Passion fruit intelligent identification and picking system based on unmanned aerial vehicle

Abstract

The invention discloses an intelligent passion fruit recognition and picking system based on an unmanned aerial vehicle, and belongs to the technical field of agricultural intelligent equipment, and the system comprises an unmanned aerial vehicle flying platform, and a multi-mode intelligent recognition module, a precise picking execution mechanism and an intelligent dynamic sorting module which are carried on the unmanned aerial vehicle flying platform. The multi-mode intelligent identification module collects color images, near infrared spectrum data and three-dimensional point cloud data of the fruits, and carries out fusion processing to judge the maturity type, the contour size and the surface defect state of the fruits. And the precise picking execution mechanism picks the fruits according to the identification result. And the intelligent dynamic sorting module automatically sorts the picked fruits into different storage bins according to the fruit information. The problems that manual passion fruit harvesting efficiency is low, quality judgment depending on experience is uneven, and existing equipment lacks integrated operation capacity are solved, sensing, decision making and execution functions are integrated through the unmanned aerial vehicle platform, and automatic, intelligent and high-quality passion fruit harvesting is achieved.

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent equipment technology, and in particular to a passion fruit intelligent identification and harvesting system based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Passion fruit, as a fruit with high economic value, relies heavily on harvesting in large-scale plantations. Traditional harvesting methods primarily rely on manual labor. Workers need to go deep into the orchards to visually assess the ripeness, size, and appearance defects of the fruit, then manually pick and initially sort them. This model has gradually revealed many limitations in dealing with the ever-expanding scale of cultivation.

[0003] First, manual harvesting is inefficient and labor-intensive. Harvesting seasons are often concentrated, requiring the hiring of a large number of temporary workers. This not only results in high labor costs, which constitute a significant proportion of total planting costs, but also increases the difficulty and safety risks of manual labor in orchards with complex terrain such as hills and mountains. Second, relying on manual experience to judge fruit ripeness and quality is subjective and lacks standardized criteria. Passion fruit ripeness is not only related to external color but also closely linked to internal indicators such as sugar content. Visual observation alone can easily lead to misjudgments, resulting in inconsistent ripeness and quality of harvested fruit, affecting the marketable fruit rate and market value. Improper handling during manual harvesting can easily cause mechanical damage to the fruit skin or damage fruiting branches, affecting subsequent yields. Harvested fruit usually needs to be collected and sorted a second time, increasing logistics and labor costs.

[0004] With the development of agricultural automation technology, some ground-based mobile robots or robotic arms have been tested for fruit and vegetable harvesting. However, for vine plants like passion fruit, where the fruit hangs irregularly from trellises, ground equipment faces challenges in terms of accessibility, field of vision, and proximity to the fruit. While existing agricultural drone technology is widely used for plant protection spraying and remote sensing monitoring, its functions are relatively limited, lacking integrated capabilities specifically for fruit harvesting, which requires precise perception, delicate operation, and real-time decision-making. In particular, given passion fruit's thin and fragile peel, near-ellipsoidal shape, and the need for comprehensive internal and external maturity assessment, there is a lack of dedicated, lightweight, high-precision identification and flexible execution mechanisms adapted to drone platforms.

[0005] Therefore, there is a need in this field for an aerial harvesting system that can adapt to complex orchard environments, achieve efficient and automated harvesting, and perform online intelligent identification and sorting of fruit quality, in order to overcome the shortcomings of manual harvesting and existing technical solutions, and improve the overall efficiency and economic benefits of passion fruit harvesting operations. Summary of the Invention

[0006] This invention overcomes the problems of low efficiency in manual passion fruit harvesting, inconsistent quality due to reliance on experience, and lack of integrated operation capabilities in existing equipment. By integrating perception, decision-making, and execution functions through a drone platform, it achieves automated, intelligent, and high-quality harvesting of passion fruit.

[0007] To achieve the above objectives, the present invention adopts the following solution: The drone-based intelligent passion fruit identification and harvesting system includes a drone flight platform and its onboard multimodal intelligent identification module, precision harvesting actuator, and intelligent dynamic sorting module, wherein: The multimodal intelligent recognition module includes an RGB camera, a near-infrared spectral sensor, and a lidar for acquiring multi-source fruit data. The multi-source fruit data includes color images, near-infrared spectral data, and 3D point cloud data of passion fruit. The multimodal intelligent recognition module processes the acquired multi-source fruit data to obtain and output fruit category information, including maturity category, outline size, and surface defect status. The processing includes: Feature data of fruit surface color in HSV color space is extracted from color images. A prediction model of fruit internal sugar content is established based on near-infrared spectral data. By fusing the feature data of HSV color space with the output of the prediction model of fruit internal sugar content, the fruit maturity category is determined. The outline size of the fruit is calculated based on 3D point cloud data. The surface defect status of the fruit is identified based on color images. The precision harvesting execution mechanism compares the various fruit information output by the multimodal intelligent recognition module with the preset qualification conditions, and performs harvesting actions on fruits that meet the conditions; The intelligent dynamic sorting module is connected to the precision picking execution mechanism and receives fruit information output by the multimodal intelligent recognition module. Based on the fruit information, it sorts the fruits picked by the precision picking execution mechanism into different storage bins.

[0008] As a preferred embodiment, the process of determining the fruit maturity category in the multimodal intelligent recognition module specifically includes: Pixels of the centroid region of each fruit are extracted from the color image and converted to the HSV color space. The average value of its hue components is calculated as the color feature value. A near-infrared spectral sensor is driven to collect diffuse reflectance spectra on the same fruit surface to obtain spectral sequences in the wavelength range of 900nm to 1700nm. The spectral sequences are input into a pre-trained partial least squares regression model, which is trained based on the known sugar content of passion fruit samples and outputs the predicted sugar content of the fruit. The obtained color feature value is compared with the preset ripe color threshold range, and the sugar content prediction value is compared with the preset ripe sugar content threshold. If the color feature value falls within the ripe color threshold range and the sugar content prediction value is greater than the ripe sugar content threshold, the ripeness category of the corresponding fruit is determined to be ripe; otherwise, it is determined to be unripe.

[0009] Preferably, the process of calculating the fruit outline size and estimating the quality in the multimodal intelligent recognition module specifically includes: The three-dimensional point cloud data of the target fruit was collected using lidar, and the point set representing the fruit surface was segmented from the three-dimensional point cloud data. Based on the obtained fruit surface point set, an ellipse fitting algorithm was used to construct a spatial ellipse model reflecting the shape of the fruit. The major axis diameter and minor axis diameter of the fruit were extracted from the spatial ellipse model as its outline dimensions. The fruit was approximated as an ellipsoid, and the approximate volume of the fruit was calculated based on its major axis diameter and minor axis diameter. The approximate volume was multiplied by the average density of the passion fruit to obtain the estimated mass of the corresponding fruit.

[0010] Preferably, the process of identifying the state of fruit surface defects in the multimodal intelligent recognition module specifically includes: High-resolution color images of the fruit captured by an RGB camera are acquired and preprocessed, and the fruit regions are segmented. Within the segmented fruit regions, a local binary mode operator is used to extract feature vectors representing the surface texture of the fruit. Specifically: A three-by-three neighborhood is constructed with each pixel as the center. An eight-bit binary pattern code is generated based on the comparison of the gray values ​​of the neighboring pixels and the center pixel. The local binary pattern feature histogram of all pixels in the selected area is used as the texture feature vector. The obtained texture feature vector is input into a pre-trained classification model based on support vector machine. The classification model outputs the probability of the corresponding fruit belonging to various defect states based on the input feature vector, and determines whether the fruit has surface defects and the defect category based on the maximum probability.

[0011] As a preferred method, the classification model is pre-trained in the following way: Based on historically collected, classified, and verified passion fruit defect information, a dedicated passion fruit defect dataset was constructed, including sample images of normal fruit, insect-damaged fruit, cracked fruit, and deformed fruit. Each image in the dataset was labeled and assigned a corresponding defect category label. During the model training phase, a transfer learning strategy was adopted, using a convolutional neural network pre-trained on a large general image dataset as the base model; this base model was fine-tuned using a passion fruit-specific defect dataset; during the fine-tuning process, an attention mechanism module designed for the characteristics of small-scale defects on the fruit surface was introduced, which focuses on enhancing the model's ability to capture features of insect-eaten holes and fine cracks. The trained model is quantized and compressed, and converted into a lightweight format that can run in real time on the embedded computing platform of the drone.

[0012] As a preferred option, the precision harvesting mechanism includes a flexible mechanical claw, which comprises three independently driven claw fingers symmetrically distributed at 120 degrees; each claw finger includes an internal skeleton and an external covering layer, the internal skeleton being made of hard alloy and the external covering layer being medical-grade silicone. Each claw is connected to the same base via a pivot at its base and is driven to rotate around the pivot by an independent micro servo motor. When the three claws are in the initial position, the minimum distance between their inner silicone surfaces is 50mm. When the three micro servo motors drive the claws to rotate to the maximum spread angle in sync, the minimum distance increases to 120mm, thus forming a gripping range that can accommodate fruits with a diameter of 30mm to 80mm.

[0013] As a preferred embodiment, the precision harvesting mechanism also includes a negative pressure assist system, which comprises at least one negative pressure nozzle and a miniature vacuum pump. The negative pressure nozzle is made of polyurethane material, with an inlet diameter of 30mm and a curvature radius of 15mm at the tip of the inlet to conform to the arc-shaped contour of the passion fruit surface. The negative pressure nozzle is connected to the miniature vacuum pump via a flexible tube and is mounted at the center of the base of the flexible mechanical claw. The inlet plane of the negative pressure nozzle is lower than the surface of the silicone coating layer on the inner side of the three claw fingers.

[0014] As a preferred embodiment, the intelligent dynamic sorting module includes a rotary sorting bin, which is driven by a stepper motor to rotate around its central axis; the interior of the sorting bin is divided into multiple independent fruit storage chambers along the circumference, and each fruit storage chamber has an opening at the top to receive a corresponding type of fruit; the top of the sorting bin is provided with a feeding guide, which is connected to a conveying pipe for receiving the picked fruit. Based on the category information of the harvested fruit output by the multimodal intelligent recognition module, the stepper motor is driven by the controller to rotate the sorting bin until the target fruit storage bin corresponding to the current fruit category information rotates to directly below the feed inlet; the fruit falls into the target fruit storage bin through the conveying pipe and the feed inlet.

[0015] As a preferred option, the sorting process of the intelligent dynamic sorting module is as follows: The fruit information output in real time by the multimodal intelligent recognition module is transmitted to the system's controller. The controller has pre-stored classification rules for dividing the fruit into multiple categories. The classification rules are based on a comprehensive judgment of the fruit's maturity category, whether the outline size is within the preset acceptable range, and the state of surface defects. The controller compares the received fruit information with the classification rules to determine the category to which the fruit belongs. Based on the preset mapping relationship between the fruit category and each storage compartment in the rotary sorting bin, the controller calculates the required rotation angle and controls the stepper motor to drive the sorting bin to rotate by the corresponding angle, so that the target storage compartment mapped to the current fruit category is aligned with the top feed inlet. The picked fruit falls into its target storage compartment through the conveying pipe and the feed inlet.

[0016] Preferably, the classification rules pre-stored in the controller are specifically a sorting decision query table. The sorting decision query table uses the fruit maturity category, the outline size qualification mark, and the surface defect status category as joint input conditions, and outputs the corresponding target fruit storage compartment number. Among them, the outline size qualification mark is generated by comparing the long axis diameter of the fruit with the preset qualification range. If it is within the qualification range, it is true; otherwise, it is false. The fruit storage compartments include mature and qualified compartments, unripe temporary storage compartments, and defective fruit compartments. Fruit maturity is checked only when the surface defect status category is no defect and the outline size is qualified. If the maturity category is mature, the output number corresponds to the mature qualified warehouse. If the maturity category is unripe, the output number corresponds to the unripe temporary storage warehouse. When the surface defect status category is defective, or the surface defect status category is no defect but the outline dimension qualified mark is false, the corresponding defective fruit bin number will be output. In the initial state, the sorting decision query table is configured with default settings, and the maturity judgment threshold and size qualification range are dynamically updated through the ground control terminal according to the actual qualified requirements of the harvested fruits.

[0017] The present invention has at least the following beneficial effects: (1) It constructs a complete aerial autonomous operation system, realizing full-process automation from fruit information perception and picking decision to physical picking and immediate sorting, significantly reducing manual intervention, greatly improving the overall efficiency of passion fruit harvesting operations, and adapting to the operational needs of large-scale orchards and complex terrains; (2) By adopting a multimodal fusion recognition strategy of RGB image analysis for color, near-infrared spectroscopy prediction of internal sugar content, and lidar measurement of three-dimensional dimensions, it realizes a comprehensive and high-precision online assessment of fruit maturity, contour size, and surface defect status, significantly improving the accuracy and reliability of fruit quality grading and picking decision, and ensuring the uniform quality of harvested fruits from the source; (3) It designs a recognition scheme based on image texture analysis and machine learning classification model specifically for minor defects on the fruit surface, enabling the system to achieve small defects on the resource-constrained drone terminal. (4) By adopting a composite picking mechanism design with flexible mechanical claws and negative pressure assistance, while ensuring sufficient gripping force, the buffering and friction-increasing properties of the silicone coating layer and the pre-stabilizing effect of negative pressure adsorption greatly reduce the risk of crushing and scratching the fragile fruit peel, and improve the stability of successfully picking the fruit in complex postures, thus achieving truly non-destructive or low-damage precise picking; (5) Through the rotary multi-compartment design, controller decision based on real-time identification information and dynamically updated sorting rules, the immediate, automatic and orderly classification and storage of the picked fruit is realized, physically isolating fruits of different qualities, simplifying subsequent processing, and through the software-configurable sorting logic, the system can flexibly adapt to the harvesting requirements of different varieties and different market standards, enhancing the system's versatility and practicality. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system module connection relationship of the present invention; Figure 2 This is a schematic diagram of the precision harvesting mechanism of the present invention; Figure 3 This is a schematic diagram of the hopper structure of the present invention.

[0019] In the diagram: 1. Unmanned aerial vehicle (UAV) flight platform; 2. Precision harvesting mechanism; 201. Claw; 202. Base; 203. Negative pressure suction nozzle; 3. Sorting chamber; 301. Central shaft; 302. Fruit storage chamber; 4. Conveying pipeline. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0021] like Figure 1-3As shown, the passion fruit intelligent identification and harvesting system based on unmanned aerial vehicles (UAVs) provided by this invention includes a UAV flight platform 1 and a multimodal intelligent identification module, a precision harvesting execution mechanism 2, and an intelligent dynamic sorting module mounted on it, wherein: The multimodal intelligent recognition module includes an RGB camera, a near-infrared spectral sensor, and a lidar for acquiring multi-source fruit data. The multi-source fruit data includes color images, near-infrared spectral data, and 3D point cloud data of passion fruit. The multimodal intelligent recognition module processes the acquired multi-source fruit data to obtain and output fruit category information, including maturity category, outline size, and surface defect status. The processing includes: Feature data of fruit surface color in HSV color space is extracted from color images. A prediction model of fruit internal sugar content is established based on near-infrared spectral data. By fusing the feature data of HSV color space with the output of the prediction model of fruit internal sugar content, the fruit maturity category is determined. The outline size of the fruit is calculated based on 3D point cloud data. The surface defect status of the fruit is identified based on color images. The precision harvesting execution mechanism 2 compares the various fruit information output by the multimodal intelligent recognition module with the preset qualification conditions, and performs the harvesting action on the fruits that meet the conditions; The intelligent dynamic sorting module is connected to the precision picking execution mechanism 2 and receives fruit information output by the multimodal intelligent recognition module. Based on the fruit information, it sorts the fruits picked by the precision picking execution mechanism 2 into different storage bins.

[0022] The system uses a drone flight platform 1 as the basic carrier for aerial movement and operations. The flight platform possesses stable flight attitude maintenance capabilities, precise hovering and positioning capabilities, and sufficient load capacity to ensure it can smoothly carry subsequent functional modules through the orchard environment and stably operate near the target fruit. Existing civilian working drones are sufficient to provide this flight performance. Three core functional modules are integrated on this flight platform: a multimodal intelligent recognition module, a precision picking execution mechanism 2, and an intelligent dynamic sorting module. These three modules are interconnected via electrical connections and a data bus, forming a closed-loop workflow of "perception-decision-execution-sorting." The drone flight platform 1 provides the recognition module with an observation position close to the fruit, provides the execution mechanism with the mobility to reach the fruit, and provides the sorting module with a carrying and moving platform.

[0023] The multimodal intelligent recognition module serves as the system's "eyes" and "brain," responsible for comprehensively acquiring and initially analyzing fruit information. Its hardware integrates at least three different types of sensors to collect complementary multi-source fruit data. An RGB camera, an imaging device capable of capturing red, green, and blue primary colors in the visible light band, is used to obtain high-definition color images of passion fruit, recording its external color, texture, and other visual features. A near-infrared spectral sensor detects the absorption and reflection characteristics of near-infrared light. Based on the different absorption characteristics of passion fruit's internal chemical components (such as sugar and water) for different wavelengths of near-infrared light, collecting near-infrared spectral data reflected or transmitted from the fruit surface can indirectly reflect its internal quality, such as sugar content. A lidar, an active detection device that measures distance by emitting a laser beam and receiving its reflected signal, can quickly acquire dense three-dimensional point coordinates on the target fruit surface, forming three-dimensional point cloud data to accurately describe the fruit's spatial contour and geometric dimensions. These three sensors work together to comprehensively characterize the fruit's state from three dimensions: external vision, internal composition, and three-dimensional geometry, providing a comprehensive and three-dimensional data foundation for subsequent intelligent judgment. During operation, the drone hovers in a suitable position, and the module's sensor array simultaneously collects data on the target fruit, acquiring its color image, near-infrared spectral sequence, and 3D point cloud. Since passion fruit is a vine, orchards typically have corresponding climbing supports to allow the vines to spread horizontally on a platform frame. This allows the passion fruit to hang below like grapes, enabling the drone's sensors to clearly capture the fruit's position and image information, and facilitating easy harvesting by the harvesting mechanism.

[0024] The multimodal intelligent recognition module fuses the collected multi-source data to obtain more reliable fruit category information. In the processing, firstly, a dual verification strategy integrating external color and internal sugar content is adopted for determining maturity category. Specifically, the system extracts color information from specific areas (e.g., the centroid region or overall average) on the fruit surface from color images acquired by an RGB camera and converts it to the HSV color space. The HSV color space decomposes color into hue, saturation, and lightness, with the hue component more stably reflecting the essential attributes of color and less affected by changes in illumination. The system calculates the characteristic value (e.g., average value) of this hue component as a key indicator characterizing the fruit's external color. Simultaneously, the system drives a near-infrared spectral sensor to be aimed at the same fruit, acquiring its spectral data within a specific wavelength range (e.g., 900nm to 1700nm). This spectral data is input into a pre-trained prediction model, which analyzes the correlation between spectral features and the actual sugar content of the fruit, outputting a predicted sugar content value as an important indicator reflecting the fruit's internal maturity. Finally, the system compares the external color feature values ​​with a preset range of ripe color thresholds (e.g., for purple passion fruit, the hue is concentrated in a specific angular range when ripe), and the internal sugar content prediction value with a preset ripe sugar content threshold (e.g., an empirical minimum sugar content standard). Only when the color features match the characteristics of ripe fruit and the sugar content prediction value reaches or exceeds the ripening threshold is the fruit's ripeness category determined to be "ripe"; otherwise, it is determined to be "unripe." This combined internal and external determination method, compared to relying solely on color, significantly reduces misjudgments caused by lighting conditions, varietal differences, or slight uneven coloring on the fruit surface.

[0025] For obtaining the fruit's outline dimensions, the system primarily relies on 3D point cloud data acquired by LiDAR. The processing involves: First, from the original point cloud including the background (such as branches, leaves, and sky), algorithms such as spatial clustering and shape matching are used to segment the point set that mainly represents the surface of the target fruit. Then, based on this fruit surface point set, a simplified geometric model that best matches the actual shape of the fruit is constructed using spatial geometric fitting algorithms (such as ellipse fitting, sphere fitting, or a combination thereof). For passion fruit, which approximates an ellipsoid, a spatial ellipse or ellipsoidal model is typically used for fitting. From this fitted model, key dimensional parameters representing the fruit's size, such as its major and minor axis diameters, can be stably extracted as the outline dimensions. Furthermore, the system can approximate the fruit as an ellipsoid defined by its major and minor axes, and calculate its approximate volume using geometric formulas. Multiplying this approximate volume by the average density of the passion fruit (a constant value or a small-range fluctuation value obtained through experimental statistics) allows for the estimation of the fruit's approximate mass. This size and quality information is of great reference value for determining whether the fruit meets commercial specifications and for the subsequent force control of the harvesting agency.

[0026] For identifying surface defects in fruits, the system primarily utilizes high-resolution color images acquired by an RGB camera. The processing typically includes the following steps: First, the original image is preprocessed, such as through noise reduction and contrast enhancement, to improve image quality. Then, image segmentation techniques (e.g., color-, edge-based, or deep learning-based segmentation models) are used to accurately separate the fruit region from the complex background. Within the segmented pure fruit region, the system employs image texture analysis algorithms to detect surface anomalies. One feasible approach is to use a local texture description operator (e.g., a local binary pattern operator) to traverse the fruit surface region. This operator generates an encoding that represents the local texture pattern by comparing the grayscale relationship between each pixel and its surrounding neighboring pixels. The texture pattern distribution of all pixels within the entire region of interest is statistically analyzed to form a histogram or vector representing the overall surface texture features of the fruit. Finally, this texture feature vector is input into a pre-trained classification model (e.g., a classifier built using machine learning algorithms such as support vector machines or neural networks). This classification model learns from a large number of labeled surface texture features of passion fruit (normal, insect-infested, cracked, deformed, etc.), analyzes the input feature vector, and outputs the probability or direct classification result of the fruit belonging to various defect states, thereby determining its surface defect state (such as no defects, insect-infested, cracked, deformed, etc.).

[0027] The precision harvesting execution mechanism 2 is responsible for completing the final physical harvesting action. This mechanism receives various fruit information output from the multimodal intelligent recognition module, including maturity category, outline size, and surface defect status. Before harvesting, the system will compare the received fruit information with preset conditions one by one. These preset conditions may include: (1) maturity is "fully mature", outline size is within the preset qualified size range (e.g., the major axis diameter is above a certain minimum value), and surface defect status is "no defects"; (2) maturity is "not fully mature" but meets the maturity condition of post-ripening after leaving the plant, outline size is within the preset qualified size range, and surface defect status is "no defects"; (3) outline size is less than the preset minimum qualified size, or surface defect status is "defective". Among them, condition (1) corresponds to fruits that can be directly marketed and are not considered for long-term storage. Such fruits are suitable for direct consumption. Condition (2) corresponds to fruits whose size is basically fixed and whose interior has accumulated enough edible components. Because passion fruit is a fruit that undergoes a ripening process after picking, if the passion fruit is not fully ripe when picked (the peel has not fully changed color and the pulp is acidic), the starch in the pulp will gradually be converted into fructose and glucose during the ripening process, the acidity will decrease, the sweetness will increase significantly, and more aromatic substances will be synthesized during the ripening process, making the unique aroma of passion fruit more prominent. This part of the fruit is not fully ripe and is marked as "unripe". Although it can be eaten directly to provide a different flavor, it is still considered unripe. The flavor of fully ripe fruit, for example, is used in the production of sour drinks, but in most cases it can be stored and eaten, and is more suitable for transportation and preservation; condition (3) corresponds to fruit that does not meet the regular market demand, which can be processed into flavored beverages and other products and sold as raw materials. Fruit with skin defects usually does not affect consumption and needs to be removed in time to avoid further consumption of plant nutrients; note that the size mentioned in this part is smaller than the qualified size, which is limited to the case where the fruit size is basically fixed and will not change. The system will analyze whether the fruit maturity has reached the fruit shape determination level. If the fruit is still in the growth period, it does not need to be picked to avoid losses caused by misjudgment. After judging the fruit information, the system will decide whether to pick it. After making the picking decision, the precision picking execution mechanism 2 will be activated, controlling its end effector (such as mechanical claw, cutter, etc.) to move to the position of the target fruit and perform a series of precise grasping and separation (if the stem is cut or broken) actions to separate the qualified fruit from the plant. For fruits that do not meet the conditions (such as completely unripe), the system will ignore them and continue to look for the next target.

[0028] The intelligent dynamic sorting module is responsible for the immediate classification and storage of harvested fruits. This module connects to the precision harvesting execution mechanism 2 via a mechanical interface or conveyor pipe 4, and can receive fruits transported after being separated by the harvesting mechanism. Simultaneously, it receives the category information corresponding to the currently harvested fruit from the multimodal intelligent recognition module via a data link (this information is determined and associated during the harvesting decision). The intelligent dynamic sorting module internally contains multiple storage compartments for storing different categories of fruits (e.g., mature and qualified fruit compartments, unripe fruit temporary storage compartments, defective fruit compartments, etc.). Based on the received fruit information, it determines the appropriate category of the fruit in real time and automatically controls internal mechanisms (such as rotating compartments, levers, slide switches, etc.) to guide the harvested fruit and place it into the corresponding storage compartment.

[0029] By integrating a multimodal intelligent recognition module and comprehensively utilizing visual, spectral, and three-dimensional geometric information, the system achieves comprehensive and high-precision online recognition of passion fruit ripeness, size, and surface defects. This significantly improves the accuracy and reliability of fruit condition assessment and overcomes the limitations of single-sensor information. By setting up a precision harvesting execution mechanism 2 and introducing a qualification comparison decision based on multi-dimensional information, selective automated harvesting is achieved. This ensures that fruits meeting different standards are harvested differently, classifying and screening them from the source, thus guaranteeing the overall uniformity of harvested fruit quality. By introducing an intelligent dynamic sorting module and linking it in real-time with the recognition and harvesting processes, "immediate sorting upon harvest" is achieved, automatically separating and storing different types of fruit. This greatly simplifies subsequent sorting processes and improves the automation and efficiency of the overall workflow. The entire system uses a drone as a mobile platform, integrating recognition, harvesting, and sorting functions into one unit. This enables automated aerial harvesting operations in passion fruit orchards, adapting to the application needs of large-scale orchards with complex terrain, and effectively reducing reliance on high-intensity, high-cost manual harvesting.

[0030] In another technical solution, the process of determining the fruit ripeness category in the multimodal intelligent recognition module specifically includes: Pixels of the centroid region of each fruit are extracted from the color image and converted to the HSV color space. The average value of its hue components is calculated as the color feature value. A near-infrared spectral sensor is driven to collect diffuse reflectance spectra on the same fruit surface to obtain spectral sequences in the wavelength range of 900nm to 1700nm. The spectral sequences are input into a pre-trained partial least squares regression model, which is trained based on the known sugar content of passion fruit samples and outputs the predicted sugar content of the fruit. The obtained color feature value is compared with the preset ripe color threshold range, and the sugar content prediction value is compared with the preset ripe sugar content threshold. If the color feature value falls within the ripe color threshold range and the sugar content prediction value is greater than the ripe sugar content threshold, the ripeness category of the corresponding fruit is determined to be ripe; otherwise, it is determined to be unripe.

[0031] The system locates and focuses on the effective area of ​​the fruit from color images captured by an RGB camera. To avoid interference from fruit edges, stem connections, or reflections from adjacent leaves, the algorithm does not simply average the pixels within the entire fruit bounding box. Instead, it preferentially extracts the pixel set of each fruit's centroid region or the central main surface area. This region typically represents the main color of the fruit surface more stably. Subsequently, the RGB color values ​​of these pixels are converted to the HSV color space. The HSV color space decomposes color information into three independent components: hue, saturation, and lightness. The hue component directly reflects the type of color (e.g., red, yellow, green) and is relatively insensitive to changes in light intensity, making it more suitable for color-based ripeness determination. The system calculates the average hue component of all pixels in the selected centroid region, using it as a key color feature value. This numerical characteristic represents the hue angle of the dominant color on the fruit's surface. For example, for passion fruit, an unripe fruit might correspond to a green hue (hue value in the range of approximately 60-90 degrees), while a ripe fruit might shift to a purple or yellow hue (hue value could be in the range of 0-15 degrees or 340-360 degrees, as well as another specific range, depending on the variety). This conversion and calculation process transforms an intuitive color image into a stable color index that can be used for quantitative comparison.

[0032] Simultaneously or immediately after acquiring a color image, the system drives the probe of a near-infrared spectral sensor to be aimed at the surface of the same target fruit. This sensor actively emits near-infrared light and receives the spectral signals diffusely reflected back from the fruit surface, thereby acquiring a spectral sequence within a specific wavelength range. The selected wavelength range, for example, 900 nm to 1700 nm, covers the characteristic absorption bands related to the chemical bond vibrations of organic molecules such as sugars and water within the fruit. The acquired spectral sequence (i.e., reflectance or absorbance data at a series of wavelengths) constitutes a spectral fingerprint reflecting the internal chemical composition of the fruit. This spectral sequence is then input into a pre-trained partial least squares regression model. This model is a multivariate statistical analysis model that establishes a mathematical correlation (i.e., a calibration model) between spectral characteristics and sugar content by analyzing spectral data from a large number of passion fruit samples with known actual sugar content values. Based on the input spectral sequence of a new fruit, the model calculates its corresponding predicted sugar content value. This predicted value is a continuous numerical value, for example, in "Brix". It objectively quantifies the level of sugar accumulation inside the fruit and is a core internal indicator for judging physiological maturity. Changes in internal sugar and other substances during fruit ripening will alter its near-infrared spectral characteristics.

[0033] The system does not rely solely on color or sugar content. First, it compares the color feature value obtained in the first stage with a preset ripe color threshold range. This threshold range is set based on the statistical range of typical hues of the target passion fruit variety at the ripening stage. For example, for purple passion fruit varieties, the average hue of ripe fruit is mainly distributed within the range of 0-20 degrees (the specific range can be adjusted according to the variety). Second, it compares the predicted sugar content value obtained in the second stage with a preset ripe sugar content threshold. This sugar content threshold is set according to the minimum sugar content standard required for fruit commercialization or the empirical value of optimal edible sugar content, for example, a value between 13 Brix and 15 Brix (such as 14 Brix). Only when the color feature value falls within the preset ripe color threshold range and the predicted sugar content value is greater than the preset ripe sugar content threshold does the system ultimately determine the fruit's ripeness category as "ripe." If the color or sugar content does not meet the standard, or both, it is determined as "unripe." This fusion strategy effectively avoids misjudgments caused by color distortion due to light (such as ripe fruit appearing darker in the shade and being mistakenly judged as unripe) or by special cases such as certain fruits having "high sugar content but poor color change" or "good color change but insufficient sugar content," significantly improving the accuracy and reliability of maturity classification.

[0034] By clearly defining the specific steps for extracting stable hue features from color images, predicting internal sugar content using near-infrared spectroscopy, and employing dual-condition fusion for determination, a multi-level, information-complementary automatic maturity identification scheme was constructed. This scheme combines the external appearance and internal quality of the fruit, overcoming the shortcomings of relying solely on color, which is easily affected by environmental interference or subjective influence, thus making the maturity determination results more objective, accurate, and robust.

[0035] In another technical solution, the process of calculating the fruit outline size and estimating the mass in the multimodal intelligent recognition module specifically includes: The three-dimensional point cloud data of the target fruit was collected using lidar, and the point set representing the fruit surface was segmented from the three-dimensional point cloud data. Based on the obtained fruit surface point set, an ellipse fitting algorithm was used to construct a spatial ellipse model reflecting the shape of the fruit. The major axis diameter and minor axis diameter of the fruit were extracted from the spatial ellipse model as its outline dimensions. The fruit was approximated as an ellipsoid, and the approximate volume of the fruit was calculated based on its major axis diameter and minor axis diameter. The approximate volume was multiplied by the average density of the passion fruit to obtain the estimated mass of the corresponding fruit.

[0036] The system uses an onboard LiDAR to scan the target fruit. The LiDAR emits a laser beam and precisely measures the time it takes for the beam to reflect back from the fruit's surface, obtaining a large number of high-precision spatial distance points. These points are aggregated to form a 3D point cloud of the fruit and its surrounding environment. Each data point contains 3D spatial coordinate information. The initial point cloud not only contains the target fruit but also typically includes information from branches, stems, adjacent fruits, and even distant background points. Therefore, the primary task is to accurately separate the set of points purely representing the target fruit's surface from the complex 3D scene. This process can be achieved through a series of point cloud processing algorithms. For example, first, the spatial range is cropped based on the relative position of the UAV and the fruit; then, a clustering algorithm based on Euclidean distance is used to group spatially clustered points belonging to the same entity into one class; finally, the prior geometric features of the fruit (such as approximate size and ellipsoidal outline) are combined to filter out the point group most likely corresponding to the target fruit from the clustering results.

[0037] Since passion fruit typically approximates an ellipsoid in shape, fitting an elliptical or ellipsoidal model is an efficient and reasonable simplification. The system uses an ellipse fitting algorithm (such as a three-dimensional extension of least-squares ellipse fitting) to calculate the 3D point set obtained from the segmented fruit surface. The goal of this algorithm is to find an elliptical (or ellipsoidal) model in space that minimizes the overall distance error between the model's surface in 3D space and the actual fruit surface point set. Through fitting, the geometric parameters of this spatial elliptical model can be obtained, including its spatial orientation, center position, and the lengths of each principal axis. For passion fruit that approximates a rotating ellipsoid, its major axis diameter (maximum size) and minor axis diameter (the smaller size perpendicular to it) are typically extracted as core parameters describing the fruit's outline dimensions. These two dimensions directly reflect the fruit's length and width, serving as a direct basis for assessing whether the fruit's size specifications meet the standards. This fitting method based on 3D point clouds is more accurate than estimating dimensions solely from 2D images and is unaffected by shooting angle and perspective distortion.

[0038] After obtaining the major and minor axis diameters, the approximate volume of the fruit needs to be calculated to estimate its mass. A common and simple method is to approximate the shape of the passion fruit as an ellipsoid of revolution with the aforementioned major axis diameter as the major axis and the minor axis diameter as the minor axis. Using the volume calculation formula for an ellipsoid of revolution (a known mathematical and geometric formula, not elaborated here), the volume of this approximate ellipsoid can be calculated based on the values ​​of the major and minor axis diameters. Subsequently, the system multiplies this calculated approximate volume by a pre-set average density value of passion fruit to estimate the mass of the fruit. The average density value can be statistically obtained by measuring the actual mass and volume of a large number of samples experimentally, and is a statistically representative constant value or a small-range fluctuation value. The system realizes a complete process from non-contact 3D scanning data to key dimension extraction and single fruit mass estimation.

[0039] This paper presents a non-contact, high-precision, and visually-insensitive method for measuring fruit physical parameters by clearly defining the technical path of acquiring 3D point cloud data using lidar, segmenting the fruit surface point set, fitting a spatial elliptical model to extract dimensions, and estimating quality based on the geometric model. It can accurately obtain the true outline dimensions of the fruit and estimate its quality, providing objective and quantitative 3D geometric evidence for fruit specification selection, grading, and operational data statistics.

[0040] In another technical solution, the process of identifying the state of fruit surface defects in the multimodal intelligent recognition module specifically includes: High-resolution color images of the fruit captured by an RGB camera are acquired and preprocessed, and the fruit regions are segmented. Within the segmented fruit regions, a local binary mode operator is used to extract feature vectors representing the surface texture of the fruit. Specifically: A three-by-three neighborhood is constructed with each pixel as the center. An eight-bit binary pattern code is generated based on the comparison of the gray values ​​of the neighboring pixels and the center pixel. The local binary pattern feature histogram of all pixels in the selected area is used as the texture feature vector. The obtained texture feature vector is input into a pre-trained classification model based on support vector machine. The classification model outputs the probability of the corresponding fruit belonging to various defect states based on the input feature vector, and determines whether the fruit has surface defects and the defect category based on the maximum probability.

[0041] The system acquires high-resolution color images of the fruit captured by an RGB camera. The resolution is sufficient to clearly reveal the subtle textures and imperfections on the fruit surface, reaching 4 megapixels or higher to ensure the ability to distinguish tiny insect holes or fine cracks. After acquiring the raw images, a series of preprocessing operations are performed to optimize image quality and prepare for subsequent analysis. Preprocessing may include: color correction to eliminate color casts caused by different lighting conditions; image denoising to reduce random noise from sensor noise or environmental interference; and contrast enhancement to make the differences between fruit surface features and the background or other features more apparent. The crucial fruit region segmentation step follows. The system accurately separates the main region of the target fruit from the complex background (such as branches, leaves, sky, soil, or other fruits) in the image. This can be achieved through various image segmentation algorithms, such as color threshold-based segmentation (utilizing the difference between the fruit and background in color space), edge detection segmentation (identifying the contour edges of the fruit), or more advanced deep learning-based semantic segmentation models (directly distinguishing the fruit and background at the pixel level). After successful segmentation, a clean image region containing only the surface pixels of the target fruit is obtained. Subsequent texture analysis will be strictly limited to this region, thereby effectively eliminating background interference and improving the accuracy of defect identification.

[0042] The Local Binary Pattern (LBP) operator is used to define a small local neighborhood, such as a common 3-pixel square, centered on each pixel in the image. Then, the gray values ​​(or intensity values ​​of a certain color channel) of the surrounding 8 pixels are compared with the gray value of the center pixel. If the gray value of the surrounding pixels is greater than or equal to the gray value of the center pixel, it is marked as 1; otherwise, it is marked as 0. This results in an 8-bit sequence of binary bits (0 or 1) around the center, i.e., an 8-bit LBP code. This code essentially reflects the texture structure of the local area around the center point, such as flat areas, edges, and corners. Each pixel within the segmented fruit region (or sampled at a certain step size) is traversed, and its LBP code is calculated for each pixel. Then, the LBP codes generated by all pixels are statistically analyzed to form a feature histogram representing the texture distribution of the entire fruit surface. This histogram records the frequency of different LBP patterns throughout the region, converting two-dimensional image texture information into a one-dimensional, quantitative feature vector. This feature vector is robust to changes in illumination and is very sensitive to small, irregular changes in local surfaces (such as holes caused by insects or linear depressions caused by cracks), making it very suitable for defect detection.

[0043] The obtained feature vectors representing the overall texture of the fruit surface are used as input data and fed into a pre-trained classification model based on Support Vector Machines (SVMs). SVMs are a classic machine learning classification algorithm that finds an optimal hyperplane in a high-dimensional feature space to best separate samples from different categories. This classification model has already learned a large number of passion fruit texture feature vectors labeled "normal fruit," "insect-damaged fruit," "cracked fruit," and "deformed fruit" during the training phase, thus mastering the feature space distribution patterns corresponding to various defects. When a new fruit feature vector is input, the model analyzes its position in the feature space and calculates the probability or confidence score of its belonging to each predefined defect category (including the "no defect" category). These probability values ​​reflect the degree of matching between the current fruit texture features and the typical features of each category. The system ultimately determines whether the fruit has a surface defect and its specific defect category based on the category corresponding to the maximum value among these probability values. For example, if the probability of the "insect-damaged" category is the highest and exceeds a certain confidence threshold, it is classified as an insect-damaged fruit; if the probability of the "no defect" category is the highest, it is classified as a normal fruit. This statistical learning-based method can handle the complexity of texture variations and achieve automated, objective defect identification.

[0044] By clearly defining a complete technical chain, including high-resolution image preprocessing and precise segmentation, local binary pattern texture feature extraction, and support vector machine classification model determination, an automated detection scheme specifically designed for subtle defects on fruit surfaces has been constructed. This scheme leverages the high sensitivity of texture analysis to local anomalies, combined with the powerful classification capabilities of machine learning models, to effectively identify minute wormholes, fine cracks, and irregular, deformed textures that are difficult for the human eye to detect immediately, thus significantly improving the automation level and accuracy of surface defect detection.

[0045] The classification model is pre-trained in the following way: Based on historically collected, classified, and verified passion fruit defect information, a dedicated passion fruit defect dataset was constructed, including sample images of normal fruit, insect-damaged fruit, cracked fruit, and deformed fruit. Each image in the dataset was labeled and assigned a corresponding defect category label. During the model training phase, a transfer learning strategy was adopted, using a convolutional neural network pre-trained on a large general image dataset as the base model; this base model was fine-tuned using a passion fruit-specific defect dataset; during the fine-tuning process, an attention mechanism module designed for the characteristics of small-scale defects on the fruit surface was introduced, which focuses on enhancing the model's ability to capture features of insect-eaten holes and fine cracks. The trained model is quantized and compressed, and converted into a lightweight format that can run in real time on the embedded computing platform of the drone.

[0046] The performance of a model largely depends on the quality and scale of the training data. Therefore, a dedicated defect dataset needs to be constructed based on a large number of passion fruit images collected historically and classified and validated manually or rigorously. This dataset needs to be representative, covering various common surface defect types, including images of samples in categories such as "normal fruit" (intact surface), "insect-damaged fruit" (holes formed by insect damage), "cracked fruit" (cracked peel), and "deformed fruit" (abnormal shape, wrinkled, or mottled). Each collected passion fruit image needs to be finely annotated manually or semi-automatically, assigning it an accurate corresponding defect category label. The annotation process goes beyond simple image-level classification, including bounding boxes or delineating defect areas to provide richer supervisory information. Constructing such a high-quality, multi-category dedicated dataset is fundamental to providing reliable learning materials for subsequent model training, ensuring that the model is exposed to sufficiently diverse and realistic defect patterns, and avoiding recognition failures in practical applications due to data bias.

[0047] Training a complex deep learning model (such as a convolutional neural network) from scratch using a small, specialized dataset can easily lead to overfitting and underperformance. Therefore, a transfer learning strategy was adopted. First, a convolutional neural network pre-trained on a large-scale general image dataset (e.g., a massive image dataset containing thousands of object classes) was selected as the base model. Such a pre-trained model has already learned the powerful ability to extract general image features (such as edges, textures, and shapes). Then, this base model was "fine-tuned" using the passion fruit-specific defect dataset constructed above. During fine-tuning, the parameters of the last few layers or all of the network were updated and optimized according to the specific task of passion fruit defect recognition, so that its feature extraction and classification capabilities adapted from the general domain to the specific domain of passion fruit surface defects. To address the characteristics of fruit surface defects (especially insect holes and micro-cracks) which are usually small in scale and have inconspicuous features, a specially designed attention mechanism module was introduced during the model fine-tuning process. This module enables the network to automatically and selectively focus more computational "attention" on image regions that may contain small-scale defect features when analyzing images, thereby enhancing the model's ability to capture and recognize these subtle but crucial features and effectively improving the detection sensitivity of small defects.

[0048] Deep learning models trained on servers can have a large number of parameters and complex computations, making them difficult to deploy directly on embedded computing platforms for drones, where computing resources, storage space, and power consumption are limited. Therefore, lightweight model processing is necessary. Quantization compression is a key technique that converts model parameters (typically 32-bit floating-point numbers) into a lower-precision numerical format (e.g., 8-bit integers). This significantly reduces the storage space occupied by the model and the memory bandwidth and computational load required for inference calculations. After compression and possible other optimizations (such as model pruning, removing redundant connections or layers that contribute little to performance), the model is converted into a lightweight format suitable for efficient operation on edge devices. This ensures that complex defect recognition algorithms can run in real-time on drones, meeting the system's stringent requirements for recognition speed (e.g., processing multiple frames per second). By systematically constructing a dedicated defect dataset, employing transfer learning combined with attention mechanisms for model fine-tuning, and deploying targeted lightweight compression, we can not only fully utilize existing general knowledge and perform precise optimization for specific tasks, significantly improving the model's recognition accuracy and robustness for small-scale defects on passion fruit surfaces, but also ensure the feasibility and real-time performance of advanced algorithms on resource-constrained embedded platforms through effective model compression technology. This provides core algorithmic support for achieving online, real-time, and high-precision defect detection on UAVs.

[0049] In another technical solution, such as Figure 2As shown, the precision picking actuator 2 includes a flexible mechanical claw, which includes three independently driven claw fingers 201 that are symmetrically distributed at 120 degrees. Each claw finger 201 includes an internal skeleton and an external covering layer. The internal skeleton is made of hard alloy, and the external covering layer is medical silicone. Each claw 201 is connected to the same base 202 via a pivot at its base and is driven to rotate around the pivot by an independent micro servo motor. When the three claws 201 are in the initial position, the minimum distance between their inner silicone surfaces is 50mm. When the three micro servo motors drive the claws 201 to rotate to the maximum spread angle in sync, the minimum distance increases to 120mm, thereby forming a gripping range that can accommodate fruits with a diameter of 30mm to 80mm.

[0050] The flexible mechanical gripper employs a stable and efficient "three-finger" symmetrical layout, specifically comprising three identical claw fingers 201 components. These three claw fingers 201 are not fixed as a single unit, but rather independent and non-interfering with each other, and are precisely distributed in a 120-degree circumferential symmetry around a common central base 202. This equiangular symmetrical arrangement ensures that when the three claw fingers 201 simultaneously retract inward, they can uniformly and stably envelop and clamp the nearly spherical or ellipsoidal passion fruit from three directions, forming a three-point force distribution. This effectively prevents the fruit from rolling or slipping due to uneven force distribution and minimizes the eccentric pressure caused by the clamping posture on the fruit.

[0051] Each claw 201 employs a composite structural design, consisting of an internal load-bearing skeleton and an external flexible covering layer. The internal skeleton is the core of the claw 201's mechanical support, possessing sufficient strength and rigidity to transmit driving torque and resist deformation. It is typically made of precision-machined hard alloys (such as high-strength aerospace aluminum alloys or titanium alloys), and its shape can be designed as an arc or a biomimetic structure with joints. The external covering layer tightly covers the outer surface of the skeleton, especially the inner surface that directly contacts the fruit. Its material is medical-grade silicone, which has excellent elasticity, softness, and a high coefficient of friction. It can gently conform to the irregular arc contours of the fruit surface, increasing the contact area and friction. At the same time, its good cushioning properties can absorb the impact energy from clamping and possible minor collisions, thus providing sufficient clamping force while greatly reducing the risk of crushing and scratching the fragile fruit peel.

[0052] Each claw 201 is connected to the central base 202 at its base via a precision rotating shaft mechanism, allowing the claw 201 to rotate within a certain angle range around this axis. Each claw 201 is driven by an independent micro servo motor, meaning that the three claws 201 can receive independent control signals to achieve synchronous or asynchronous movement, thereby completing actions such as opening and closing. In the initial state (i.e., unpowered or standby state), a specific minimum distance is maintained between the inner silicone surfaces of the three claws 201. This initial opening size facilitates the drone's rough positioning when approaching the fruit. When it is necessary to grasp the fruit, the control system calculates the required closing angle based on the fruit size information (such as diameter) provided by the multimodal recognition module, and then drives the three micro servo motors to move synchronously and precisely, causing the claws 201 to close inward. When the motor drives the claws 201 to the predetermined maximum unfolding angle (corresponding to the fully open state of the claws 201), the minimum distance between the inner silicone surfaces of the three claws 201 can be increased to a larger value, such as 120 mm. In this way, by precisely controlling the rotation angle of the servo motor, the gripping range of the mechanical claw (i.e., the range of fruit diameters it can stably grasp) can be continuously or segmentally adjustable from tens of millimeters to over one hundred millimeters, for example, it can adapt to passion fruit fruits with diameters between 40 and 80 millimeters. This design allows a single actuator to cover the main size specifications of the target fruit. By adopting a flexible composite structure design with three-finger symmetrical independent drive, the mechanical claw achieves adaptive envelope and stable, compliant grasping of the fruit. Its application of rigid-flexible materials significantly reduces the risk of mechanical damage to the fruit surface; while the independently controllable drive method combined with the adjustable gripping range gives the actuator a high degree of flexibility and precise control capability, enabling it to reliably complete the task of non-destructive harvesting of fruits of different sizes and positions. It is the core hardware guarantee for achieving the two major goals of "precision" and "non-destructive" in the entire automated harvesting system.

[0053] The precision harvesting actuator 2 also includes a negative pressure auxiliary system, which includes at least one negative pressure nozzle 203 and a miniature vacuum pump. The negative pressure nozzle 203 is made of polyurethane material, with a suction inlet diameter of 30mm and a curvature radius of 15mm at the tip of the suction inlet to conform to the arc-shaped contour of the passion fruit surface. The negative pressure nozzle 203 is connected to the miniature vacuum pump through a flexible pipeline. The negative pressure nozzle 203 is installed at the center of the base 202 of the flexible mechanical claw, and the suction inlet plane of the negative pressure nozzle 203 is lower than the surface of the silicone coating layer on the inner side of the three claw fingers 201.

[0054] The negative pressure auxiliary system mainly consists of a negative pressure suction nozzle 203 that generates suction force and a miniature vacuum pump that provides vacuum power, connected by flexible tubing. The negative pressure suction nozzle 203 is cleverly integrated and installed at the center of the central base 202 of the flexible mechanical gripper. This arrangement ensures that the suction point of the nozzle essentially coincides spatially with the gripping center formed by the three gripper fingers 201. The plane of the suction inlet of the negative pressure suction nozzle 203 (i.e., the end face that contacts the fruit surface for suction) is set slightly lower than the surface of the silicone coating layer inside the three gripper fingers 201. This means that when the drone drives the entire actuator close to the fruit, the suction nozzle will contact the fruit surface before the silicone fingers of the mechanical gripper. This "pre-positioned suction nozzle" design allows the system to first apply a pre-fixing force to the fruit through negative pressure suction before the mechanical claw closes and grips it. This initially and stably "pulls" or "sticks" the fruit to the suction nozzle, thereby correcting any slight positional shifts in the fruit and providing a stable reference for the subsequent closing and gripping of the mechanical claw. This greatly increases the probability of successful harvesting in complex environments (if the fruit is slightly shaky).

[0055] The nozzle body is made of polyurethane, a material that combines good elasticity, abrasion resistance, and a certain degree of food contact safety. Its softness allows it to fit tightly against the fruit surface without damaging the peel. The suction inlet of the nozzle is designed with a specific diameter, such as approximately 30 mm. This size needs to match the typical curvature of the target fruit to ensure sufficient adsorption sealing area, while avoiding air leakage or interference with the movement of adjacent claws 201 due to an excessively large diameter. The tip of the suction inlet is machined into an arc shape with a specific radius of curvature, such as a sphere with a radius of 15 mm or an ellipsoid similar to the curvature of a passion fruit cross-section. This contoured arc design allows the tip of the nozzle to better conform to the natural curved contour of the passion fruit surface. When negative pressure is applied, the flexible material adapts and forms a more effective sealing ring with the fruit surface, reducing air leakage. This results in greater effective adsorption force at the same vacuum level and avoids localized stress concentration in the peel caused by hard edge contact. A miniature vacuum pump serves as the power source and is connected to the nozzle via a flexible tube (such as a PU tube or silicone tube). The flexible tubing allows for relative displacement between the nozzle and pump body during gripping and rotating actions without affecting functionality. In the harvesting process, after the drone hovers near the target fruit and completes final positioning, the control system first activates the micro-vacuum pump, rapidly establishing negative pressure (a pressure below atmospheric pressure) within the nozzle cavity via the flexible tubing. As the actuator continues to approach the fruit, once the tip of the already working nozzle contacts the fruit surface and forms a preliminary seal, the negative pressure "adsorbs" the fruit surface, generating a pre-tightening force perpendicular to the contact surface. Next, with the fruit initially secured by the negative pressure, the control system instructs the three claws 201 of the flexible mechanical gripper to simultaneously retract, encircling the fruit from the sides. At this point, the fruit is firmly held at the end of the actuator by the combined action of the central suction force and the flexible clamping forces from the three lateral directions. Subsequently, the actuator performs a lifting or rotating cutting action, separating the fruit from the stem. Throughout the harvesting and moving process, negative pressure adsorption and mechanical clamping work together to ensure that the fruit will not fall off due to inertia or vibration during the transfer process, until it is safely sent into the conveying pipe 4.

[0056] By using a negative pressure assist system integrated with the center of the flexible mechanical gripper, a pilot-guided, non-contact stabilizing force is added to the simple mechanical gripping. The contour-following flexible design of the suction nozzle enhances its adaptive adsorption and sealing effect on the fruit surface. The coordinated work of this assist system and the mechanical gripper significantly improves the stability and success rate of the harvesting actuator in the initial grasping stage, reducing the risk of grasping failure due to smooth fruit surfaces, dew, or slight misalignment, making the entire harvesting operation smoother and more reliable.

[0057] In another technical solution, such as Figure 3As shown, the intelligent dynamic sorting module includes a rotary sorting bin 3, which is driven by a stepper motor to rotate around its central axis 301. The interior of the sorting bin 3 is divided into multiple independent fruit storage chambers 302 along the circumference. Each fruit storage chamber 302 has an opening at the top to receive a corresponding type of fruit. The top of the sorting bin 3 is provided with a feeding guide, which is connected to a conveying pipe 4 for receiving the picked fruit. Based on the category information of the harvested fruit output by the multimodal intelligent recognition module, the stepper motor is driven by the controller to rotate the sorting bin 3 until the target fruit storage bin 302 corresponding to the current fruit category information rotates to directly below the feed inlet; the fruit falls into the target fruit storage bin 302 through the conveying pipe 4 and the feed inlet.

[0058] The main body of the intelligent dynamic sorting module is a rotary sorting bin 3. Through the physical rotation of the bin, different fruit storage compartments 302 are sequentially aligned with a fixed feeding port, thereby achieving automatic classification and storage of the fruit. The bin is driven by a precision stepper motor, which, through a reduction mechanism or direct connection, drives the entire sorting bin 3 to rotate precisely around its central axis 301. The stepper motor can receive digital pulse signals to control its rotation angle, thus allowing for very precise control of the bin to stop at multiple preset angular positions. The interior of the sorting bin 3 is divided into multiple independent fruit storage compartments 302 along its circumference by sturdy partitions, for example, into three, four, or more compartments. Each compartment is physically isolated and used to store different types of fruit, such as ripe and qualified fruit, unripe fruit, defective fruit, or fruit requiring special treatment. Each fruit storage compartment 302 has an open entrance at the top to receive fruit falling from above. This rotating multi-compartment design replaces a complex multi-channel distributor with simple mechanical motion, resulting in a compact structure that is particularly suitable for space- and load-constrained drone platforms.

[0059] A fixed feed inlet is positioned directly above the sorting bin 3. This feed inlet is a stationary funnel-shaped or tubular structure, with its lower outlet always aligned with the delivery area near the central axis 301 of the sorting bin 3. The upper end of the feed inlet is connected to the fruit release point of the precision picking actuator 2 via a conveying pipe 4. After the actuator completes picking and releases the fruit, the fruit first falls into this conveying pipe 4. The pipe can be made of a lightweight, smooth material, and its inner wall can be designed with a cushioning structure or guide spiral to ensure that the fruit slides smoothly and steadily, reducing collision damage. Under the influence of gravity, the fruit eventually reaches the fixed feed inlet via the conveying pipe 4. At this time, the sorting bin 3 has been pre-rotated into position under the command of the controller, so that the target storage compartment 302 corresponding to the current fruit category is rotated directly below the feed inlet. In this way, the fruit falling from the feed inlet will fall into the designated target compartment without obstruction and accurately. A small gap needs to be left between the feed inlet and the top of the rotating chamber to avoid friction, but at the same time the design must ensure that the fruit does not fall into the wrong chamber.

[0060] During the sorting process, the multimodal intelligent recognition module generates complete category information (such as maturity, size, and defect status) of the fruit being picked, while identifying and deciding on the harvest. When the fruit is successfully picked and enters the conveyor pipe 4, its corresponding category information is transmitted in real time to the system's central controller or the dedicated controller of the sorting module. The controller has pre-stored the fruit category code (i.e., mapping relationship) corresponding to each fruit storage compartment 302. After receiving the current fruit category information, the controller immediately performs a query and comparison to determine which numbered fruit storage compartment 302 (i.e., the target compartment) the fruit should be sorted into. Then, based on the current angular position of the compartment and the required angular position of the target compartment (i.e., directly below the feed inlet), the controller calculates the rotation direction and precise number of steps (or angles) that the stepper motor needs to drive. Subsequently, the controller sends the corresponding pulse sequence to the stepper motor. The stepper motor responds quickly, driving the entire sorting compartment 3 to rotate smoothly, rapidly and accurately rotating and positioning the target compartment below the feed inlet before the fruit has completely fallen from the conveyor pipe 4. Once positioning is complete, the fruit falls precisely, completing a accurate delivery. The system then resets, ready to receive the next fruit's classification information and execute the next rotational sorting. The entire process is fast, accurate, and coordinated, ensuring highly efficient continuous operation. By employing a rotary multi-compartment structure combined with precise stepper motor drive and a fixed feed guide design, a simple, reliable, and space-efficient dynamic sorting device has been constructed. It enables immediate, automatic, and orderly classification and storage of harvested fruit within the limited payload space of a drone, physically separating fruits of different qualities, greatly simplifying subsequent processing steps, and improving the automation level and continuity of the entire harvesting operation.

[0061] The sorting process of the intelligent dynamic sorting module is as follows: The fruit information output in real time by the multimodal intelligent recognition module is transmitted to the system's controller. The controller has pre-stored classification rules for dividing the fruit into multiple categories. The classification rules are based on a comprehensive judgment of the fruit's maturity category, whether the outline size is within the preset acceptable range, and the state of surface defects. The controller compares the received fruit information with the classification rules to determine the category to which the fruit belongs. Based on the preset mapping relationship between the fruit category and each fruit storage compartment 302 in the rotary sorting bin 3, the controller calculates the required rotation angle and controls the stepper motor to drive the sorting bin 3 to rotate by the corresponding angle, so that the target fruit storage compartment 302 mapped to the current fruit category is aligned with the top feed inlet. The picked fruit falls into the target fruit storage compartment 302 through the conveying pipe 4 and the feed inlet.

[0062] The system's controller (which can be a main controller or a separate sorting controller) serves as the key control core of the sorting process, continuously receiving real-time fruit information streams from the multimodal intelligent recognition module. This information corresponds to each freshly picked fruit currently being transported, encompassing multi-dimensional data such as its maturity category (e.g., ripe, unripe), outline dimensions (e.g., major axis diameter), and surface defect status (e.g., no defects, insect damage, cracked fruit, etc.). Simultaneously, the controller's internal non-volatile memory pre-stores one or more sets of configurable classification rules. These rules define how to combine the aforementioned multi-dimensional fruit information to comprehensively determine the final category to which the fruit should belong. The classification rules are based on at least three core dimensions: first, the fruit's maturity category, which is the primary factor determining the fruit's intended use (immediate market or post-ripening); second, whether the fruit's outline dimensions are within a preset acceptable range, which relates to the fruit's commercial specifications; and finally, the fruit's surface defect status, which directly affects the fruit's commercial appearance and shelf life. Rules can be simple AND / OR logic combinations or more complex decision trees or lookup tables. The purpose is to map continuous or discrete identification results to a limited number of sorting categories.

[0063] Once the controller receives all the information for a specific fruit, its internal processing unit begins executing the classification rules. The controller compares each received fruit information entry with the conditions set in the pre-stored rules. For example, the rules might stipulate that ripeness is only further checked if "surface defect status = no defects" and "outline dimensions are within acceptable range." If it is "ripe," it is classified as category A; if it is "unripe," it is classified as category B. However, if "surface defect status = defective (any type)" or "surface defect status = no defects but outline dimensions are unacceptable," it is directly classified as category C regardless of ripeness. Through this logical operation, the controller determines the final category (e.g., category A, B, or C) of the current fruit. Next, the controller needs to convert this logical category into a physical execution instruction. For this purpose, the controller also pre-stores a preset mapping relationship between fruit categories and the numbers of each fruit storage compartment 302 in the rotary sorting bin 3. For example, category A fruits are mapped to warehouse 1 (ripe and qualified warehouse), category B fruits are mapped to warehouse 2 (temporary storage warehouse for unripe fruits), and category C fruits are mapped to warehouse 3 (defective fruit warehouse). Based on the newly determined logical category of the fruit, the controller immediately obtains the physical number of the target fruit storage warehouse 302 by querying this mapping relationship.

[0064] After determining the number of the target fruit storage compartment 302, the controller needs to direct the stepper motor to drive the compartment to rotate to the correct position. The controller knows the current angular position of the sorting compartment 3 (obtained through motor step count recording or initial position sensor), and also knows the absolute angular position relative to the feed inlet for each fruit storage compartment 302 (these positions were calibrated and stored during system initialization). Based on the angular position of the target compartment and the current position of the compartment, the controller calculates the shortest path to be rotated and the corresponding angle value (e.g., 120 degrees clockwise). Since the step angle of the stepper motor is known, this angle value is quickly converted into the number of pulses to be sent. Subsequently, the controller sends control commands containing parameters such as direction, speed, and number of pulses to the stepper motor driver. Upon receiving the command, the stepper motor immediately starts with high acceleration, driving the sorting compartment 3 to rotate smoothly and stop precisely when it reaches the target position. The entire rotation and positioning process needs to be completed in a short time to ensure that the target compartment is accurately positioned while the fruit is sliding down the conveyor pipe 4. Once the fruit falls into the correct storage compartment, one sorting cycle ends, the system status is updated, and preparations are made for the next decision and execution. A logically clear, responsive, and accurate intelligent sorting control system has been constructed. The complex multimodal recognition results from the front end are transformed into clear and executable sorting action instructions, ensuring that each picked fruit is automatically allocated to the correct storage space based on its overall quality. This achieves complete automation and intelligence in the sorting process, significantly improving the level of refined management and overall efficiency of harvesting operations.

[0065] The classification rules pre-stored in the controller are specifically a sorting decision query table. The sorting decision query table uses the fruit maturity category, the outline size qualification mark, and the surface defect status category as joint input conditions, and outputs the corresponding target fruit storage compartment 302 number. Among them, the outline size qualification mark is generated by comparing the long axis diameter of the fruit with the preset qualification range. If it is within the qualification range, it is true; otherwise, it is false. The fruit storage compartment 302 includes a mature qualified compartment, a waiting-to-ripen temporary storage compartment, and a defective fruit compartment. Fruit maturity is checked only when the surface defect status category is no defect and the outline size is qualified. If the maturity category is mature, the output number corresponds to the mature qualified warehouse. If the maturity category is unripe, the output number corresponds to the unripe temporary storage warehouse. When the surface defect status category is defective, or the surface defect status category is no defect but the outline dimension qualified mark is false, the corresponding defective fruit bin number will be output. In the initial state, the sorting decision query table is configured with default settings, and the maturity judgment threshold and size qualification range are dynamically updated through the ground control terminal according to the actual qualified requirements of the harvested fruits.

[0066] The pre-installed sorting decision lookup table in the controller is a data structure that directly maps input conditions to output actions, enabling efficient and unambiguous sorting logic judgments. The lookup table uses three key fruit characteristics as joint input conditions: First, the fruit's "maturity category," a discrete value, such as directly classifying it as "ripe" or "unripe"; second, the "outline size qualification mark," a Boolean value (true or false), which is not the original size data but a simplified conclusion generated after pre-comparison. The generation process for this mark is as follows: the system compares the measured value of the fruit's long axis diameter provided by the multimodal recognition module with a preset size qualification range. This qualification range is usually set according to the target market or variety standards, for example, a long axis diameter between 50 mm and 80 mm. If the measured value is within this range, the mark is "true" (qualified); otherwise, it is "false" (unqualified). The third input condition is the "surface defect status category," which is also a discrete value, and may include "no defects," "insect infestation," "cracked fruit," "deformed," etc. The output of the lookup table is a simple "target fruit storage compartment number," which directly corresponds to a specific physical compartment within the rotary sorting unit 3, such as compartment 1, compartment 2, etc. This structured lookup table approach transforms complex multi-condition logical judgments into fast table lookup operations, significantly improving decision-making speed and reliability.

[0067] The specific logical rules embedded in the sorting decision query table reflect the sorting priorities and comprehensive considerations. The core principle of the rules is: surface integrity and size specifications are the basic thresholds, and maturity is the grading basis. It is clearly stipulated that the system will only further check the "maturity category" attribute when the "surface defect status category" is "no defects" and the "outline size qualified mark" is "true". In this case, if the maturity category is "mature", the query table outputs the number of the corresponding "mature qualified warehouse"; if it is "awaiting ripening", it outputs the number of the corresponding "awaiting ripening temporary storage warehouse". This ensures that only fruits with intact appearance and qualified size will be stored according to their maturity. The rules also stipulate two situations in which the fruit is directly judged as defective, which have higher priority: the first situation is that as long as the "surface defect status category" is "defective" (regardless of whether it is insect infestation, cracking, or deformity), the number of the corresponding "defective fruit warehouse" will be directly output regardless of its maturity and size. The second scenario is that even if the fruit surface is defect-free, if its "outline size qualification mark" is "false" (i.e., the size is too large or too small), it will also be directly classified into the defective fruit bin. This rule design ensures that fruits with significant appearance defects or serious size discrepancies are immediately removed and do not participate in the maturity grading, which simplifies the logic and guarantees the basic quality baseline of commercial fruit.

[0068] Before shipment or operation, the system performs a "default configuration" on the sorting decision query table. This default configuration is based on the general harvesting standards for mainstream passion fruit varieties (such as Tainong No. 3). For example, the default ripeness sugar threshold can be set to 14 Brix, and the default size acceptable range can be set to 5-7 cm. However, to adapt to the differentiated needs of different orchards, varieties, harvest batches, and even different market standards, this decision rule is "dynamically updatable." This function is achieved through a ground control terminal. Orchard operators or agronomists can flexibly adjust key parameters on the software interface of the ground control terminal according to the current specific fruit qualification requirements. For example, they can change the sugar threshold on which the ripeness determination depends from 14 Brix to 13 Brix (e.g., when sugar accumulation is insufficient during the rainy season), or adjust the size acceptable range from 5-7 cm to 6-8 cm (e.g., for large-fruited varieties or specific order requirements). Once these new parameters are transmitted to the controller on the drone via wireless communication, the controller immediately recalculates the "outline size qualification mark" based on the new thresholds and ranges, and updates the judgment criteria in the internal lookup table. This ensures that all subsequent sorting decisions are executed immediately according to the new standards. This mechanism gives the system a high degree of flexibility and adaptability, allowing for rapid response to changing agricultural needs without modifying hardware or complex code. By adopting clear and prioritized sorting decision lookup table rules and enabling them to be dynamically updated via ground terminals, a highly efficient, stable, and flexibly configurable intelligent sorting decision-making scheme is achieved. This scheme ensures that the sorting results meet strict quality stratification requirements with its concise logic, while its dynamic update feature allows a single hardware system to easily adapt to various differentiated harvesting standards and application scenarios, significantly improving the practicality and economy of the entire harvesting system and reducing adjustment costs caused by standard changes.

[0069] It should be noted that although the steps are described in a specific order above, this does not mean that they must be performed in that order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required functionality is achieved. The number of devices and processing scale described herein are for simplification of the invention; applications, modifications, and variations of this invention will be readily apparent to those skilled in the art.

[0070] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A drone-based intelligent passion fruit identification and harvesting system, characterized in that, This includes the drone flight platform and its onboard multimodal intelligent recognition module, precision harvesting actuator, and intelligent dynamic sorting module, among which: The multimodal intelligent recognition module includes an RGB camera, a near-infrared spectral sensor, and a lidar for acquiring multi-source fruit data. The multi-source fruit data includes color images, near-infrared spectral data, and 3D point cloud data of passion fruit. The multimodal intelligent recognition module processes the acquired multi-source fruit data to obtain and output fruit category information, including maturity category, outline size, and surface defect status. The processing includes: Feature data of fruit surface color in HSV color space is extracted from color images. A prediction model of fruit internal sugar content is established based on near-infrared spectral data. By fusing the feature data of HSV color space with the output of the prediction model of fruit internal sugar content, the fruit maturity category is determined. The outline size of the fruit is calculated based on 3D point cloud data. The surface defect status of the fruit is identified based on color images. The precision harvesting execution mechanism compares the various fruit information output by the multimodal intelligent recognition module with the preset qualification conditions, and performs harvesting actions on fruits that meet the conditions; The intelligent dynamic sorting module is connected to the precision picking execution mechanism and receives fruit information output by the multimodal intelligent recognition module. Based on the fruit information, it sorts the fruits picked by the precision picking execution mechanism into different storage bins.

2. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The process of determining fruit maturity category in the multimodal intelligent recognition module specifically includes: Pixels of the centroid region of each fruit are extracted from the color image and converted to the HSV color space. The average value of its hue components is calculated as the color feature value. A near-infrared spectral sensor is driven to collect diffuse reflectance spectra on the same fruit surface to obtain spectral sequences in the wavelength range of 900nm to 1700nm. The spectral sequences are input into a pre-trained partial least squares regression model, which is trained based on the known sugar content of passion fruit samples and outputs the predicted sugar content of the fruit. The obtained color feature value is compared with the preset ripe color threshold range, and the sugar content prediction value is compared with the preset ripe sugar content threshold. If the color feature value falls within the ripe color threshold range and the sugar content prediction value is greater than the ripe sugar content threshold, the ripeness category of the corresponding fruit is determined to be ripe; otherwise, it is determined to be unripe.

3. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The process of calculating fruit outline dimensions and estimating mass in the multimodal intelligent recognition module specifically includes: The three-dimensional point cloud data of the target fruit was collected using lidar, and the point set representing the fruit surface was segmented from the three-dimensional point cloud data. Based on the obtained fruit surface point set, an ellipse fitting algorithm was used to construct a spatial ellipse model reflecting the shape of the fruit. The major axis diameter and minor axis diameter of the fruit were extracted from the spatial ellipse model as its outline dimensions. The fruit was approximated as an ellipsoid, and the approximate volume of the fruit was calculated based on its major axis diameter and minor axis diameter. The approximate volume was multiplied by the average density of the passion fruit to obtain the estimated mass of the corresponding fruit.

4. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The process of identifying the state of fruit surface defects in the multimodal intelligent recognition module specifically includes: High-resolution color images of the fruit captured by an RGB camera are acquired and preprocessed, and the fruit regions are segmented. Within the segmented fruit regions, a local binary mode operator is used to extract feature vectors representing the surface texture of the fruit. Specifically: A three-by-three neighborhood is constructed with each pixel as the center. An eight-bit binary pattern code is generated based on the comparison of the gray values ​​of the neighboring pixels and the center pixel. The local binary pattern feature histogram of all pixels in the selected area is used as the texture feature vector. The obtained texture feature vector is input into a pre-trained classification model based on support vector machine. The classification model outputs the probability of the corresponding fruit belonging to various defect states based on the input feature vector, and determines whether the fruit has surface defects and the defect category based on the maximum probability.

5. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 4, characterized in that, The classification model is pre-trained in the following way: Based on historically collected, classified, and verified passion fruit defect information, a dedicated passion fruit defect dataset was constructed, including sample images of normal fruit, insect-damaged fruit, cracked fruit, and deformed fruit. Each image in the dataset was labeled and assigned a corresponding defect category label. During the model training phase, a transfer learning strategy was adopted, using a convolutional neural network pre-trained on a large general image dataset as the base model; this base model was then fine-tuned using a passion fruit-specific defect dataset. During the fine-tuning process, an attention mechanism module designed for small-scale defects on the fruit surface was introduced. This attention mechanism module focuses on enhancing the model's ability to capture features such as insect holes and fine cracks. The trained model is quantized and compressed, and converted into a lightweight format that can run in real time on the embedded computing platform of the drone.

6. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The precision harvesting mechanism includes a flexible mechanical claw, which consists of three independently driven claw fingers symmetrically distributed at 120 degrees. Each claw finger includes an internal skeleton and an external covering layer. The internal skeleton is made of hard alloy, and the external covering layer is made of medical-grade silicone. Each claw is connected to the same base via a pivot at its base and is driven to rotate around the pivot by an independent micro servo motor. When the three claws are in the initial position, the minimum distance between their inner silicone surfaces is 50mm. When the three micro servo motors drive the claws to rotate to the maximum spread angle in sync, the minimum distance increases to 120mm, thus forming a gripping range that can accommodate fruits with a diameter of 30mm to 80mm.

7. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 6, characterized in that, The precision harvesting mechanism also includes a negative pressure assist system, which consists of at least one negative pressure nozzle and a miniature vacuum pump. The negative pressure nozzle is made of polyurethane material, with a 30mm diameter inlet and a 15mm radius of curvature at the tip to conform to the arc-shaped contour of the passion fruit surface. The negative pressure nozzle is connected to the miniature vacuum pump via a flexible tube and is mounted at the center of the base of the flexible mechanical claw. The inlet plane of the negative pressure nozzle is lower than the surface of the silicone coating layer on the inner side of the three claw fingers.

8. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, The intelligent dynamic sorting module includes a rotary sorting bin, which is driven by a stepper motor to rotate around its central axis. The interior of the sorting bin is divided into multiple independent fruit storage chambers along the circumference. Each fruit storage chamber has an opening at the top to receive a corresponding type of fruit. The top of the sorting bin is equipped with a feeding guide, which is connected to a conveying pipe for receiving the picked fruit. Based on the category information of the harvested fruit output by the multimodal intelligent recognition module, the stepper motor is driven by the controller to rotate the sorting bin until the target fruit storage bin corresponding to the current fruit category information rotates to directly below the feed inlet; the fruit falls into the target fruit storage bin through the conveying pipe and the feed inlet.

9. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 8, characterized in that, The sorting process of the intelligent dynamic sorting module is as follows: The fruit information output in real time by the multimodal intelligent recognition module is transmitted to the system's controller. The controller has pre-stored classification rules for dividing the fruit into multiple categories. The classification rules are based on a comprehensive judgment of the fruit's maturity category, whether its outline size is within the preset acceptable range, and the state of its surface defects. The controller compares the received fruit information with the classification rules to determine the category to which the fruit belongs. Based on the preset mapping relationship between the fruit category and each storage compartment in the rotary sorting bin, the controller calculates the required rotation angle and controls the stepper motor to drive the sorting bin to rotate by the corresponding angle, so that the target storage compartment mapped to the current fruit category is aligned with the top feed inlet. The picked fruit falls into its target storage compartment through the conveying pipe and the feed inlet.

10. The intelligent passion fruit identification and harvesting system based on unmanned aerial vehicles (UAVs) according to claim 9, characterized in that, The classification rules pre-stored in the controller are specifically a sorting decision query table. The sorting decision query table uses the fruit maturity category, the outline size qualification mark, and the surface defect status category as joint input conditions, and outputs the corresponding target fruit storage compartment number. Among them, the outline size qualification mark is generated by comparing the long axis diameter of the fruit with the preset qualification range. If it is within the qualification range, it is true; otherwise, it is false. The fruit storage compartments include mature and qualified compartments, unripe temporary storage compartments, and defective fruit compartments. Fruit maturity is checked only when the surface defect status category is no defect and the outline size is qualified. If the maturity category is mature, the output number corresponds to the mature qualified warehouse. If the maturity category is unripe, the output number corresponds to the unripe temporary storage warehouse. When the surface defect status category is defective, or the surface defect status category is no defect but the outline dimension qualified mark is false, the corresponding defective fruit bin number will be output. In the initial state, the sorting decision query table is configured with default settings, and the maturity judgment threshold and size qualification range are dynamically updated through the ground control terminal according to the actual qualified requirements of the harvested fruits.

Images