A multi-target tracking based method for sequence planning of orange picking by unmanned aerial vehicle

Abstract

The application discloses a kind of unmanned plane citrus picking sequence planning methods based on multi-target tracking, comprising the following steps: S1, acquisition and processing citrus image dataset;S2, construct lightweight YOLOv8-Clementine target detection model, and carry out target detection to citrus;S3, based on lightweight YOLOv8-Clementine target detection model, adopts BotSort multi-target tracking algorithm to carry out real-time dynamic tracking to citrus;S4, adopts Corner-Kmeans clustering algorithm to divide the region of citrus tree;S5, picking sequence planning based on global strategy optimization.The application is aimed at real orchard complex scene, based on the efficient optimization of picking sequence of citrus visual perception and tracking model, to make the optimal picking order for picking unmanned plane, improve the safety and efficiency of picking.

Description

Technical Field

[0001] This invention belongs to the field of UAV visual perception technology, specifically relating to a UAV citrus harvesting sequence planning method based on multi-target tracking. Background Technology

[0002] Currently, citrus harvesting mainly relies on manual labor. Although researchers have developed fruit-picking robots, robots for picking individual fruits from tall citrus trees still require further research and development. Therefore, harvesting drones with flexible operating ranges have attracted the attention of researchers.

[0003] However, drones have limited computing power and battery life, making it impossible to spend significant time processing complex network models to obtain citrus location information. Furthermore, harvesting multiple clusters of single fruits requires designing a reasonable harvesting sequence plan based on fruit distribution characteristics to achieve precise and efficient citrus harvesting by drones. Harvesting drones cannot successfully complete citrus harvesting tasks relying solely on visual perception models, and traditional object detection algorithms have limited ability to describe the appearance changes and occlusion features of individual citrus fruits. In addition, existing object detection algorithms have limitations such as easily losing features of small targets and lacking multi-target task sequence planning mechanisms. Therefore, how to achieve high-precision visual perception and efficient sequence planning for multiple targets in complex environments has become a key technical challenge that urgently needs to be solved. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and to propose a UAV citrus harvesting sequence planning method based on multi-target tracking.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for planning a citrus harvesting sequence using a drone based on multi-target tracking includes the following steps:

[0007] S1. Collect and process citrus image datasets;

[0008] S2. Construct a lightweight YOLOv8-Clementine object detection model to perform object detection on citrus fruits;

[0009] S3. Based on the lightweight YOLOv8-Clementine target detection model, the BotSort multi-target tracking algorithm is used to perform real-time dynamic tracking of citrus fruits.

[0010] S4. Use the Corner-Kmeans clustering algorithm to divide citrus trees into regions;

[0011] S5. Harvesting sequence planning based on global strategy optimization.

[0012] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0013] 1. This invention achieves high-precision and lightweight citrus detection using a target detection method based on an adaptive attention mechanism. This reduces computational resource consumption and extends the drone's flight time, ensuring that the harvesting drone can complete more harvesting tasks with limited battery power. The method also utilizes a multi-target real-time tracking method based on global motion estimation to achieve real-time dynamic tracking of citrus fruits, employs dynamic clustering to achieve regional statistical division, and finally achieves scientific and reasonable harvesting sequence planning based on global strategy optimization. This method can determine the optimal harvesting order for the harvesting drone based on information such as the location and quantity of individual citrus fruits, reducing interference from the drone to other unharvested fruits during the harvesting process and improving the safety and efficiency of harvesting. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention.

[0015] Figure 2 This is a schematic diagram of the overall structure of the lightweight YOLOv8-Clementine object detection model.

[0016] Figure 3 This is a schematic diagram of the Adaptive Attention (AMA) mechanism.

[0017] Figure 4 This is a partial structural diagram of the DenseAMA module.

[0018] Figure 5 This is a schematic diagram of the VOVGSCSPC module.

[0019] Figure 6 This is a schematic diagram of the Inner-SIoU loss function.

[0020] Figure 7 This is a flowchart illustrating the harvest sequence planning process. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0022] Examples; such as Figure 1 As shown, a method for planning a citrus harvesting sequence using a drone based on multi-target tracking includes the following steps:

[0023] S1. Collect and process the citrus image dataset; specifically:

[0024] Citrus images were acquired using an RGB-D camera to obtain a citrus image dataset. This dataset includes images taken under various lighting conditions to ensure that the acquired images are not affected by artificial shadows or lighting.

[0025] The collected citrus images were manually screened and reviewed, selecting images with high clarity and rich details, and removing images with poor quality, severe exposure, excessive blurriness, or a large number of duplicates or images containing only one bunch of fruit.

[0026] The X-AnyLabelImg image annotation tool was used to manually annotate and define classification labels for citrus images. The image annotation information of individual citrus fruits was saved in JSON format and converted into corresponding txt files.

[0027] The labeled image dataset is divided into training, testing, and validation sets according to the specified proportions.

[0028] In this embodiment, the drone used is the DJI M300 RTK drone, which integrates an Intel RealSense D435i depth sensing camera, an NVIDIA Jetson TX3 NX processor, a Livox Avia LiDAR, and an integrated clipper actuator.

[0029] This embodiment uses an Intel RealSense D435i RGB-D camera to collect images from citrus bunches at distances of 400-700mm (close-up) and 700-1100mm (distant view) as a dataset. The dataset includes images taken under various lighting conditions, such as full sunlight and backlighting. The 1084 images of individual citrus fruits in complex orchard backgrounds were manually screened and reviewed.

[0030] The X-AnyLabelImg image annotation tool was used to manually annotate and define classification labels for 1084 images. The image annotation information of individual citrus fruits was saved in JSON format and converted into corresponding TXT files. The annotated dataset was divided into training, testing, and validation sets in a ratio of 8:1:1, resulting in 867 training images, 109 testing images, and 109 validation images.

[0031] S2. Construct a lightweight YOLOv8-Clementine object detection model to perform object detection on citrus fruits;

[0032] The lightweight YOLOv8-Clementine object detection model is as follows:

[0033] A lightweight YOLOv8s-Clementine model is built based on the YOLOv8 model. The model includes an adaptive attention mechanism (AMA). Figure 2 As shown, it consists of three parts: the backbone network, the neck network, and the head network.

[0034] The backbone network includes the DenseAMA module, the C2f-Faster-AMA module, and the SPPF module. The DenseAMA module replaces the first Conv and C2f combination module of the backbone network in the YOLOv8 model with a dense connection structure, and the C2f-Faster-AMA module replaces the C2f module of the backbone network in the YOLOv8 model.

[0035] The Neck network serves as a pyramid-shaped multi-scale feature fusion structure for the model. The VOVGSCSPC module, a cross-stage local network structure designed using the aggregation method, replaces the C2f module in the Neck part of the YOLOv8 model.

[0036] The Head network serves as the detection output of the model, and the Inner-SIoU loss function is used as the loss function for the bounding box of a single citrus fruit.

[0037] In this embodiment, all models have 100 epochs, an initial learning rate of 0.01, a termination learning rate of 1e-4, and all images are 848×480. The mean precision (mAP), number of model parameters (Params), total floating-point operations (GFLOPs), and frames per second (FPS) are used as evaluation metrics for model performance.

[0038] like Figure 3 As shown, the adaptive attention mechanism (AMA) is as follows:

[0039] The original feature vector X of the citrus target with dimensions C×H×W is used as input; C is the channel dimension, H is the feature height, and W is the feature width. The feature vector X contains complex background noise in the orchard, interference from overlapping fruits, and redundant information. The processing steps of the adaptive attention mechanism (AMA) include:

[0040] First, global max pooling (GMP) and global average pooling (GAP) are simultaneously performed on the input features to extract the most prominent citrus features and the mean representation of the overall features. Then, the kernel size K is calculated based on the channel dimension C, and a 1×1 convolutional layer is used to replace the traditional fully connected layer. The 1×1 convolution operation is performed on each channel feature with the K value to generate intermediate features of 1×1×C dimension. Next, the intermediate features are processed by the H-swish activation function and 1×1 convolution to generate channel weight coefficients, which are then fused to obtain the final attention weight. This attention weight is then weighted and operated on channel by channel with the original input feature vector X to enhance the effective target features of citrus and suppress negative features such as background and occlusion. Finally, the output dimension is still C×H×W, and the citrus target feature vector is recalibrated. This feature has stronger discriminative power and can accurately support the planning of citrus picking sequence in multi-target tracking scenarios.

[0041]

[0042]

[0043] Where, parameters a=2, b=1, =4.

[0044] like Figure 4 As shown, the DenseAMA module is as follows:

[0045] Based on the original DenseNet121 network, the DenseAMA module is obtained by adding the H-swish activation function and the adaptive attention mechanism (AMA) to the DenseBlock and Transition layers. The DenseAMA module consists of three stages, where the first and second stages form the DenseLayerAMA layer and the third stage forms the TransitAMA layer.

[0046] The C2f-Faster-AMA module, specifically:

[0047] The Bottleneck module in the C2f module of the YOLOv8 model is replaced with Faster-Block. Faster-Block consists of PConv and Conv modules. The PConv module only uses Conv to extract citrus single fruit features on some input feature channels, so as to significantly reduce the number of model parameters and computation.

[0048] like Figure 5 As shown, the VOVGSCSPC module is specifically as follows:

[0049] First, a 1×1 convolutional layer is used to extract features from the input citrus single-fruit feature map with 1 channel. Then, the feature extraction result is input into the GS Bottleneck module. The result processed by the GS Bottleneck module is concatenated with the result of the 1×1 convolution of the input citrus single-fruit feature map with 1 channel, and then a 1×1 convolution is performed again to obtain an output with 2 channels.

[0050] GS Bottleneck is a lightweight residual module that processes input features in two ways: the main branch uses two Ghost convolutions to compress and restore channels, while the branch uses depthwise convolutions to directly extract features. The two paths are added together to output the final features.

[0051] The output of the GS Bottleneck module is:

[0052]

[0053] The output of the VOVGSCSPC module is:

[0054]

[0055] in, These are the input features after channel compression. It is the number of input features. It is a transformation function. It's channel splicing. This indicates grouped channel-wise convolution.

[0056] like Figure 6 As shown, the Inner-SIoU loss function is as follows:

[0057] In Inner-SIoU, auxiliary bounding boxes are used to calculate the loss to accelerate the bounding box regression process, and a scale factor is introduced to control the scale of the auxiliary bounding boxes. Inner-SIoU is expressed as:

[0058] ;

[0059] in, Represents the Inner-SIoU loss function; This is the overlap height of the two inner frames. inter is the area of ​​the overlapping part of the two inner frames, and union is the total area covered by the two inner frames; For distance loss, γ is a hyperparameter that controls the weights of the distance loss. The difference between the x / y coordinates of the center points of the two boxes; For shape loss, , These represent the predicted width and height of the bounding box, respectively. Hyperparameters are used to control the nonlinearity of shape loss.

[0060] S3. Based on the lightweight YOLOv8-Clementine target detection model, the BotSort multi-target tracking algorithm is used to perform real-time dynamic tracking of citrus fruits; specifically:

[0061] The input video frames are processed frame by frame using a lightweight YOLOv8-Clementine object detection model. The shape features of individual citrus fruits are extracted using a YOLOv8 deep convolutional neural network, and the target bounding box set is output.

[0062]

[0063] in, Contains the center coordinates of the target citrus fruit. width high and detection confidence level ;

[0064] Simultaneously, multi-scale appearance feature vectors are extracted using the adaptive attention mechanism (AMA):

[0065] ;

[0066] in, Represents the state transition matrix. This represents the appearance feature vector of the i-th target;

[0067] The Kalman filter is used to model the motion state of the target, and the target's state vector is defined as follows:

[0068] ;

[0069] in, The first derivative (velocity) represents the position coordinates of a single citrus fruit. Indicates the rate of change of dimensions;

[0070] The target state is predicted using a state transition equation, which is:

[0071] ;

[0072] in, It is the state transition matrix. It is a control input matrix. It is a control input;

[0073] The bounding box information detected by the lightweight YOLOv8-Clementine object detection model is mapped to the state space through observation equations:

[0074] ;

[0075] Where H is the observation matrix, It is a state estimate. It is observation noise;

[0076] Cosine similarity is used to measure appearance consistency, and feature vectors of two consecutive frames are calculated. and Similarity:

[0077] ;

[0078] For successfully matched trajectories (i.e., the paths along which the position of each citrus fruit target changes over time in consecutive frames of the video sequence), their state vectors and appearance models are updated to incorporate the latest detection information to adapt to changes in the target's appearance; for unmatched detection boxes, new trajectories are initialized to respond to newly emerging targets.

[0079] S4. The Corner-Kmeans clustering algorithm is used to divide the citrus fruit trees into regions; specifically:

[0080] The Corner-Kmeans clustering algorithm is used for citrus clustering partitioning, selecting the four corners of the target region as initial cluster centers to effectively address the sensitivity of traditional K-means algorithms to initial center selection. Firstly, by selecting the four corners of the target region as initial cluster centers, a more uniform distribution of cluster centers can be provided for the data in rectangular (or approximately rectangular) regions, avoiding the suboptimal clustering results caused by random initial center selection in traditional K-means. This makes the initial cluster centers more representative, thereby reducing the instability caused by random selection. The algorithm specifically includes the following steps:

[0081] S41. Input 3D coordinate data using the depth camera of the harvesting drone; extract the x, y, and z axis coordinates of the 3D coordinate data to form a 3D dataset. Each data point ;

[0082] S42. Select the four corner points of region R ( , , , As the initial cluster centers, ensure that the initial cluster centers cover all four directions of the entire region. , Let x and y represent the minimum and maximum values ​​of the x-axis and y-axis, respectively.

[0083] S43. For each data point x1 in the dataset, calculate its Euclidean distance to the four cluster centers. , As the current cluster center, For data points;

[0084] S44. For each cluster k, calculate the mean C of all data points within that cluster. k And use this mean as the new cluster center:

[0085]

[0086] in, It is the set of all data points belonging to cluster k. Indicates the size of the cluster;

[0087] S45. Repeat steps S43 and S44 until the cluster centers no longer change or the maximum number of iterations is reached, then the clustering partitioning ends.

[0088] S5. Harvesting sequence planning based on global strategy optimization;

[0089] The picking sequence planning uses the nearest neighbor algorithm, specifically:

[0090] After using the Corner-Kmeans clustering algorithm for clustering and partitioning, the citrus fruit tree area was divided into multiple sub-regions, and the starting point or picking point of each sub-region was identified.

[0091] Starting from the beginning of each sub-region, calculate the distance from the current point to all unpicked points, select the single fruit closest to the current point as the next picking target, update the current point to the target; mark the single citrus fruit as picked, and add it to the picking path;

[0092] After ensuring all fruit trees have been harvested, output the harvesting path for each sub-region, indicating the harvesting order, and finally send the harvesting path to the harvesting drone to execute the harvesting task. For example... Figure 7 The diagram shown is a schematic diagram of the harvest sequence planning.

[0093] Euclidean distance is used to measure the distance from the current point to all unpicked points;

[0094] Suppose that at a certain moment, the current location of the harvesting drone is... The set of individual citrus fruits to be harvested is as follows: The goal of the nearest neighbor algorithm is to select a cluster of target single fruit points. This makes it possible to start from the current position To the target point The distance is the smallest, that is .

[0095] The nearest neighbor algorithm makes decisions by minimizing a distance function, which is the path length. These are two adjacent target points on the path.

[0096] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0097] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for planning a citrus harvesting sequence using unmanned aerial vehicles (UAVs) based on multi-target tracking, characterized in that, Includes the following steps: S1. Collect and process citrus image datasets; S2. Construct a lightweight YOLOv8-Clementine object detection model to perform object detection on citrus fruits; S3. Based on the lightweight YOLOv8-Clementine target detection model, the BotSort multi-target tracking algorithm is used to perform real-time dynamic tracking of citrus fruits. S4. Use the Corner-Kmeans clustering algorithm to divide citrus trees into regions; S5. Harvesting sequence planning based on global strategy optimization.

2. The method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 1, characterized in that, Step S1 is as follows: Citrus images were acquired using an RGB-D camera to obtain a citrus image dataset. This dataset includes images taken under various lighting conditions to ensure that the acquired images are not affected by artificial shadows or lighting. The collected citrus images were manually screened and reviewed, selecting images with high clarity and rich details, and removing images with poor quality, severe exposure, excessive blurriness, or a large number of duplicates or images containing only one bunch of fruit. The X-AnyLabelImg image annotation tool was used to manually annotate and define classification labels for citrus images. The image annotation information of individual citrus fruits was saved in JSON format and converted into corresponding txt files. The labeled image dataset is divided into training, testing, and validation sets according to the specified proportions.

3. The method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 1, characterized in that, In step S2, the lightweight YOLOv8-Clementine object detection model is specifically as follows: A lightweight YOLOv8s-Clementine model was built based on the YOLOv8 model. The model has an adaptive attention mechanism (AMA) and consists of three parts: the backbone network, the neck network, and the head network. The backbone network includes the DenseAMA module, the C2f-Faster-AMA module, and the SPPF module. The DenseAMA module replaces the first Conv and C2f combination module in the YOLOv8 model's backbone network with a dense connection structure, and the C2f-Faster-AMA module replaces the C2f module in the YOLOv8 model's backbone network. The Neck network serves as a pyramid-shaped multi-scale feature fusion structure for the model. The VOVGSCSPC module, a cross-stage local network structure designed using the aggregation method, replaces the C2f module in the Neck part of the YOLOv8 model. The Head network serves as the detection output of the model, and the Inner-SIoU loss function is used as the loss function for the bounding box of a single citrus fruit.

4. The method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 3, characterized in that, The adaptive attention mechanism (AMA) is as follows: The input consists of the original feature vector X of the citrus target with dimensions C×H×W; C is the channel dimension, H is the feature height, and W is the feature width. X contains complex background noise in the orchard, interference from overlapping and occlusion of fruits, and redundant information. The processing steps of the adaptive attention mechanism (AMA) include: First, global max pooling (GMP) and global average pooling (GAP) are simultaneously performed on the input features to extract the most prominent citrus features and the mean representation of the overall features. Then, the kernel size K is calculated based on the channel dimension C, and a 1×1 convolutional layer is used to replace the traditional fully connected layer. The 1×1 convolution operation is performed on each channel feature with the K value to generate intermediate features of 1×1×C dimension. The intermediate features are processed by the H-swish activation function and 1×1 convolution to generate channel weight coefficients, which are then fused to obtain the final attention weight. This attention weight is then weighted and operated on channel by channel with the original input feature vector X to enhance the effective target features of citrus and suppress background and occlusion. The final output is still a citrus target feature vector of C×H×W dimension, which has been recalibrated. This feature has stronger discriminative power and can accurately support the planning of citrus picking sequence in multi-target tracking scenarios.

5. The method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 3, characterized in that, The DenseAMA module, specifically: Based on the original DenseNet121 network, the DenseAMA module is obtained by adding the H-swish activation function and the adaptive attention mechanism (AMA) to the DenseBlock and Transition layers. The DenseAMA module consists of three stages, where the first and second stages form the DenseLayerAMA layer and the third stage forms the TransitAMA layer. The C2f-Faster-AMA module, specifically: The Bottleneck module in the C2f module of the YOLOv8 model is replaced with Faster-Block. Faster-Block consists of PConv and Conv modules. The PConv module only uses Conv to extract citrus single fruit features on some input feature channels, so as to significantly reduce the number of model parameters and computation.

6. The method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 3, characterized in that, The VOVGSCSPC module is specifically as follows: First, a 1×1 convolutional layer is used to extract features from the input citrus single fruit feature map with 1 channel. Then, the feature extraction result is input into the GS Bottleneck module. The result processed by the GS Bottleneck module is concatenated with the result of the 1×1 convolution of the input citrus single fruit feature map with 1 channel, and then a 1×1 convolution is performed again to obtain an output with 2 channels. GS Bottleneck is a lightweight residual module that processes input features in two ways: the main branch uses two Ghost convolutions to compress and restore channels, while the branch uses depthwise convolutions to directly extract features. The two paths are added together to output the final features. The output of the GS Bottleneck module is: The output of the VOVGSCSPC module is: in, These are the input features after channel compression. It is the number of input features. It is a transformation function. It's channel splicing. This indicates grouped channel-wise convolution.

7. The method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 3, characterized in that, The Inner-SIoU loss function is as follows: In Inner-SIoU, auxiliary bounding boxes are used to calculate the loss to accelerate the bounding box regression process, and a scale factor is introduced to control the scale of the auxiliary bounding boxes. Inner-SIoU is expressed as: ； in, Represents the Inner-SIoU loss function; This is the overlap height of the two inner frames. inter is the area of ​​the overlapping part of the two inner frames, and union is the total area covered by the two inner frames; For distance loss, γ is a hyperparameter that controls the weights of the distance loss. The difference between the x / y coordinates of the center points of the two boxes; For shape loss, , These represent the predicted width and height of the bounding box, respectively. Hyperparameters are used to control the nonlinearity of shape loss.

8. A method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 1, characterized in that, Step S3 is as follows: The input video frames are processed frame by frame using a lightweight YOLOv8-Clementine object detection model. The shape features of individual citrus fruits are extracted using a YOLOv8 deep convolutional neural network, and the target bounding box set is output. in, Contains the center coordinates of the target citrus fruit. width high and detection confidence level ; Simultaneously, multi-scale appearance feature vectors are extracted using the adaptive attention mechanism (AMA): ； in, Represents the state transition matrix. This represents the appearance feature vector of the i-th target; The Kalman filter is used to model the motion state of the target, and the target's state vector is defined as follows: ； in, The first derivative representing the position coordinates of a single citrus fruit. Indicates the rate of change of dimensions; The target state is predicted using a state transition equation, which is: ； in, It is the state transition matrix. It is a control input matrix. It is a control input; The bounding box information detected by the lightweight YOLOv8-Clementine object detection model is mapped to the state space through observation equations: ； Where H is the observation matrix, It is a state estimate. It is observation noise; Cosine similarity is used to measure appearance consistency, and feature vectors of two consecutive frames are calculated. and Similarity: ； For successfully matched trajectories, i.e. the paths through which the position of each citrus fruit target changes over time in consecutive frames of the video sequence, their state vectors and appearance models are updated, incorporating the latest detection information to adapt to changes in the target's appearance; for unmatched detection boxes, new trajectories are initialized to respond to newly emerging targets.

9. A method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Input 3D coordinate data using the depth camera of the harvesting drone; extract the x, y, and z axis coordinates of the 3D coordinate data to form a 3D dataset. Each data point ; S42. Select the four corner points of region R. , , and As initial cluster centers, ensure that the initial cluster centers cover all four directions of the entire region. , Let x and y represent the minimum and maximum values ​​of the x-axis and y-axis, respectively. S43. For each data point x1 in the dataset, calculate its Euclidean distance to the four cluster centers. , As the current cluster center, For data points; S44. For each cluster k, calculate the mean C of all data points within that cluster. k And use this mean as the new cluster center: in, It is the set of all data points belonging to cluster k. Indicates the size of the cluster; S45. Repeat steps S43 and S44 until the cluster centers no longer change or the maximum number of iterations is reached, then the clustering partitioning ends.

10. A method for planning a UAV citrus harvesting sequence based on multi-target tracking according to claim 1, characterized in that, In step S5, the picking sequence planning uses the nearest neighbor algorithm, specifically as follows: After using the Corner-Kmeans clustering algorithm for clustering and partitioning, the citrus fruit tree area was divided into multiple sub-regions, and the starting point or picking point of each sub-region was identified. Starting from the beginning of each sub-region, calculate the distance from the current point to all unpicked points, select the single fruit closest to the current point as the next picking target, update the current point to the target; mark the single citrus fruit as picked, and add it to the picking path; After ensuring that all fruit trees have been harvested, output the harvesting path for each sub-region, indicating the harvesting order, and finally send the harvesting path to the harvesting drone to execute the harvesting task.

Images