Fruit stem shearing and picking robot and picking method thereof

Abstract

The invention relates to the field of intelligent agricultural equipment, and discloses a fruit stem shearing and picking robot and a picking method thereof.The fruit stem shearing and picking robot comprises a movable chassis, an operation platform, a control device, an execution device, a visual navigation device and a recognition and positioning device; the working platform is mounted on the movable chassis; the control device, the execution device and the visual navigation device are all installed on the work platform. The visual navigation device is used for environment perception, path planning, coarse positioning of fruits and identification of fruit dense areas; the identifying and positioning device is mounted on the executing device; the execution device is used for cleaning sheltering objects and cutting fruit stems; and the control device controls the mobile chassis, the execution device, the visual navigation device and the identification positioning device to work cooperatively. The sheltering problem can be effectively solved, and the picking speed and the success rate are increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent agricultural equipment, in particular to a fruit stem shearing and picking robot and a picking method thereof. BACKGROUND

[0002] With the development of agricultural intelligence and automation technology, fruit picking robots have gradually become an important tool to improve the efficiency of agricultural production. In particular, in the picking operation of economic fruits such as apples, peaches, and pears, the traditional manual picking method is low in efficiency and high in labor intensity, which is difficult to meet the high production demand of modern orchards.

[0003] In a natural growing environment, an orchard has a high degree of unstructured and uncertainty. Fruits are usually partially or completely obscured by leaves, other fruits, or supporting objects, and changes in light and weather conditions make it difficult for the vision system to accurately identify the location of the target fruit and fruit stem. In the prior art, most robots rely solely on the vision system for positioning, and when there is an obstruction, they cannot effectively adjust the operation path or avoid obstacles, resulting in high failure rate of picking, increased fruit damage, and even damage to the plant. Although some technologies attempt to introduce auxiliary devices such as brushes and air nozzles, these devices often lack coordinated control with the vision system and are difficult to adapt to complex obscured environments, and the obscured handling capability is still insufficient. In addition, most existing picking robots use a single precise positioning strategy, i.e., high-precision image processing and coordinate calculation are performed before each picking, which ensures the picking accuracy but significantly prolongs the cycle time of single picking, making it difficult to meet the demand of large-scale orchard high-efficiency operation.

[0004] Therefore, it is urgent to propose a fruit stem shearing and picking robot and a picking method thereof to solve the above problems. SUMMARY

[0005] The purpose of the present application is to propose a fruit stem shearing and picking robot and a picking method thereof, which can effectively handle the obstruction problem and improve the picking speed and success rate.

[0006] To solve the above technical problems, the present application provides a fruit stem shearing and picking robot, comprising a mobile chassis, a work platform, a control device, an execution device, a vision navigation device, and an identification and positioning device. The work platform is installed on the mobile chassis; the control device, the execution device, and the vision navigation device are all installed on the work platform. The vision navigation device is used for environment perception, path planning, and coarse positioning of fruits, and identification of fruit dense areas. The identification and positioning device is installed on the execution device and is used for precise positioning of fruits and fruit stems to obtain the spatial position information of the fruit to be picked and the fruit stem. The execution device is used for clearing obstructions and shearing fruit stems. The control device is used for judging the shielding degree according to the spatial position information, controlling the execution device to adopt a corresponding picking strategy according to different shielding degrees, and controlling the mobile chassis, the execution device, the visual navigation device and the identification and positioning device to work cooperatively.

[0007] Further, the execution device comprises a mechanical arm, a servo motor, a first shell, a second shell, a rotating ring and an executor. The mechanical arm is installed on the working platform and connected with the first shell at the tail end. The servo motor is installed on the first shell and connected with an external gear at the output shaft. The rotating ring is arranged around the outer periphery of the second shell and fixedly connected with the second shell, and the inner side of the rotating ring is provided with an internal gear; the second shell is rotationally installed at the end of the first shell away from the mechanical arm; the second shell is provided with a groove, the output shaft of the servo motor passes through the groove so that the external gear can engage with the internal gear; and the executor is installed on the second shell.

[0008] Further, the executor comprises a mounting frame, a driving assembly, a pair of moving knife tables and a pair of movable blocking rods. The mounting frame is installed on the second shell; the driving assembly is installed inside the mounting frame and connected with one end of the moving knife table to drive a pair of the moving knife tables to move towards or away from each other; each moving knife table is provided with a blade; and each movable blocking rod is connected with the corresponding moving knife table through a sliding connection mode, and the movable blocking rod can move outward synchronously with the moving away of the moving knife table.

[0009] Further, the collecting device is installed on the working platform and located below the executor for receiving and buffering the picked fruits. The collecting device comprises an upper ring cover, a fruit bag, a buffer net and a collecting frame; the outer wall of the rotating ring is provided with an annular sliding rail; the upper ring cover is slidingly connected with the annular sliding rail and can slide along the annular sliding rail; under the action of gravity, the upper ring cover always slides to the lowest position of the rotating ring; one end of the fruit bag is connected with the upper ring cover, and the other end of the fruit bag is communicated with the collecting frame; the collecting frame is installed on the working platform; and the fruit bag is internally provided with a plurality of buffer nets.

[0010] In addition, the present application also proposes a fruit stem shearing picking method using the fruit stem shearing picking robot as described above, which specifically comprises the following steps. Collecting working environment images, planning a motion path and controlling the fruit stem shearing picking robot to move to a target picking area; The target harvesting area is roughly located and divided into regions to determine the priority harvesting areas; The fruit stem cutting and picking robot is controlled to move to the target area, obtain the spatial position information of the fruit and fruit stem, and judge the degree of occlusion of the fruit stem of each effective picking target based on the spatial position information to generate a picking sequence. Fruits in the target area are harvested according to the priority order of the harvesting sequence, and corresponding harvesting strategies are adopted for different degrees of shading.

[0011] Furthermore, the process of acquiring images of the working environment, planning the motion path, and controlling the fruit stem cutting and harvesting robot to move to the target harvesting area specifically includes: The visual navigation device collects images of the working environment and transmits them to the control device. The control device corrects the errors in the images, fills in missing information, integrates multi-source data to generate a continuous global environment map, and collects dynamic and static obstacle information in real time, updates the global environment map, and determines the optimal movement trajectory. Subsequently, the mobile chassis moves to the target picking area according to the optimal trajectory.

[0012] Furthermore, the step of roughly locating and dividing the target harvesting area to determine the priority harvesting area specifically includes: The coarse localization of the fruit uses a lightweight target detection network, which extracts multi-scale fruit features with a lightweight backbone network to achieve rapid detection and region segmentation. Based on the size parameters of the working environment image, the detection area is uniformly divided into multiple equal-width regions horizontally and multiple equal-height regions vertically, forming multiple rectangular detection sub-regions. Simultaneously, all detected fruit bounding boxes are traversed, the center point coordinates of each fruit bounding box are calculated, and the detection sub-region to which it belongs is determined based on the center point coordinates. The number of fruits in each sub-region is accumulated and statistically analyzed to generate a fruit density distribution matrix. Based on the density distribution matrix, the number of fruits in each region is sorted and compared, and the region with the most fruits is automatically identified and selected as the priority target region for harvesting.

[0013] Furthermore, the step of acquiring the spatial location information of the fruit and its stem, and judging the degree of occlusion of the stem of each valid harvesting target based on the spatial location information to generate a harvesting sequence, specifically includes: The identification and positioning device collects target fruit images and depth images in the target area and transmits them to the control device. The control device uses a multimodal fusion detection network to fuse the spatial location information of the depth image with the texture features of the visible light image through a feature complementary mapping mechanism, and strengthens the modeling of the spatial relationship between the fruit and the stem to locate their two-dimensional coordinates and three-dimensional spatial positions. Then, the detection results are verified for completeness, and effective target pairs that simultaneously meet the confidence thresholds of the fruit and stem and have a reasonable spatial distance are selected. Next, the depth gradient features of the area in front of the stem are analyzed, and the targets are divided into three levels: no occlusion, slight occlusion, and severe occlusion, based on the occlusion score. Finally, the harvesting priority score of each target is calculated by combining the fruit confidence, stem confidence, fruit size, and occlusion level, and the optimal harvesting sequence from no occlusion to severe occlusion is dynamically generated to guide the execution device to perform the harvesting operation in priority order.

[0014] Furthermore, the semantic features are first processed through channel interaction to generate channel weights for enhancing spatial features, and each channel is processed independently using depthwise separable convolution. Spatial features are spatially interacted to generate spatial weights for enhancing semantic features, and spatial information is aggregated to generate a spatial attention map. After obtaining the channel weights and spatial weights, the final features are generated through bidirectional weighting and fusion. The final features contain both semantic information and spatial location information, and the detection results are output. Based on the detection results, integrity verification is performed on each detected target to ensure that there are matching pairs of fruit and stem simultaneously. Specifically, this includes: calculating the fruit-stem matching degree; judging target integrity; filtering targets to obtain valid target pairs; for the filtered valid target pairs, extracting the region of interest in front of the stem and calculating the depth gradient discontinuity and depth variance features within the region, generating an occlusion score through weighted fusion, and classifying the occlusion state into multiple occlusion levels according to a preset threshold; calculating the picking priority score for each target fruit based on the occlusion level, combined with fruit detection confidence, fruit size, and spatial location information; and sorting all target fruits from high to low according to the picking priority score to generate the optimal picking sequence.

[0015] Furthermore, the step of harvesting fruits in the target area according to the priority order of the harvesting sequence, and adopting corresponding harvesting strategies for different degrees of shading, specifically includes: Based on the occlusion level and analysis results, differentiated obstacle avoidance and removal strategies are adopted: Unobstructed fruit: Follow the standard picking path directly; Slightly obstructed fruit: Adjust the actuator's approach angle and use a trial-and-error contact strategy to clear the obstructions and pick the fruit; After each clearing operation, the identification and positioning device re-acquires images and assesses the obstruction level. If the obstruction level decreases, proceed with the picking process; if the obstruction level does not decrease, mark it as a target to be processed and skip it; Severely obstructed fruit: Mark it as a target to be processed, record its corresponding spatial location, and process it after the surrounding fruit has been picked or after intervention. The control device calculates the clamping force required based on the shape and material of the fruit stalk, and then cuts the fruit stalk. During the cutting process, the identification and positioning device detects the distance between itself and the target fruit in real time to ensure that the fruit stem is cut successfully. If a cutting failure is detected, the control device will terminate the remaining actions through a feedback mechanism and repeat the above steps until the harvest is successful. Then the fruit is moved to the collection box through the fruit bag. After harvesting the target fruit in the field of view of the identification and positioning device, repeat all the steps to continue harvesting the target fruit in the remaining area.

[0016] Through the above technical solution, the present invention has the following beneficial effects: By employing a visual navigation device for environmental perception, path planning, and coarse fruit positioning, and a positioning device for precise location of fruits and stems, the control device determines the degree of occlusion based on spatial location information and adopts corresponding harvesting strategies. This effectively improves the harvesting success rate and overall operational efficiency in complex orchard environments. Furthermore, by using an execution device to clear obstructions and cut fruit stems, the system's adaptability to different occlusion conditions is enhanced, reducing the rate of missed harvests.

[0017] Furthermore, a multimodal fusion detection network is used to fuse spatial location information from depth images with texture features from visible light images, enhancing the spatial relationship modeling between fruits and stems and improving target localization accuracy. By dividing the detection area into multiple sub-regions and calculating fruit density, priority is given to harvesting densely fruited areas. Combined with occlusion level judgment and harvesting priority ranking, an optimal harvesting sequence is dynamically generated, significantly improving harvesting efficiency. The rotation adjustment mechanism and movable baffle in the execution device clear obstructions, along with the buffer net and adaptively oriented fruit bags in the collection device, reduce fruit damage rate and improve harvesting quality. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall structure of the fruit stalk cutting and harvesting robot in one embodiment of the present invention; Figure 2 This is a partial structural diagram of a fruit stalk cutting and harvesting robot according to an embodiment of the present invention; Figure 3This is a partial three-dimensional structural diagram of a fruit stalk cutting and harvesting robot in one embodiment of the present invention; Figure 4 This is a schematic diagram of the rotating ring structure in a fruit stalk cutting and harvesting robot according to an embodiment of the present invention; Figure 5 This is a top view of the actuator in a fruit stem cutting and harvesting robot according to an embodiment of the present invention; Figure 6 This is a partial structural diagram of the actuator in a fruit stem cutting and harvesting robot according to an embodiment of the present invention; Figure 7 This is a top view of the collection device in a fruit stem cutting and harvesting robot according to an embodiment of the present invention; Figure 8 This is a flowchart of a fruit stalk cutting and harvesting method in one embodiment of the present invention; Figure 9 This is an overall flowchart of a fruit stalk cutting and harvesting method according to an embodiment of the present invention; Figure 10 This is a diagram of the deep learning network model structure for fine localization of fruit and fruit stalk in a fruit stalk cutting and picking method according to an embodiment of the present invention. Figure 11 This is a network structure diagram of the spatial semantic feature interaction module (SFFI) in the fruit stalk cutting and harvesting method of an embodiment of the present invention. Figure 12 This diagram illustrates the specific steps of adaptive harvesting based on fine positioning of the fruit and fruit stalk in a fruit stalk cutting and harvesting method according to an embodiment of the present invention.

[0019] In the diagram, 1. Mobile chassis; 2. Visual navigation device; 3. Working platform; 4. Robotic arm; 5. Identification and positioning device; 6. Actuator; 7. Control device; 8. Collection device; 9. Collection box; 601. Movable blade holder; 60101. Blade; 60102. Flexible clamping block; 602. Movable stop bar; 6021. Groove; 603. Drive assembly; 604. Mounting frame; 605. Servo motor; 606. Rotating ring; 607. First housing; 608. Second housing; 801. Upper ring cover; 802. Buffer net; 803. Fruit bag. Detailed Implementation

[0020] Based on the teachings of this specification, those skilled in the art can form new technical solutions through cross-combination of different implementation methods without creating technical contradictions. Such variations should all be considered to fall within the protection scope of this invention.

[0021] The following description, in conjunction with the accompanying drawings, provides a more detailed account of a fruit stalk cutting and harvesting robot and its harvesting method according to the present invention, which illustrates preferred embodiments of the invention. It should be understood that those skilled in the art can modify the invention described herein while still achieving its advantageous effects. Therefore, the following description should be understood as being of general knowledge to those skilled in the art and is not intended to limit the invention.

[0022] The invention is described more specifically by way of example in the following paragraphs with reference to the accompanying drawings. The advantages and features of the invention will become clearer from the following description. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the invention.

[0023] like Figures 1-3 As shown in the figure, an embodiment of the present invention proposes a fruit stem cutting and harvesting robot, including a mobile chassis 1, a working platform 3, a control device 7, an execution device, a visual navigation device 2, and an identification and positioning device 5.

[0024] Specifically, the working platform 3 is mounted on the mobile chassis 1; the control device 7, the execution device, and the visual navigation device 2 are all mounted on the working platform 3; the visual navigation device 2 is used for environmental perception, path planning, and coarse positioning of the fruit to identify areas with dense fruit; the identification and positioning device 5 is mounted on the execution device to accurately position the fruit and fruit stems and obtain spatial location information of the fruit and fruit stems to be picked; the execution device is used to clear obstructions and cut the fruit stems; the control device 7 is used to determine the degree of obstruction based on the spatial location information, control the execution device to adopt corresponding picking strategies according to different degrees of obstruction, and control the mobile chassis 1, the execution device, the visual navigation device 2, and the identification and positioning device 5 to work together.

[0025] In this embodiment, the mobile chassis 1 is a tracked chassis, which can adapt to complex orchard terrain and provide stable mobility. The working platform 3 also integrates a power supply system, a drive module, and a high-capacity battery pack, enabling the robot to operate continuously and stably in complex orchard environments.

[0026] In this embodiment, the control device 7 includes a control cabinet, which houses a lower-level control system responsible for real-time control and execution, and an upper-level control system responsible for decision analysis, path planning, and human-computer interaction. The lower-level control system receives various sensor information and feeds it back to the upper-level control system. After receiving the information and making a decision, the upper-level control system outputs control commands to the lower-level control system to control each component to perform corresponding actions.

[0027] In this embodiment, the visual navigation device 2 includes a navigation system base, a first camera, and a lidar. The navigation system base is fixed on the working platform 3, and the first camera and lidar are respectively installed above the navigation system base. The two work together to achieve 3D modeling of the orchard environment, obstacle detection, and path planning. The first camera also performs coarse positioning. When the navigation system controls the robot to reach the designated picking area, the first camera activates coarse positioning mode, using the lightweight target detection algorithm YOLO-IP to quickly identify areas with dense fruit, obtain their approximate coordinates, and send this position information to the execution device, guiding the execution device to move to the target working area (i.e., the target picking area mentioned below), providing initial positioning data for subsequent fine picking operations.

[0028] In this embodiment, the identification and positioning device 5 is a second camera installed on the execution device. After the execution device reaches the designated working area based on the coarse positioning result, it performs fine detection by introducing the YOLO-PP detection network optimized by feature extraction and attention mechanism to obtain the spatial location information of the fruit to be picked and the fruit stem, such as three-dimensional coordinate information.

[0029] In this embodiment, combined with Figures 2-4 As shown, the execution device includes a robotic arm 4, a servo motor 605, a first housing 607, a second housing 608, a rotating ring 606, and an actuator 6.

[0030] Specifically, the robotic arm 4 is mounted on the work platform 3, and its end is connected to the first housing 607; the servo motor 605 is mounted on the first housing 607, and its output shaft is connected to an external gear; the rotating ring 606 surrounds the second housing 608 and is fixedly connected to the second housing 608, and an internal gear is provided on the inner side of the rotating ring 606; the second housing 608 is rotatably mounted on the end of the first housing 607 away from the robotic arm 4; the second housing 608 has a groove, and the output shaft of the servo motor 605 passes through the groove so that the external gear can mesh with the internal gear; the actuator 6 is mounted on the second housing 608. In this embodiment, the servo motor 605 drives the meshing transmission of the external gear and the internal gear to realize the rotation adjustment of the rotating ring 606 and the second housing 608, thereby driving the actuator 6 to rotate as a whole and adjusting the working angle of the actuator 6.

[0031] Preferably, the robotic arm 4 is a multi-degree-of-freedom robotic arm 4.

[0032] In this embodiment, combined with Figures 2-6 As shown, the actuator 6 includes a mounting frame 604, a drive assembly 603, a pair of movable tool holders 601, and a pair of movable stop levers 602.

[0033] Specifically, the mounting frame 604 is mounted on the second housing 608; the drive assembly 603 is mounted inside the mounting frame 604 and connected to one end of the movable blade 601 to drive the pair of movable blades 601 to move towards or away from each other; each movable blade 601 is provided with a blade 60101; each movable stop bar 602 is connected to the corresponding movable blade 601 by a sliding connection, and the movable stop bar 602 can move outward synchronously with the opposite movement of the movable blades 601, so as to actively push away the obstructing branches and leaves and remove obstacles around the target fruit during the harvesting process.

[0034] Preferably, each of the movable blades 601 is further provided with a flexible clamping block 60102, which is located around the blade 60101 and is used to firmly clamp the fruit stem before cutting.

[0035] In this embodiment, the drive component 603 is preferably an electric gripper, which can precisely control the distance and clamping force between the two moving blades 601 according to the instructions of the control device 7, so as to achieve adaptive clamping and cutting of fruit stems of different thicknesses.

[0036] In this embodiment, each of the movable blade holders 601 is provided with a slot 6021. One end of the movable stop bar 602 is inserted into the slot 6021 and can slide within the slot 6021, allowing the movable stop bar 602 to extend and retract with the opening and closing movement of the movable blade holders 601. When the movable blade holders 601 move in opposite directions (open), the movable stop bar 602 extends outward; when the movable blade holders 601 move towards each other (close), the movable stop bar 602 retracts inward. This arrangement allows the movable stop bar 602 to both serve as a mechanical auxiliary cleaning device to remove obstructions and to avoid interfering with the cutting action during shearing.

[0037] Furthermore, such as Figure 1 , Figure 2 , Figure 3 and Figure 7 As shown, this embodiment also includes a collection device 8; the collection device 8 is installed on the working platform 3 and located below the actuator 6, and is used to receive and buffer the harvested fruit.

[0038] More specifically, the collecting device 8 includes an upper ring cover 801, a fruit bag 803, a buffer net 802, and a collecting frame 9. The outer wall of the rotating ring 606 is provided with an annular slide rail; the upper ring cover 801 is slidably connected to the annular slide rail and can slide along the annular slide rail. Under the action of gravity, the upper ring cover 801 always slides to the lowest position of the rotating ring 606; one end of the fruit bag 803 is connected to the upper ring cover 801, and the other end of the fruit bag 803 is connected to the collecting frame 9; the collecting frame 9 is installed on the working platform 3; multiple buffer nets 802 are provided inside the fruit bag 803.

[0039] Preferably, the buffer net 802 has different sizes, and the inner diameter gradually increases with the distance from the upper ring cover 801, forming a multi-level buffer structure, which effectively slows down the falling speed of the fruit and reduces fruit damage.

[0040] In addition, such as Figures 8-12 As shown, this embodiment also proposes a method for cutting and harvesting fruit stalks, using the fruit stalk cutting and harvesting robot described above, specifically including the following steps: S1. Collect images of the working environment, plan the motion path, and control the fruit stem cutting and picking robot to move to the target picking area; S2. Perform rough fruit location and area division on the target picking area to determine the target area for priority picking; S3. Control the fruit stem cutting and picking robot to move to the target area, obtain the spatial position information of the fruit and fruit stem, and judge the degree of occlusion of the fruit stem of each effective picking target according to the spatial position information, and generate a picking sequence. S4. Harvest the fruits in the target area according to the priority order of the harvesting sequence, and adopt corresponding harvesting strategies for different degrees of shading.

[0041] Before step S1, step S0 is included: first, images of the target fruit are acquired using the recognition and positioning device 5, and then key points are manually labeled. The key points include the fruit bounding box key points and the fruit stem position key points. The labeled image data is compiled into a dataset and trained using the YOLO V11 network, i.e., using the YOLO-Initialpositioning (YOLO-IP) model, to obtain a key point recognition model for the target fruit and its stem for subsequent actual detection and recognition.

[0042] In step S1, the first camera and lidar of the visual navigation device 2 acquire images and depth information of the working environment and transmit them to the control device 7. The control device 7 corrects the errors in the images of the working environment, fills in missing information, fuses multi-source data to generate a continuous global environment map, and collects dynamic and static obstacle information in real time, updates the global environment map, and determines the optimal movement trajectory. Subsequently, the control device 7 generates navigation commands and sends them to the mobile chassis 1, controlling the mobile chassis 1 to move to the target picking area according to the optimal trajectory.

[0043] In step S2, after the robot reaches the designated location, it performs coarse positioning of the fruit target, intelligent area division, and priority picking decision based on the image data collected by the first camera of the visual navigation device 2.

[0044] The coarse localization unit uses an optimized YOLO V11 network to detect fruit targets in the image. The specific improvement scheme is as follows: Step 2.1: Replace the original feature extraction module of YOLO V11 with MobileNetV4 as the backbone network. Specifically, the Backone feature extraction network in YOLO V11 is replaced with the lightweight feature extraction architecture of MobileNetV4, while retaining the inverted residual structure and efficient attention mechanism of MobileNetV4. A multi-scale feature extraction path is constructed through depthwise separable convolutions and linear bottleneck layers. This module achieves a balance between accuracy and efficiency by fusing depthwise separable convolutions with an efficient attention mechanism.

[0045] Feature enhancement operations are as follows: ; in, , These represent the dimension-up transformation and dimension-down transformation achieved through pointwise convolution, respectively. For depthwise convolution, A lightweight attention module is used to generate a feature-weighted map. Subsequently, residual connections are used to enhance the network's generalization ability. ; In the formula, This indicates random depth regularization to enhance the model's generalization ability. When the input / output dimensions match, If it is an identity mapping, then it is a projection transformation to ensure that the residual connection holds.

[0046] Step 2.2: In terms of feature fusion, a pyramid structure is adopted to connect the multi-scale feature maps output by MobileNetV4. The feature pyramid is constructed through upsampling and downsampling operations to enhance the detection capability of fruits of different sizes. At the same time, an adaptive feature fusion module is introduced to dynamically adjust the weights of features at different levels according to the complexity of the input image, thereby improving the robustness of the network in complex orchard environments.

[0047] Step 2.3: The network structure is designed to be lightweight in the detection head part, and depthwise separable convolutions are used to replace traditional convolutional layers, reducing the number of model parameters and computational complexity; at the same time, a channel attention mechanism is introduced to enhance the network's ability to extract key features of the fruit, and significantly improve the inference speed while maintaining high detection accuracy.

[0048] Step 2.4: Based on the size parameters of the working environment image, the detection area is uniformly divided into multiple equal-width areas horizontally and equal-height areas vertically, forming multiple rectangular detection sub-regions; simultaneously, all detected fruit bounding boxes are traversed, the center point coordinates of each fruit bounding box are calculated, and the detection sub-region to which it belongs is determined based on the center point coordinates, and the number of fruits in each sub-region is accumulated and counted to generate a fruit density distribution matrix; based on the density distribution matrix, the number of fruits in each region is sorted and compared, and the region with the most fruits is automatically identified and selected as the priority picking target region.

[0049] For example, the detection area is evenly divided into three equal-width regions horizontally and two equal-height regions vertically, forming six rectangular detection sub-regions. Simultaneously, all detected fruit bounding boxes are traversed, the center point coordinates of each fruit bounding box are calculated, and the detection sub-region to which it belongs is determined based on the center point coordinates. The number of fruits in each sub-region is accumulated and counted to generate a fruit density distribution matrix. Based on this density distribution matrix, the number of fruits in each region is sorted and compared, and the region with the most fruits is automatically identified and selected as the priority picking target region for the robotic arm 4.

[0050] The specific process is as follows: Step 2.4.1: Divide the detection area into three equal parts horizontally and two equal parts vertically, forming a grid. Each sub-region... The mathematical definition of is: ; In the formula, , . This represents the floor function, ensuring the boundaries are integer pixels; W and H are the total width and total height of the image, respectively. This definition makes... The interval is left-closed and right-open, thus making all sub-regions mutually exclusive and collectively covering the entire image.

[0051] Step 2.4.2: Density Distribution Matrix Each element The calculation is as follows: ; in, It is an indicator function that returns 1 when the condition is true, and 0 otherwise.

[0052] Step 2.4.3: Solve for the density distribution matrix The maximum index is used to automatically determine the target area where the fruit is most densely packed. ; Ultimately, the region The area was selected as the target area for priority harvesting operations by robotic arm 4.

[0053] Based on the target area for priority harvesting, the control device 7 drives the robotic arm 4 to move, causing the actuator 6 and the identification and positioning device 5 to move synchronously to the optimal observation posture. The control device 7 calculates the motion trajectory based on the inverse kinematics algorithm and achieves precise positioning through position closed-loop control, while avoiding obstacles. This allows the identification and positioning device 5 to obtain clear and unobstructed images, providing reliable visual input for subsequent fruit identification and positioning.

[0054] In step S3, the identification and positioning device 5 acquires images and depth images of the target fruit in the target area and transmits them to the control device 7; the control device 7 employs a multimodal fusion detection network, namely YOLO-Precise Localization (YOLO-Precise Localization). The YOLO-PP positioning network performs target detection. This network fuses the spatial location information of the depth image with the texture features of the visible light image through a feature complementary mapping mechanism. The multi-core perception unit in the control device 7 enhances the modeling of the spatial relationship between the fruit and the stem, locating their two-dimensional coordinates and three-dimensional spatial positions. Subsequently, the detection results are verified for completeness, and valid target pairs that simultaneously meet the confidence thresholds of the fruit and stem and have a reasonable spatial distance are selected. Next, the multi-level occlusion judgment module in the control device 7 analyzes the depth gradient features of the area in front of the stem and classifies the targets into three levels: no occlusion, slight occlusion, and severe occlusion based on the occlusion score. Finally, the harvesting priority score of each target is calculated by combining the fruit confidence, stem confidence, fruit size, and occlusion level, and the optimal harvesting sequence from no occlusion to severe occlusion is dynamically generated to guide the robotic arm 4 of the execution device to perform the harvesting operation in priority order.

[0055] Specifically as follows: Step 3.1: Fruit and pedicel detection and location: In the feature complement mapping stage, the YOLO-PP network first performs channel interaction processing on semantic features: each channel is processed independently through depthwise separable convolutions, generating intermediate features, which are then derived by global average pooling and a sigmoid (non-linear) activation function to derive channel attention weights. Simultaneously, spatial features are spatially interacted, generating spatial weights using a spatial attention module constructed with 1×1 convolutions, BN layers, and a sigmoid function. Subsequently, a bidirectional weighted fusion mechanism is used to connect the weighted semantic features and spatial features, forming a fused feature that simultaneously contains rich semantic information and precise spatial location, mitigating the spatial information loss problem caused by downsampling. Meanwhile, the multi-kernel perceptual unit introduced by the network adopts a parallel convolutional structure, using multi-scale convolutional kernels such as 3×3, 5×5, and 7×7 to collaboratively capture contextual information, strengthening the modeling of the spatial relationship between the fruit and its stem. Finally, the network outputs high-precision two-dimensional coordinates of the target fruit and stem based on the fused multi-dimensional features, and generates three-dimensional spatial location information by combining depth data. The specific process is as follows: First, semantic features Perform channel interactions to generate channel weights for enhancing spatial features. Using depthwise separable convolution to process each channel independently, we obtain... , its first The calculation for each channel is as follows: ; in, This represents a depthwise convolution operation. It is the first One convolutional kernel, It is the first One input channel. Then for... Global average pooling is performed and channel attention weights are generated using the Sigmoid activation function. : ; in, , For the Sigmoid function, GAP represents global average pooling.

[0056] Spatial features Perform spatial interactions to generate spatial weights for enhancing semantic features. A lightweight module consisting of 1×1 convolutions, BN layers, and a sigmoid function is used to aggregate spatial information and generate a spatial attention map. ; in, In obtaining channel weights Spatial weights After that, through two-way weighting and fusion, the final features are generated. Feature aggregation is achieved through the following formula: ; where, represents element-wise multiplication, represents the feature concatenation operation. The finally output fused feature contains both rich semantic information and accurate spatial position information. This module has lightweight computation. By explicitly transmitting shallow spatial information to the deep layers of the network, it effectively alleviates the problem of spatial information loss caused by downsampling in the backbone network, thereby enhancing the model's perception ability of object positions.

[0057] Step 3.2: Object integrity detection and screening: Based on the detection results output by the YOLO-PP network, the integrity of each detected object is verified to enable the existence of matching pairs of fruits and fruit stalks simultaneously. Specifically, it includes: Calculation of fruit-fruit stalk matching degree: By analyzing the spatial position relationship, calculate the distance and angle relationship between each fruit and the surrounding fruit stalks to establish fruit-fruit stalk association pairs. Specifically, calculate the Euclidean distance between the center point of the fruit and the center point of the fruit stalk, as well as the angular deviation of the fruit stalk relative to the fruit, and screen out the matching pairs whose distance and angle are both within a reasonable range.

[0058] Judgment of object integrity: Only when the detection confidence of the fruit exceeds the threshold and the detection confidence of the fruit stalk exceeds the threshold , and at the same time the spatial distance between the two is within a reasonable range (D_min < D_fruit-stem < D_max), it is determined as a valid picking target.

[0059] Object screening: Eliminate the detection targets with only fruits without fruit stalks, only fruit stalks without fruits, or insufficient confidence of both, so that this device only processes complete target pairs, that is, valid target pairs.

[0060] Step 3.3: The multi-level occlusion judgment module analyzes the occlusion situation in the area in front of the fruit stalk: For the valid target pairs screened in Step 3.2, extract the region of interest (ROI, referring to a specific region in the image that needs to be focused on, processed, or analyzed) in front of the fruit stalk, and calculate the depth gradient discontinuity and depth variance features in this region by combining with the depth image. Generate an occlusion score through weighted fusion, and divide the occlusion state into three levels: unoccluded, lightly occluded, and heavily occluded according to a preset threshold, where: Unoccluded: The depth in the area in front of the fruit stalk is uniform, without obvious depth mutations, that is ; Lightly occluded: There are a small number of depth mutations in the area in front of the fruit stalk, that is ; Severe shading: The depth of the area in front of the fruit stalk changes drastically, that is... ; T_1 and T_2 are preset occlusion judgment thresholds, which are calibrated according to the actual orchard environment and fruit characteristics. The occlusion score S_occ is calculated by comprehensively considering the depth gradient magnitude, depth variance, and the number and distribution density of depth abrupt change points within the ROI region.

[0061] Step 3.4: Dynamically adjust picking priority based on occupancy level: Based on the occlusion level obtained in step 3.3, and combined with the fruit detection confidence level, fruit size, and spatial location information, the harvesting priority score for each target fruit is calculated using the following formula: ; in The confidence level of the fruit as determined by YOLO-PP. The area of ​​the fruit in pixels. To determine the confidence level for fruit stalk detection, The occlusion level is 3. , , , These are weighting coefficients; the system is based on... All target fruits are sorted from highest to lowest score to generate the optimal picking sequence. Unobstructed fruits are picked first, while heavily obstructed fruits are either not picked or marked as pending.

[0062] In step S4, the harvesting strategy is optimized based on the occlusion situation, specifically as follows: Step S4.1: Based on the occlusion level determined in steps 3.3 and 3.4 and the analysis results, the control device 7 adopts a differentiated obstacle avoidance and clearance strategy: Unobstructed Fruit: The standard picking path is directly executed to complete the picking with maximum efficiency. The control device 7 commands the robotic arm 4 to rotate directly to the optimal clamping angle. At the same time, the servo motor 605 drives the rotating ring 606 and the second housing 608 to rotate through the meshing of the external and internal gears, so that the moving blade 601 on the actuator 6 is directly facing the fruit, that is, the fruit is located directly below the moving blade 601, in preparation for subsequent cutting.

[0063] Slightly obstructed fruit: Adjust the approach angle of actuator 6 and use a tentative contact strategy to clear the obstruction and harvest the fruit.

[0064] The control device 7 first drives the robotic arm 4, which in turn drives the actuator 6 to approach the target via a non-direct "conservative path," thereby avoiding collisions with the main branches. At the same time, the actuator 6 adjusts its own posture so that the opening plane of its movable stop bar 602 is aligned with or at a preset angle with the main obstruction area, preparing for the cleaning action.

[0065] After positioning the predetermined cleaning location, the control device 7 does not immediately instruct the drive assembly 603 to perform the closing motion necessary for shearing. Instead, it commands it to execute a specific large-amplitude opening procedure (opening angle of 70%–100% of the maximum opening angle). The main purpose of this opening action is to allow the movable stop bars 602 on both sides to unfold and extend forward and laterally along with the moving blade table 601. The movable stop bars 602, as dedicated cleaning components, will perform the following tasks: Sweeping: Using its plate-like structure, it pushes aside thin obstructions (such as leaves or tendrils) hanging in front of the fruit or stalk, forward or to the side.

[0066] Push away: Uses its physical structure to push smaller branches away from the working path.

[0067] Separation: In areas with dense fruit, the movable baffle 602 can effectively separate adjacent fruits, creating favorable conditions for precise positioning and cutting of the target fruit.

[0068] To expand the cleaning range, the robotic arm 4 can drive the actuator 6 to swing in small amplitudes in multiple directions, so that the movement trajectory of the movable stop 602 changes from a static position to a dynamic "sweeping" area, thereby more effectively clearing obstacles on the side.

[0069] After each clearing operation, the identification and positioning device 5 re-acquires the image and evaluates the occlusion level. If the occlusion level is reduced to no occlusion or within an acceptable range, the picking process begins. If the occlusion level is not reduced or still exceeds the acceptable threshold after a predetermined number of clearing operations, it is marked as a target to be processed and skipped, moving on to the next target fruit.

[0070] Severely obscured fruit: Mark it as a target to be processed, record its spatial location, and process it after the surrounding fruit has been harvested or through human intervention.

[0071] Through the coordinated operation of the above sub-steps, the system achieves accurate screening of complete fruit-stalk targets, judgment of occlusion in front of the fruit stalk, and intelligent optimization of harvesting strategies, ensuring that only complete target pairs are harvested.

[0072] Step S4.2: The control device 7 controls the movement of the robotic arm 4, which in turn drives the actuator 6 to move, causing the blade 60101 on the movable blade table 601 to move to the fruit stem picking point of the target fruit; the control device 7 calculates the clamping force required according to the shape and material of the fruit stem, and drives the two movable blade tables 601 to move closer to each other through the drive component 603; the flexible clamping block 60102 on the movable blade table 601 fits against the surface of the fruit stem, so that the fruit stem is firmly clamped; then the blade 60101 on the movable blade table 601 cuts the fruit stem.

[0073] In this embodiment, the control device 7, by recognizing the three-dimensional spatial position information of the fruit stalk obtained by the positioning device 5 and combining it with the inverse kinematics model of the robotic arm 4, calculates the optimal motion trajectory of the blade 60101 to reach the fruit stalk cutting point. Simultaneously, based on the fruit stalk thickness information and material characteristics output by the YOLO-PP network, the control device 7 calculates the clamping force of the drive component 603 through a preset force control model, ensuring a stable grip on the fruit stalk without damaging its tissue. The flexible clamping block 60102 is made of soft material, capable of adapting to fruit stalks of different thicknesses and shapes, improving the adaptability and reliability of the clamping.

[0074] Step S4.3: During the cutting process, the depth camera of the positioning device 5 detects the change in distance between the actuator 6 and the target fruit in real time. By monitoring whether the fruit moves with the actuator 6, it is determined whether the fruit stem has been successfully cut, thus indicating successful cutting. If a cutting failure is detected, the control device 7 will terminate the remaining actions through a feedback mechanism and repeat steps S3, S4.1, S4.2, and S4.3 until the harvest is successful. After successful cutting, the drive assembly 603 remains in a clamping state, and the robotic arm 4 moves the actuator 6 above the fruit bag 803. Then, the drive assembly 603 releases the moving blade 601, the fruit is released from the actuator 6, and then the fruit moves through the fruit bag 803 and is buffered by the various levels of buffer nets 802 into the collection box 9, completing a single harvesting cycle.

[0075] Step S4.4: After harvesting all target fruits in the current harvesting sequence within the field of view of the identification and positioning device 5, the control device 7 determines whether there are any unharvested fruits in the current target area. If so, the robotic arm 4 drives the identification and positioning device 5 to adjust its viewing angle and re-executes steps S3 and S4 until all harvestable fruits in the current target area have been processed.

[0076] Step S4.5: After harvesting all the target fruits in the current target area within the field of view of the identification and positioning device 5, repeat steps S1 to S4 to continue harvesting target fruits in other areas until the harvesting task of the entire work area is completed.

[0077] Throughout the harvesting process, the control device 7 monitors real-time feedback from various sensors, including the position and torque of each joint of the robotic arm 4, the clamping status of the actuator 6, and the image quality of the identification and positioning device 5, ensuring the safety and reliability of the harvesting process. When an abnormality is detected, such as obstruction of the robotic arm 4, malfunction of the actuator 6, or obstruction of the field of view of the identification and positioning device 5, the control device 7 immediately triggers the protection mechanism, suspends the current operation, generates an alarm message, and resumes operation after manual intervention or automatic recovery.

[0078] In addition, to improve overall operational efficiency, the control device 7 adopts a dynamic task scheduling strategy. While executing the current picking task, the visual navigation device 2 scans the next target area in advance, performs pre-planning and priority evaluation, and realizes the continuous operation of the picking task in an assembly line manner, minimizing waiting and idle time.

[0079] In this embodiment, adaptive harvesting in complex orchard environments is achieved through a collaborative operation of "coarse positioning-fine positioning" hierarchical visual perception and mechanically assisted cleaning. First, the visual navigation device 2 collects environmental information, and the control device 7 plans a path to drive the chassis to the target area. Second, in coarse positioning mode, a lightweight YOLO-IP network is used to quickly identify all fruits and divide them into a 3×2 grid, selecting the area with the highest fruit density as the priority target. In fine positioning mode, a YOLO-PP network is used to fuse depth maps and visible light images to accurately obtain the three-dimensional coordinates of the fruits and stems. Subsequently, the control device 7 verifies the integrity of the target, calculates an occlusion score using depth gradient and variance features, and classifies the target into three levels: unoccluded, slightly occluded, and heavily occluded. A harvesting priority sequence is generated by comprehensively considering confidence level, size, and occlusion level. Finally, differentiated strategies are adopted for different occlusion levels: unoccluded fruits are harvested directly; slightly occluded fruits are harvested after being cleaned using the movable baffle 602; and heavily occluded fruits are marked for further processing. During the shearing process, the movable blade holder 601 precisely controls the clamping force, and the flexible clamping block 60102 securely holds the blade 60101 to complete the shearing. The camera monitors and provides feedback on the harvesting results in real time. Through the coordinated work of hierarchical positioning, intelligent decision-making, adaptive execution, and closed-loop feedback, efficient and precise adaptive harvesting operations are achieved.

[0080] In summary, the fruit stalk cutting and harvesting robot and its harvesting method proposed in this invention have the following advantages: The visual navigation device 2 performs environmental perception, path planning, and coarse fruit positioning, while the identification and positioning device 5 precisely locates the fruit and fruit stems. The control device 7 determines the degree of occlusion based on spatial location information and adopts corresponding harvesting strategies, effectively improving the harvesting success rate and overall operational efficiency in complex orchard environments. By using an execution device to clear obstructions and cut fruit stems, the system's adaptability to different occlusion conditions is enhanced, reducing the missed harvest rate.

[0081] Furthermore, a multimodal fusion detection network is used to fuse spatial location information from depth images with texture features from visible light images, enhancing the spatial relationship modeling between fruits and stems and improving target localization accuracy. By dividing the detection area into multiple sub-regions and calculating fruit density, priority is given to harvesting densely populated fruit areas. Combining occlusion level judgment and harvesting priority ranking, an optimal harvesting sequence is dynamically generated, significantly improving harvesting efficiency. The rotation adjustment mechanism and movable baffle 602 in the execution device clear obstructions, along with the buffer net 802 and adaptively oriented fruit bag 803 in the collection device 8, reduce fruit damage rate and improve harvesting quality.

[0082] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A fruit stalk cutting and harvesting robot, characterized in that, It includes a mobile chassis, a work platform, a control device, an execution device, a visual navigation device, and an identification and positioning device; The work platform is mounted on the mobile chassis; the control device, the execution device, and the visual navigation device are all mounted on the work platform. The visual navigation device is used for environmental perception, path planning, and coarse positioning of fruits to identify areas with dense fruit distribution. The identification and positioning device is installed on the execution device and is used to accurately locate the fruit and fruit stalk, and obtain the spatial location information of the fruit and fruit stalk to be picked. The actuator is used to clear obstructions and cut fruit stems; The control device is used to determine the degree of occlusion based on the spatial location information, control the execution device to adopt corresponding picking strategies according to different degrees of occlusion, and control the mobile chassis, execution device, visual navigation device and identification and positioning device to work together.

2. The fruit stalk cutting and harvesting robot as described in claim 1, characterized in that, The execution device includes a robotic arm, a servo motor, a first housing, a second housing, a rotating ring, and an actuator; The robotic arm is mounted on the work platform and its end is connected to the first housing; The servo motor is mounted on the first housing, and its output shaft is connected to an external gear; The rotating ring is arranged around the outer periphery of the second housing and is fixedly connected to the second housing. An internal gear is provided on the inner side of the rotating ring. The second housing is rotatably mounted on the end of the first housing away from the robotic arm. The second housing is provided with a groove, through which the output shaft of the servo motor passes so that the external gear can mesh with the internal gear. The actuator is mounted on the second housing.

3. The fruit stalk cutting and harvesting robot as described in claim 2, characterized in that, The actuator includes a mounting frame, a drive assembly, a pair of movable tool holders, and a pair of movable stop levers; The mounting frame is mounted on the second housing; the drive assembly is mounted inside the mounting frame and connected to one end of the movable tool post to drive the pair of movable tool posts to move towards or away from each other; each movable tool post is provided with a blade; each movable stop is connected to the corresponding movable tool post by a sliding connection, and the movable stop can move outward synchronously with the opposite movement of the movable tool posts.

4. The fruit stalk cutting and harvesting robot as described in claim 2, characterized in that, It also includes a collection device; the collection device is installed on the working platform and located below the actuator, for receiving and buffering the harvested fruit; The collection device includes an upper ring cover, a fruit bag, a buffer net, and a collection frame; the outer wall of the rotating ring is provided with an annular slide rail; the upper ring cover is slidably connected to the annular slide rail and can slide along the annular slide rail, and under the action of gravity, the upper ring cover always slides to the lowest position of the rotating ring; One end of the fruit bag is connected to the upper ring cover, and the other end of the fruit bag is connected to the collection frame; the collection frame is installed on the working platform; multiple buffer nets are provided inside the fruit bag.

5. A method for cutting and harvesting fruit stalks, using a fruit stalk cutting and harvesting robot as described in any one of claims 1-4, characterized in that, Specifically, it includes the following: The robot collects images of the working environment, plans a motion path, and controls the fruit stem cutting and picking robot to move to the target picking area. The target harvesting area is roughly located and divided into regions to determine the priority harvesting areas; The fruit stem cutting and picking robot is controlled to move to the target area, obtain the spatial position information of the fruit and fruit stem, and judge the degree of occlusion of the fruit stem of each effective picking target based on the spatial position information to generate a picking sequence. Fruits in the target area are harvested according to the priority order of the harvesting sequence, and corresponding harvesting strategies are adopted for different degrees of shading.

6. The method for cutting and harvesting fruit stalks as described in claim 5, characterized in that, The process of acquiring images of the working environment, planning a motion path, and controlling the fruit stem cutting and harvesting robot to move to the target harvesting area specifically includes: The visual navigation device collects images of the working environment and transmits them to the control device. The control device corrects the errors in the images, fills in missing information, integrates multi-source data to generate a continuous global environment map, and collects dynamic and static obstacle information in real time, updates the global environment map, and determines the optimal movement trajectory. Subsequently, the mobile chassis moves to the target picking area according to the optimal trajectory.

7. The method for cutting and harvesting fruit stalks as described in claim 5, characterized in that, The process of roughly locating and dividing the target harvesting area to determine the priority harvesting area specifically includes: The coarse localization of the fruit uses a lightweight target detection network, which extracts multi-scale fruit features with a lightweight backbone network to achieve rapid detection and region segmentation. Based on the size parameters of the working environment image, the detection area is uniformly divided into multiple equal-width regions horizontally and multiple equal-height regions vertically, forming multiple rectangular detection sub-regions. Simultaneously, all detected fruit bounding boxes are traversed, the center point coordinates of each fruit bounding box are calculated, and the detection sub-region to which it belongs is determined based on the center point coordinates. The number of fruits in each sub-region is accumulated and statistically analyzed to generate a fruit density distribution matrix. Based on the density distribution matrix, the number of fruits in each region is sorted and compared, and the region with the most fruits is automatically identified and selected as the priority target region for harvesting.

8. The method for cutting and harvesting fruit stalks as described in claim 7, characterized in that, The process of acquiring spatial location information of the fruit and its stem, and determining the degree of occlusion of the stem for each valid harvesting target based on the spatial location information to generate a harvesting sequence, specifically includes: The identification and positioning device collects target fruit images and depth images in the target area and transmits them to the control device. The control device uses a multimodal fusion detection network to fuse the spatial location information of the depth image with the texture features of the visible light image through a feature complementary mapping mechanism, and strengthens the modeling of the spatial relationship between the fruit and the stem to locate their two-dimensional coordinates and three-dimensional spatial positions. Then, the detection results are verified for completeness, and effective target pairs that simultaneously meet the confidence thresholds of the fruit and stem and have a reasonable spatial distance are selected. Next, the depth gradient features of the area in front of the stem are analyzed, and the targets are divided into three levels: no occlusion, slight occlusion, and severe occlusion, based on the occlusion score. Finally, the harvesting priority score of each target is calculated by combining the fruit confidence, stem confidence, fruit size, and occlusion level, and the optimal harvesting sequence from no occlusion to severe occlusion is dynamically generated to guide the execution device to perform the harvesting operation in priority order.

9. The method for cutting and harvesting fruit stalks as described in claim 8, characterized in that, First, channel interaction is performed on semantic features to generate channel weights for enhancing spatial features, and depthwise separable convolution is used to process each channel independently. Spatial features are spatially interacted to generate spatial weights for enhancing semantic features, and spatial information is aggregated to generate a spatial attention map. After obtaining the channel weights and spatial weights, the final features are generated through bidirectional weighting and fusion. The final features contain both semantic information and spatial location information, and the detection results are output. Based on the detection results, integrity verification is performed on each detected target to ensure that there are matching pairs of fruit and stem simultaneously. Specifically, this includes: calculating the fruit-stem matching degree; judging target integrity; filtering targets to obtain valid target pairs; for the filtered valid target pairs, extracting the region of interest in front of the stem and calculating the depth gradient discontinuity and depth variance features within the region, generating an occlusion score through weighted fusion, and classifying the occlusion state into multiple occlusion levels according to a preset threshold; calculating the picking priority score for each target fruit based on the occlusion level, combined with fruit detection confidence, fruit size, and spatial location information; and sorting all target fruits from high to low according to the picking priority score to generate the optimal picking sequence.

10. The method for cutting and harvesting fruit stalks as described in claim 9, characterized in that, The process of harvesting fruits in the target area according to the priority order of the harvesting sequence, and adopting corresponding harvesting strategies for different degrees of shading, specifically includes: Based on the occlusion level and analysis results, differentiated obstacle avoidance and removal strategies are adopted: Unobstructed fruit: Follow the standard picking path directly; Slightly obstructed fruit: Adjust the actuator's approach angle and use a trial-and-error contact strategy to clear the obstructions and pick the fruit; After each clearing operation, the identification and positioning device re-acquires images and assesses the obstruction level. If the obstruction level decreases, proceed with the picking process; if the obstruction level does not decrease, mark it as a target to be processed and skip it; Severely obstructed fruit: Mark it as a target to be processed, record its corresponding spatial location, and process it after the surrounding fruit has been picked or after intervention. The control device calculates the clamping force required based on the shape and material of the fruit stalk, and then cuts the fruit stalk. During the cutting process, the identification and positioning device detects the distance between itself and the target fruit in real time to ensure that the fruit stem is cut successfully. If a cutting failure is detected, the control device will terminate the remaining actions through a feedback mechanism and repeat the above steps until the harvest is successful. Then the fruit is moved to the collection box through the fruit bag. After harvesting the target fruit in the field of view of the identification and positioning device, repeat all the steps to continue harvesting the target fruit in the remaining area.

Images