared reflectance of each stable candidate pixel with a preset reference reflectance value. The difference in the rate is used to construct a steady-state canopy thickness index. Using the steady-state canopy thickness index as a global weight, the visible light gradient amplitude of each pixel in the stable candidate pixel region is spatially weighted to obtain a weighted gradient field. The feature extraction unit further calculates the ratio of the gradient energy in the main axis direction of the weighted gradient field to the gradient energy in the secondary axis direction, and multiplies the ratio by the temporal variation coefficient of the near-infrared reflectance of each stable candidate pixel. The product is used as the spatial distribution adjustment factor of the geometric and spectral coupling feature. The adjustment factor is used to correct the final value of the geometric and spectral coupling feature to form an initial-order multidimensional feature vector.

6. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 5, characterized in that: Multidimensional feature vectors also include chromaticity distribution features; The feature extraction unit divides the stable candidate pixel set into a highlight stable subset, a shadow stable subset, and a transitional stable subset based on the temporal mean of the near-infrared reflectance of each pixel. It then extracts the hue mean of each of the three stable subsets in the HSV color space. Using the hue mean of the transitional stable subset as a benchmark, it normalizes the ratio of the hue mean of the highlight stable subset to the hue mean of the shadow stable subset. The normalized hue mean of the three subsets is then merged as the base value of the chromaticity distribution feature. The base value is then weighted using the reciprocal of the temporal variation coefficient of the near-infrared reflectance of each stable candidate pixel as a chromaticity confidence weight. The weighted chromaticity distribution feature is then incorporated into the initial-order multidimensional feature vector to form a complete multidimensional feature vector.

7. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 1, characterized in that: The winter shoot discrimination unit is further configured to: after the winter shoot discrimination of the current frame is completed, use the multi-dimensional feature vectors corresponding to the regions determined to be winter shoots and the regions determined to be non-winter shoots as incremental samples, and update the pre-stored winter shoot cluster centers and non-winter shoot cluster centers respectively using an exponential weighted moving average method; The forgetting factor adopts an exponential decay function λ(Δn)=λ0·exp(−k·Δn), where λ0 is the initial forgetting factor, k is the decay coefficient, and Δn is the difference in frame number between the current frame and the initial frame.

8. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 1, characterized in that: The winter shoot discrimination unit is further configured to: calculate the inter-class scatter matrix trace of the two types of regions in the feature space based on the multi-dimensional feature vectors of the winter shoot region and the non-winter shoot region determined in the previous frame; when the inter-class scatter matrix trace is less than a preset stability threshold, increase the preset bias by a preset first fixed step size; when it is greater than or equal to the preset stability threshold, decrease the preset bias by a preset second fixed step size.

9. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 1, characterized in that: The drive signal unit is further configured to: extract the branch edge contour in the neighborhood according to the corresponding coordinates in the visible light image based on the position information before solving the target rotation angle of each joint of the pruning actuator; use the branch edge contour as the collision constraint boundary, take the current joint angle of the pruning actuator as the starting point and the target rotation angle as the ending point, and use a fast search random tree algorithm to plan a collision-free trajectory in the joint space; the search step size of the fast search random tree algorithm is set according to the pixel ratio of the branch edge contour in the neighborhood according to a preset segmented mapping table.

10. The litchi winter shoot detection and pruning control system based on digital signal processing according to claim 1, characterized in that: The data processing center includes: after the pruning execution mechanism completes the pruning, it controls the image acquisition terminal to acquire visible light and near-infrared images of the pruning section at a preset close-range imaging focal length; it performs correlation matching between the spatial distribution of the near-infrared reflectance of the pruning section and a pre-stored spectral feature template of the xylem of winter shoots; when the matching degree is greater than a preset matching threshold, it determines that the winter shoots have not been completely removed, outputs a supplementary cutting signal to the drive signal unit, and drives the pruning execution mechanism to perform secondary pruning on the remaining winter shoots.

Images