Unmanned aerial vehicle biological control method and system based on real-time visual perception and accurate delivery decision

Abstract

The invention provides an unmanned aerial vehicle biological control method and system based on real-time visual perception and accurate delivery decision. The method comprises the following steps: collecting farmland environment data through multiple sensors, and performing low-illumination enhancement on a visible light image; utilizing an improved target detection network to synchronously identify diseases and insect pests and position a putting point; the two-dimensional image, thermal infrared and three-dimensional point cloud features are fused to realize accurate positioning of the target; a visual language model is introduced for cross-modal reasoning, target disambiguation and priority analysis are completed in combination with a task instruction, and a collision-free delivery sequence is generated; and finally, planning an optimal flight path, and realizing precise drug delivery through trajectory optimization and closed-loop control. The intelligent level and the operation precision of pest control in a complex farmland environment are improved.

Description

Technical Field

[0001] This invention relates to the fields of agricultural intelligence and drone application technology, and in particular to a drone-based biological control method and system based on real-time visual perception and precise delivery decision-making. Background Technology

[0002] Traditional agricultural pest and disease control mainly relies on manual inspections and large-scale pesticide spraying, which suffers from problems such as low efficiency, pesticide overuse, environmental pollution, and threats to human health. With the development of precision agriculture, drones are widely used in plant protection operations due to their high efficiency and flexibility. However, existing drone-based pest control technologies still have significant limitations.

[0003] First, the varying lighting conditions in farmland mean that conventional vision systems produce poor image quality under low light, severely impacting the accuracy of subsequent target recognition. Second, pests and diseases are typically small and well-hidden, making it easy for general target detection models to miss or falsely detect them. Furthermore, single two-dimensional visual information is insufficient for precise three-dimensional spatial positioning of application points, and the fusion of multimodal data, such as visible light, thermal infrared, and point clouds, often remains at a simple stitching level, failing to achieve deep feature complementarity. At the decision-making level, existing methods lack semantic understanding of task instructions, making it impossible to perform target disambiguation and intelligent priority planning in complex scenarios, resulting in rigid workflows and an inability to cope with real, dynamic farmland environments.

[0004] Therefore, there is an urgent need for a drone-based biological control method that integrates high-performance perception, multimodal deep fusion, cross-modal semantic understanding, and intelligent decision-making to achieve truly automated, precise, and intelligent pest control. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for biological control using unmanned aerial vehicles (UAVs) that can achieve all-weather, high-precision, and intelligent operation.

[0006] In a first aspect, embodiments of this application provide a method for biosecurity using unmanned aerial vehicles (UAVs) based on real-time visual perception and precise deployment decision-making, the method comprising: The drone carries a multi-sensor system to collect farmland environmental data synchronously and perform low-light enhancement processing on visible light images. The enhanced image is input into the target detection network to simultaneously identify pest and disease targets and detect the release posture, and establish the correspondence between targets and release points. Spatially align the 2D detection results with the point cloud data to achieve precise 3D positioning, and fuse image features, thermal infrared features and 3D spatial features; A visual language model is introduced to perform cross-modal reasoning on multimodal perception data. Combined with task instructions, target disambiguation and delivery priority analysis are performed to generate a collision-free delivery task sequence. The optimal flight path is planned according to the delivery mission sequence, and the precise delivery of the agent is achieved through trajectory optimization and speed control, and a closed-loop feedback control is formed based on real-time environmental data.

[0007] Secondly, embodiments of this application provide a drone-based biological control system based on real-time visual perception and precise deployment decision-making, applied to the drone-based biological control method based on real-time visual perception and precise deployment decision-making as described in the first aspect, the system comprising: The data acquisition and enhancement module is used to synchronously acquire farmland environmental data through the multi-sensor system carried by the UAV and to perform low-light enhancement processing on visible light images. The target detection and matching module is used to input the enhanced image into the target detection network, simultaneously identify pest targets and detect the release location, and establish the correspondence between targets and release points. The 3D perception and fusion module is used to spatially align the 2D detection results with the point cloud data to achieve precise 3D positioning, and to fuse image features, thermal infrared features and 3D spatial features. The intelligent decision-making and planning module is used to introduce a visual language model to perform cross-modal reasoning on multimodal perception data, combine task instructions to perform target disambiguation and delivery priority analysis, and generate a collision-free delivery task sequence. The precision execution and control module is used to plan the optimal flight path according to the delivery mission sequence, achieve precise drug delivery through trajectory optimization and speed control, and form a closed-loop feedback control based on real-time environmental data.

[0008] Thirdly, embodiments of this application provide an electronic device, including: processor; Memory used to store processor-executable instructions; The processor is configured to implement the UAV-based biological control method based on real-time visual perception and precise deployment decision-making as described in the first aspect when executing the instructions.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the unmanned aerial vehicle (UAV) biological control method based on real-time visual perception and precise deployment decision-making as described in the first aspect.

[0010] The beneficial effects of this invention are as follows: 1. Enhanced perception capability in complex environments: The visible light image is enhanced in low light by a physical-guided multi-scale Transformer enhancement network, and combined with thermal infrared and 3D point cloud data, which effectively overcomes environmental interference such as light changes and target occlusion, and significantly improves the perception robustness and target detection accuracy of the system in complex farmland environments.

[0011] 2. Enhanced small target detection and localization accuracy: An improved small target detection network is adopted, embedding a coordinate attention mechanism and a bidirectional feature pyramid, which strengthens the feature extraction capability for small targets such as pests and diseases. Through spatial alignment of two-dimensional detection results with point cloud data and multi-scale voxel feature fusion, sub-meter-level three-dimensional precise localization of pest and disease targets is achieved.

[0012] 3. Achieve intelligent decision-making and precise planning: By introducing a visual language model for cross-modal reasoning and deeply integrating natural language instructions with multimodal perception data, semantic-level target disambiguation and autonomous analysis of deployment priorities are achieved. This enables the generation of collision-free operation sequences, significantly improving the system's intelligence level and decision rationality.

[0013] 4. Optimized operational efficiency and resource utilization: Through global path planning, trajectory smoothing optimization, and speed adaptive control, the UAV's flight maneuvers are made highly efficient and stable. Combined with real-time closed-loop feedback control, the accuracy of pesticide application is ensured, improving pest and disease control while effectively reducing pesticide usage, lowering operational costs, and minimizing environmental impact.

[0014] 5. Forming a complete technological closed loop: From data collection, environmental perception, target recognition, multimodal fusion to intelligent decision-making and precise execution, a complete closed-loop technological system has been formed. The various modules of the system are highly coordinated, providing reliable technical support for achieving all-weather, fully autonomous precision agricultural biological control. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a drone-based biological control method based on real-time visual perception and precise deployment decision-making, provided as an embodiment of this application.

[0016] Figure 2 This application provides an architecture diagram of a drone-based biological control system based on real-time visual perception and precise delivery decision-making.

[0017] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0019] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0020] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Example 1

[0022] Figure 1 This is a schematic diagram of a drone-based biological control method based on real-time visual perception and precise deployment decision-making, provided as an embodiment of this application. Figure 1 As shown, a drone-based biological control method based on real-time visual perception and precise deployment decision-making includes: The S100 uses a multi-sensor system mounted on a drone to simultaneously collect farmland environmental data and performs low-light enhancement processing on visible light images. This step is fundamental to all subsequent processing. Its first function is to provide heterogeneous, synchronous, and spatiotemporally aligned raw data for subsequent multimodal fusion by simultaneously acquiring data from multiple sensors—visible light, thermal infrared, and lidar / GPS-IMU—and specifically to enhance the image quality of visible light images under typical harsh farmland conditions such as early morning, evening, and cloudy days. This ensures the input quality of subsequent target detection modules and overcomes the technical obstacle of detection failure due to insufficient light.

[0023] Specifically, in this embodiment, step S100 includes: The drone, equipped with a visible light camera, an infrared thermal imager, and a GPS / IMU integrated navigation system, simultaneously acquires RGB images, thermal infrared images, and six-DOF pose data of the farmland environment during flight operations. The visible light camera provides high-resolution color and texture information, such as RGB images. The infrared thermal imager provides thermal radiation data unaffected by visible light, crucial for detecting living organisms, such as insects, or areas with temperature anomalies, such as diseased plants. The GPS / IMU provides high-precision six-DOF pose data, an absolutely essential spatiotemporal reference for subsequent spatial alignment and coordinate transformation between two-dimensional images and three-dimensional point clouds.

[0024] The acquired visible light images are input into a physically guided multi-scale Transformer enhancement network for processing, providing an advanced and dedicated low-light image enhancement solution. This network is designed with clearly defined sub-functions: Decomposition networks decouple input images into illumination and reflection components, thereby decoupling image content. Their function is to decompose an image into illumination components, such as overall brightness and shadows, and reflection components, such as the essential color and texture of objects. This physically-based decomposition allows for targeted treatment of illumination problems and detail loss issues that degrade image quality.

[0025] A multi-scale Transformer encoder is used to estimate the illumination distribution and enhance the brightness of the illumination components, solving the problems of global and local uneven illumination. Its function is to process the illumination components, utilize the global attention mechanism of the Transformer to understand the illumination distribution of the entire scene, and perform adaptive, physically-compliant brightness enhancement and smoothing to avoid local overexposure or underexposure.

[0026] The detail restoration module is used to enhance the texture and suppress noise of the reflection component, solving the problems of detail loss and noise amplification during the enhancement process. Its function is to process the reflection component, enhance the texture edge information that is obscured by weak light, and suppress noise amplified in dark areas, ensuring that the enhanced image has clear details and a high signal-to-noise ratio.

[0027] The reconstruction network is used to fuse the processed illumination and reflection components, outputting an enhanced visible light image, thus achieving information fusion and output. Its function is to resynthesize the processed and optimized illumination and reflection components into a high-quality, visually natural enhanced image.

[0028] The enhancement network is optimized during the training phase using an adaptive illumination-aware loss function. This loss function combines illumination smoothing constraints, color fidelity, and structural similarity metrics to ensure that the enhanced image maintains natural colors and realistic details while increasing brightness. This ensures the quality and realism of the enhancement effect from the optimization objective level. The design of this loss function is crucial to ensuring the network learns from good images; each term has a specific function: the illumination smoothing constraint term... This prevents unnatural lighting abrupt changes and artifacts from appearing in the enhanced image. Its function is to constrain lighting changes after enhancement to be smooth, conforming to the continuity of a natural scene. Color fidelity item. To prevent color distortion in images, a loss is calculated using the Lab color space, which is more consistent with human visual perception. This ensures that the colors of objects remain consistent before and after enhancement, avoiding color shifts. (Structural similarity metric) Preserving the structural information of an image. Its function is to force the enhanced image to remain similar to an ideal reference image or itself in the case of unsupervised operation in terms of local structure, brightness, and contrast, avoiding blurring or distortion of details.

[0029] Specifically, the adaptive illumination sensing loss function is defined by the following formula:

[0030] in The illumination smoothing constraint term is defined as the gradient difference between the illumination components before and after enhancement.

[0031] in Gradient operators for images, such as Sobel and Prewitt, are used to extract changes in the horizontal and vertical directions of an image. For enhanced, or reconstructed images. This is the original, unprocessed image. This represents the square of the L2 norm over all pixels.

[0032] For color fidelity, color difference is calculated using the Lab color space:

[0033] in CIELAB, or Lab color space for short, has three channels. Lightness: The value ranges from approximately 0 to 100, representing the brightness or darkness of an image. The green-red axis has positive values ​​biased towards red and negative values ​​biased towards green. : Blue-Yellow axis, positive values ​​are biased towards yellow, negative values ​​are biased towards blue. enh: The corresponding channel in Lab space for the enhanced or reconstructed image. org: The corresponding channel in Lab space for the original, unprocessed image. The image obtained after enhancing the corresponding channel in Lab color space. : The original image of the corresponding channel in Lab space.

[0034] The structural similarity measure is calculated based on the structural information of local image patches:

[0035] in, These are weighting coefficients, with values ​​of 1.0, 0.8, and 0.5 respectively. To enhance the image, The image is the original image, and SSIM(·) is the structural similarity calculation function.

[0036] The overall function of the above steps is to construct an end-to-end low-light image enhancement subsystem based on a combination of physical priors and deep learning, with a clear optimization objective. It does more than simply brighten images; rather, through a decoupling, separate optimization, and resynthesis approach, it systematically addresses multiple problems inherent in low-light images, such as insufficient brightness, loss of detail, significant noise, and color distortion. This provides high-quality, highly reliable input for downstream object detection tasks, fundamentally improving the robustness of the entire system under adverse lighting conditions.

[0037] S110 inputs the enhanced image into the target detection network, simultaneously identifying pests and diseases and detecting their application posture, establishing a correspondence between targets and application points. The core function of this step is precise target identification and location matching. It not only needs to outline the pests and diseases in the image and identify the targets, but also simultaneously find the most suitable precise location for pesticide application, such as the entrance to the insect nest or the center of a diseased leaf. Ultimately, by establishing a one-to-one correspondence, the abstract target is transformed into a concrete, actionable point, providing direct evidence for precise action.

[0038] Specifically, in this embodiment, in step S110: The enhanced image is input into an improved small target detection network, which uses EfficientNet-B0 as a shared backbone. It achieves multi-scale feature fusion through a bidirectional feature pyramid network and embeds a coordinate attention mechanism to enhance the feature representation capability for small targets such as pests and diseases. Specifically, the EfficientNet-B0 backbone network provides efficient and powerful basic feature extraction capabilities, achieving a good balance between computational efficiency and feature richness. The bidirectional feature pyramid network enables multi-scale feature fusion. Its function is to fully combine the semantic information of deep networks, helping to identify what something is and the high-resolution detail information of shallow networks, aiding in location and significantly improving the detection capability of small targets, preventing them from being submerged in deep feature maps. The coordinate attention mechanism enhances the perception of the spatial location of small targets. This is one of the core innovations of this step. Ordinary channel attention ignores positional information, while coordinate attention, through global pooling along height and width, encodes spatial coordinate information into channel descriptors, enabling the network to more accurately focus on the precise region where the small target is located, greatly improving feature representation capability.

[0039] The coordinate attention mechanism achieves spatial information modeling in the following ways: First, global pooling is performed along the height and width directions respectively to obtain orientation-aware feature descriptors. Then, long-range dependencies in the vertical and horizontal directions are captured to generate a pair of orientation-aware feature maps.

[0040]

[0041]

[0042] in, , The first Global feature descriptors for the channel in the height and width directions. , The first The passage is in the location , The input feature map is located at h, and w is the height index (H) and width index (W), respectively.

[0043] The pooled features are then concatenated and convolved, and attention weights are generated using sigmoid activation. This process interacts and integrates feature information from both directions, resulting in the learning of a comprehensive spatial attention map.

[0044] in, It is the Sigmoid activation function. For convolution operations, This involves concatenating feature descriptors along the height and width directions; Sigmoid activation generates attention weights, normalizing the integrated features to weights between 0 and 1, representing the importance of each position.

[0045] Finally, the attention weights are multiplied by the original feature map to enhance the feature response of the target region. Specifically, by multiplying with the original feature map, the feature map is recalibrated to strengthen the feature response of the target region while suppressing irrelevant background noise.

[0046] When establishing the correspondence between targets and application points, an optimized matching method is adopted. Its weighted scoring function is defined as follows: This addresses the matching problem of selecting the most suitable and unique application point for each pest or disease target, and introduces a more intelligent matching criterion.

[0047] in Represents the target bounding box With the placement box The intersection and union ratio, For the confidence level of delivery quality, α is an adjustable balancing parameter. A comprehensive matching evaluation criterion is established through a weighted scoring function S. It no longer relies solely on the overlap (IoU) of the bounding boxes, but introduces the confidence level of delivery quality. . The probability of successful delivery at a delivery point can be predicted based on image features around the delivery point, such as whether it is flat or whether there are obstacles.

[0048] By employing a locally optimal selection strategy, targets are sorted from highest to lowest weighted score, and the best delivery location box is matched sequentially for each target bounding box, ensuring a precise one-to-one correspondence. Matching conflicts are resolved to ensure that each target is assigned a delivery point without duplication. Stable and executable matching output is achieved. A greedy algorithmic strategy prioritizes binding the best match, such as the highest-scoring target-delivery point pair, then removes the matched option and binds the next optimal pair. This guarantees that each instruction in the final generated delivery task sequence is explicit and conflict-free.

[0049] Through the above steps, a target recognition and localization subsystem with more accurate detection and better matching is constructed. It improves the sensitivity of small target detection by introducing a coordinate attention mechanism, utilizes multi-scale contextual information through a bidirectional feature pyramid, and finally, introduces a weighted IoU matching strategy based on delivery quality factors. This ensures that the system not only finds the target but also identifies the most spatially suitable and operationally feasible delivery point for that target, laying a solid foundation for subsequent accurate execution.

[0050] S120 spatially aligns the 2D detection results with point cloud data to achieve precise 3D positioning, and fuses image features, thermal infrared features, and 3D spatial features. This step achieves two key functions: first, precise spatial positioning, converting 2D pixel coordinates on the image into 3D coordinates in the real world, allowing the UAV to know the exact location of the target in 3D space; second, deep fusion of multimodal features, by fusing visual appearance such as images, thermal radiation characteristics such as thermal infrared, and geometric structure such as 3D spatial information, to generate a more comprehensive and robust integrated perception feature, thereby overcoming the limitations of a single modality. For example, visually camouflaged targets may appear abnormal in thermal infrared or geometric shape.

[0051] Specifically, in this embodiment, in step S120: Based on the real-time pose of the UAV and the intrinsic and extrinsic parameters of the sensors, the 2D detection box obtained in S2 is projected onto the point cloud coordinate system through a coordinate transformation matrix, establishing the correspondence between image pixels and the 3D point cloud. This achieves the coordinate transformation from 2D pixels to the 3D world, providing a foundation for precise positioning. This is the premise of the entire process. Its function is to utilize the pose data provided by GPS / IMU and the sensor parameters obtained through calibration, such as intrinsic and extrinsic parameters, to project each identified target pixel in the image onto the 3D point cloud coordinate system through a coordinate transformation matrix. This directly transforms the abstract insect in the image into a target point in space with real 3D coordinates (X, Y, Z), providing a unique spatial target point for the UAV's precise flight and deployment.

[0052] A multi-scale voxelization strategy is employed to construct a hierarchical spatial feature representation, including a coarse-grained voxel layer for extracting global spatial context information, a medium-grained voxel layer for preserving key geometric features, and a fine-grained voxel layer for capturing local detail features. This addresses the problem of sparse, irregular 3D point cloud data containing multi-scale geometric information by transforming it into a regular, hierarchical feature representation. An adaptive voxelization strategy is used to construct a multi-scale voxel feature pyramid, specifically including: Coarse-scale voxel layer: Using a voxel size of 0.1m × 0.1m × 0.1m, this layer extracts global spatial context features to understand the macroscopic layout of the scene, such as the overall terrain of the target area and the height of the crop canopy, providing global spatial context for perception. Mesoscale voxel layer: Using a voxel size of 0.05m × 0.05m × 0.05m, this layer preserves key geometric features, capturing the overall shape and main structure of the target, such as the outline of a diseased plant or the approximate volume of an insect nest. Fine-scale voxel layer: Using a voxel size of 0.02m × 0.02m × 0.02m, this layer captures fine local details, describing the target's fine geometric features, such as the curling of leaf edges and the precise shape of insect holes. This hierarchical representation ensures that geometric information from macroscopic to microscopic levels is effectively preserved and utilized.

[0053] This paper employs a cross-modal attention fusion mechanism to adaptively weight and fuse features from visible light images, thermal infrared images, and 3D spatial features. A multi-head attention architecture is used to process feature information from different modalities in parallel, achieving intelligent and adaptive fusion between these modalities and generating a complementary enhancement effect. This is the core and innovative aspect of this step. It includes: First, layer normalization is performed on the features of each modality, and then the cross-modal attention weights are calculated:

[0054] Where Q is the query vector, K is the key vector, and V is the value vector, each derived from a feature representation of a different modality. The cross-modal attention mechanism allows a feature from one modality to query and focus on relevant important information in other modalities. For example, a blurry target in the visible light feature can be identified by focusing on the high-temperature region in the thermal infrared feature and the protruding geometric structure at that location in the 3D feature, thus comprehensively determining whether it is a real target.

[0055] By employing a multi-head attention architecture, the relationships between modalities are learned in parallel from different representation subspaces. This can be understood as simultaneously focusing on multiple correlation patterns such as color-geometry and texture-temperature, resulting in richer and more robust fused features. Adaptive weighting dynamically evaluates the contribution of different modalities in different scenes and locations. For example, visible light is trusted in well-lit conditions, while thermal infrared is relied upon more at night; for oddly shaped targets, 3D geometric features are more heavily relied upon.

[0056] The final result is a comprehensive perceptual feature map that integrates multi-source information. This feature map simultaneously includes the target's visual appearance, thermal radiation characteristics, and 3D geometric features, providing perceptual data with spatial context for subsequent intelligent decision-making. Specifically, multiple sets of attention weights are computed in parallel using a multi-head attention mechanism and then fused to generate a comprehensive perceptual feature map with spatial context. This feature map simultaneously includes visual appearance, thermal radiation characteristics, and 3D geometric information. The output is a complete joint feature representation that can be used for high-level decision-making. This comprehensive perceptual feature map is no longer a single image or point cloud, but a super feature vector integrating color, texture, temperature, precise 3D position, and shape. It provides an unprecedented perceptual data foundation rich in spatial context for subsequent visual language models to make inference decisions, making decisions, such as identifying a high-risk disease that needs to be prioritized, more accurate and reliable.

[0057] Through the above steps, an advanced sensing subsystem capable of 2D-to-3D conversion and multimodal deep fusion is constructed. It solves the localization problem through spatial registration, addresses the structuring of 3D data through multi-scale voxelization, and ultimately achieves highly intelligent information fusion through a cross-modal attention mechanism. This generates a more comprehensive and accurate joint sensing result that describes the physical world, serving as a solid bridge for the entire invention from perception to cognition and decision-making.

[0058] S130 introduces a visual language model to perform cross-modal reasoning on multimodal perception data, combining it with task instructions for target disambiguation and deployment priority analysis to generate a collision-free deployment task sequence. This is the core of the invention's intelligence. Its functions include: Cross-modal reasoning: understanding task instructions, such as prioritizing the semantics of the severe aphid infestation in the northwest corner, and combining it with the perceived specific scene. Target disambiguation: distinguishing between different target instances with similar appearances to avoid duplicate operations or incorrect targets. Priority analysis: comprehensively considering factors such as target severity, spatial location, operational efficiency, and safety to determine which target to process first. Sequence generation: outputting an ordered, collision-free sequence of operational instructions, making the drone's actions logical and safe.

[0059] Specifically, in this embodiment, in step S130: S1301. Introduce a visual language model as the core inference engine, aligning multimodal perceptual features with natural language task instructions across modalities to achieve semantic-level understanding of pest and disease targets; endowing the system with advanced cognitive abilities of understanding and thinking. This step is the core architecture of the decision-making layer.

[0060] By using a visual-language model as the inference engine, a pre-trained large model capable of simultaneously processing visual and linguistic information is provided. This model has already learned a wealth of world knowledge, such as that aphids are green and clustered, and that mold causes spots on leaves, as well as logical relationships. Through cross-modal alignment, the multimodal perceptual features generated in the above steps, such as visual, thermal, and geometric features, are mapped to the same semantic space along with user-input natural language task instructions, such as prioritizing the treatment of severe corn borer infestations on the east side. This enables the system to understand which visual targets and spatial regions correspond to the words "east side," "corn borer," and "severe" in a specific perceptual scenario, thus achieving a leap from pixel-level perception to semantic-level understanding.

[0061] S1302. A unique identifier is assigned to each detected target through a visual tagging mechanism. Combining the target's visual features, spatial location, and semantic description, similarity calculation is performed in a unified embedding space to resolve the ambiguity in identifying similar targets. This addresses the problem of identifying and distinguishing which target is which. This is crucial for achieving accurate operation and preventing duplicate operations on the same target or confusion between different targets. Specifically, a unique digital identity is assigned to each detected target instance through a visual tagging mechanism. Similarity calculation in the unified embedding space comprehensively judges the similarity between two targets. It compares not only appearance, such as visual features, but also location, such as spatial location, and semantic information, such as semantic description. For example, even if two targets look similar, if they are spatially far apart, they will be judged as different instances. This effectively solves the technical challenge of distinguishing multiple similar-looking pests and diseases in complex scenarios.

[0062] S1303. Based on spatial constraints and mission requirements, and comprehensively considering the semantic importance, spatial accessibility, and operational safety factors of the target, a deployment priority evaluation model is established. A multi-factor decision-making model is constructed to intelligently determine the order of operations. This goes beyond simple left-to-right or near-to-far rules. Semantic importance is used to assess the urgency of the target itself; for example, targets deemed severely damaged by the VLM (Vehicle Modeling Unit) have higher priority than those with minor damage. Spatial accessibility is used to assess whether the path for the UAV to reach the target is smooth and whether energy consumption is minimized. Operational safety is used to assess whether there are risks associated with operating near the target, such as near a utility pole. The function of this evaluation model is to quantify these factors and calculate a comprehensive priority score, providing a basis for generating the optimal task sequence.

[0063] Specifically, based on prior knowledge of the target shape and constraints of orientation perception, a deployment feasibility assessment model is established:

[0064] in For semantic matching degree, Score the spatial accessibility. To score the safe distance, The weighting coefficients are used to quantify the feasibility of each deployment location by combining semantic matching, spatial accessibility, and safe distance through target shape priors and orientation perception constraints.

[0065] The calculation of semantic similarity includes: S1.1 Target Shape Prior. Shape Template Library: Pre-collect binary masks of common pests, crop parts, or target objects, such as circles, ellipses, and elongated shapes, and encode them using shape descriptors, such as Hu moments, Fourier descriptors, and ShapeNet features from deep learning. Deployment Region Extraction: Use segmentation networks, such as U-Net or DeepLabV3+, on the enhanced image to obtain masks of candidate deployment regions.

[0066] S1.2 Semantic similarity calculation. Through feature extraction, vectors are obtained for both the target mask and the candidate region mask using shape descriptors. and Similarity measures are obtained by normalizing using cosine similarity or Euclidean distance.

[0067] The value ranges from [0,1], with a larger value indicating a better shape match.

[0068] S1.3 Integrate semantic context. If semantic segmentation labels exist, such as leaves, branches, and soil, the similarity can be multiplied by the label matching coefficient. This further enhances semantic consistency.

[0069] Spatial accessibility score The calculations include: S2.1 Direction-Aware Constraints. Direction-Aware Features: Global pooling (Coordinate Attention) is performed on the feature map along the height and width directions to obtain the direction-aware vector. , Reachability vector: Generates a reachability heatmap by combining the kinematic model of the robot / robotic arm. , which represents the reachability probability from the current pose to each pixel.

[0070] S2.2 Accessibility Score Calculation:

[0071] Where C is the set of pixels in the candidate delivery area, and if A is binary (e.g., reachable / unreachable), then... Let A be the reachability ratio. If A is a probability, then it is... Average achievable confidence level.

[0072] S2.3 Path Planning Constraints. Paths are planned in the 3D workspace using RRT* or A*. If the path length L is less than a threshold... An additional reward will be given: .

[0073] Safe distance rating The calculations include: S3.1 Safety Distance Model: Safety Radius Minimum safe distances are set based on the object of the operation, such as seedlings, mature crops, and the machinery being operated. (Distance diagram) : Calculate the Euclidean distance to the nearest obstacle, such as other plants, supports, or the robot itself, for each pixel.

[0074] S3.2 Scoring Function:

[0075] in, Use a Sigmoid or ReLU normalization function such that when the actual distance ≥ hour, A score close to 1 indicates safety; when the distance is insufficient, the score drops rapidly.

[0076] S3.3 Dynamic Safety Constraints: If there are moving obstacles in the working environment, such as walking robots or branches blown by the wind, a predicted safety margin can be added to D, based on the prediction of speed and acceleration.

[0077] By integrating environmental observation information, task instructions, spatial constraints, and a predefined skill library through a structured Prompt template, the visual language model is guided to generate a collision-free delivery subtask sequence that conforms to physical constraints, with the output format being an ordered chain of operation instructions.

[0078] S1304. By integrating environmental information, task instructions, and operational constraints through structured prompt templates, a collision-free task delivery sequence conforming to the rules of the physical world is generated. The reasoning results are transformed into a standardized and safe set of operational instructions that can be directly executed by downstream systems. This is the final output of the decision-making process.

[0079] Specifically, a structured prompt template is set up to standardize the VLM's thought process and output format. Like a fill-in-the-blank exercise or checklist, it guides the VLM to systematically consider the environmental context, mission objectives, and safety constraints, ensuring that the generated plan is comprehensive and reliable. It generates a collision-free deployment mission sequence, producing an ordered list of instructions, such as sequence number 1: fly to target A1 (coordinates XYZ) and deploy; sequence number 2: fly to target B3 (coordinates XYZ) and deploy. This sequence ensures that the UAV will not experience path conflicts while traversing between targets, nor will it repeatedly visit the same area, thus achieving efficient and safe automated operation.

[0080] The above steps construct an intelligent decision-making brain with semantic understanding, instance differentiation, multi-factor evaluation, and standardized planning capabilities. This upgrades the unmanned aerial vehicle (UAV) system from a passive perception-execution machine into an intelligent agent capable of understanding complex instructions, distinguishing individual objectives, weighing pros and cons, and autonomously generating safe and efficient operational plans. This is a key breakthrough in achieving fully autonomous and intelligent agricultural operations.

[0081] Specifically, step S140 further includes: constructing a multimodal task understanding and planning framework based on a visual language model, which includes the following processing mechanisms: Task semantic parsing and embodiment mapping mechanisms are used to deconstruct abstract natural language instructions into specific task objectives associated with the current perceived scene. This addresses the problem of how to translate human language instructions into machine-understandable action goals bound to a specific environment. This is the starting point for intelligent decision-making. Specifically, task semantic parsing understands the keywords, intent, and logical structure within the instructions. For example, parsing the core elements such as "priority," "severe," and "aphids" in "prioritize handling severe aphid infestations." Embodiment mapping then places these parsed abstract elements into the current perceived scene. It maps the abstract concept of "severe aphid infestation" to specific infested areas on certain plants identified by the VLM from the camera footage that match the characteristics of "severe" and "aphids." This mechanism ensures that instructions are no longer isolated text but rather executable goals closely linked to the physical world.

[0082] A cross-modal instance disambiguation mechanism based on visual tags achieves accurate differentiation of similar-looking targets in a unified semantic space by assigning unique identifiers to detected targets. This step addresses the problem of establishing and maintaining a unique identity for each object in complex scenes. This is the foundation for achieving accurate job operations. Its core function is identity management. By assigning a unique visual tag to each target instance and comprehensively comparing its appearance, location, and semantics in a unified space, the system can continuously track each target, even those that are very similar in appearance. This prevents the system from misidentifying two different but similar-looking pest or disease points as the same, thus avoiding missed or duplicate job operations.

[0083] A spatial-semantic collaborative prioritization reasoning mechanism integrates the semantic importance, spatial accessibility, and operational safety of the target to generate a task sequence that conforms to physical constraints. It addresses the problem of how to intelligently determine the order of tasks by considering multiple factors, much like an experienced agronomist. This is the core reasoning process of decision-making. It explicitly identifies four key factors considered in the reasoning: Semantic importance: the intrinsic value or urgency of the target; Spatial accessibility: the efficiency and feasibility of physical paths; Temporal dependence: whether certain tasks have a sequential order, such as processing the upwind area before the downwind area; and Safety constraints: avoiding any dangerous operations. Its function is to execute a multi-objective optimization decision, outputting a task sequence that is optimal in terms of efficiency, effectiveness, and safety.

[0084] The structured prompting engineering-driven planning generation mechanism integrates environmental context, task specifications, and skill knowledge through hierarchical prompt templates, outputting a standardized, executable task sequence. It addresses the problem of how to standardize and guide large models to produce reliable and standardized outputs. This is the guarantee for transforming intelligent reasoning into reliable action. Specifically, structured prompting engineering designs standardized processes or thought frameworks for interacting with large models, ensuring that each decision systematically considers all necessary information, avoiding arbitrariness or illusions inherent in the large model. Through hierarchical prompt templates, input information is organized modularly, for example: Environmental context layer: providing the current scene perception results; Task specification layer: clarifying user instructions and operational requirements; Skill knowledge layer: embedding domain knowledge, such as drone flight specifications and pesticide spraying common sense. The ultimate function of this mechanism is to generate a task sequence with a fixed format, complete content, no ambiguity, and compliance with safety standards, ensuring that downstream execution systems can parse and execute it without errors.

[0085] This paper defines a complete and robust intelligent decision-making system at the architectural level. It decomposes the complex decision-making process into four core mechanisms: an understanding mechanism, task semantic parsing and embodiment mapping, a differentiation mechanism, cross-modal instance disambiguation based on visual tags, a decision-making mechanism, spatial-semantic collaborative priority reasoning, an output mechanism, and structured prompting engineering-driven plan generation. These four mechanisms work together to ensure the automation and intelligence of the entire process, from receiving ambiguous human instructions to generating accurate, reliable, and executable machine work plans.

[0086] Specifically, the mechanisms within the multimodal task understanding and planning framework include: The task semantic parsing and embodiment mapping mechanism parses key semantic elements in instructions through a pre-trained language understanding module, mapping abstract task descriptions to specific perceptual features and operational requirements. This clarifies the specific technical path for achieving task understanding. This step decomposes the mechanism into executable algorithm modules. Specifically, the pre-trained language understanding module provides powerful, readily available natural language processing capabilities for accurately extracting entities from instructions, such as aphids, attributes like severity, actions like processing, and relationships like priority. By mapping to perceptual features and operational requirements, a bridge between language, perception, and action is established. It transforms the parsed semantic elements into queries to the perception system; for example, mapping severity to a specific temperature threshold in a thermal infrared image or a specific texture feature in a visible light image, and requirements for the execution system; for example, mapping processing to administering X milliliters of a specific agent. This transforms abstract instructions into specific filtering conditions and action parameters for the current scenario.

[0087] The visually labeled cross-modal instance disambiguation mechanism achieves accurate target recognition in complex scenes by calculating the similarity of visual features, text descriptions, and spatial locations in a unified embedding space. It reveals the core algorithm behind instance disambiguation—multi-dimensional similarity calculation. This step clarifies how to utilize the unified embedding space. The unified embedding space provides a mathematical space that can jointly measure visual, textual, and spatial data. In this space, different types of data, such as image patches, words, and coordinates, can be represented as vectors, and the distance or similarity between them can be calculated. Comprehensive identity determination is performed by calculating multi-dimensional similarity. The system determines that two targets with similar visual features are different instances if their spatial locations are far apart. A target with a higher similarity to its text description, such as an aphid on a rolled leaf, is identified as the instance. This achieves more robust target identification using semantic and spatial information beyond the pixel level.

[0088] The spatial-semantic collaborative priority reasoning mechanism comprehensively considers semantic importance, spatial accessibility, temporal dependence, and security constraints to generate an optimal task sequence. It explicitly lists all core dimensions that must be considered in priority decision-making. This step is an extension and refinement of the aforementioned mechanism. It explicitly adds temporal dependence to the existing semantic importance, spatial accessibility, and security constraints. This feature is added to handle tasks with a logical sequence. For example, when wind direction changes, tasks should be performed upwind first, then downwind, to prevent drones from passing through the pesticide fog; or the source of infection should be treated first, then the infected area. This reflects the decision-making system's adherence to the rules of the complex real world, making the generated plan more logical and practical.

[0089] The structured prompting engineering-driven planning generation mechanism ensures the generation of task sequences that can be directly parsed and executed by downstream systems by designing a hierarchical template structure comprising an environmental context layer, a task specification layer, a skill knowledge layer, and an output specification layer. The internal hierarchical design of the structured prompts is concretized to ensure the reliability and usability of its output. This step defines the fuzzy template concept as a four-layer architecture with clear input-output relationships. Specifically, the environmental context layer inputs the current world state to the VLM, i.e., where we are now and what we see. Input: multimodal perception features. The task specification layer clarifies the task objectives and constraints to the VLM, i.e., what we need to do and what the requirements are. Input: parsed user instructions. The skill knowledge layer injects domain expertise and operational rules into the VLM, i.e., how we should do it, what we can do and what we cannot do. Input: built-in UAV operation manual and agronomic knowledge base. The output specification layer strictly defines the response format that the VLM must follow, i.e., please answer according to this format. For example: JSON format, containing fields such as sequence_id, target_id, 3d_coordinates, and action. The functions of these four layers work together to ensure that the inference process of VLM is based on evidence and reason, and that its output is structured, machine-readable, unambiguous, and can be seamlessly invoked by the precise execution and control modules.

[0090] S140: Based on the delivery task sequence, plan the optimal flight path, achieve precise drug delivery through trajectory optimization and speed control, and form a closed-loop feedback control based on real-time environmental data. This step is the final execution stage of the method. Its functions include: Path planning and trajectory optimization: Transform the task sequence into an energy- and time-optimal and smooth flight path to ensure flight efficiency and stability. Speed ​​control: Perform precise speed management in different sections of the path, especially at curves, to ensure optimal attitude and speed at the delivery point. Closed-loop feedback control: Solve dynamic adaptability problems. By monitoring the environment in real time, such as gusts and system status, dynamically adjust the flight trajectory and delivery timing, compensate for various disturbances, ensure the accuracy and robustness of the final delivery, and form a complete closed loop from perception to decision-making to execution.

[0091] Specifically, in this embodiment, in step S140: Based on the delivery sub-task sequence generated by S4, and considering the UAV's current pose, remaining battery power, payload status, and operational area constraints, an improved A* algorithm is used to calculate the globally optimal flight path. Specifically, it receives the delivery task sequence from the intelligent decision-making module, an ordered list of target points, and considers the UAV's current state (e.g., battery power, position) and environmental constraints (e.g., no-fly zones, obstacles) using the A* algorithm. 、DOr a similar algorithm. Calculate a global flight path that connects all drop points with the shortest total distance or the least time. This ensures the basic efficiency of the entire operation.

[0092] The trajectory smoothing optimization employs cubic B-spline curves for path fitting, generating smooth transition curves at path inflection points by inserting auxiliary control points. This ensures the continuity of the trajectory in terms of position, velocity, and acceleration, eliminating sharp corners in the path. The original path search algorithm typically generates a broken line composed of line segments, requiring the UAV to instantly change direction at inflection points, which is impractical in real-world flight. By inserting transition curves, such as B-spline curves or polynomial curves, the function is to ensure that the generated trajectory changes continuously in terms of position, zero-order velocity, first-order acceleration, and second-order acceleration. This avoids sudden stops, sharp turns, or shaking of the UAV at inflection points, resulting in smoother flight, reduced airframe wear and energy waste, and a stable platform for precise delivery.

[0093] Based on the dynamic characteristics of the UAV, a speed adaptive control strategy is designed. In the path inflection point region, segmented speed planning is adopted, including three stages: deceleration upon entry, constant speed passage, and acceleration upon exit. Before entering the inflection point, uniform deceleration is performed to gradually reduce the speed to a safe level. Maintain a constant safe speed during the transition phase at the inflection point. ,in The maximum centripetal acceleration is r, and the turning radius is r. After leaving the inflection point, the vehicle undergoes uniform acceleration and gradually returns to cruising speed.

[0094] The flight speed is dynamically adjusted based on the curvature of the trajectory, i.e., the sharpness of the turn, and the dynamic limits of the UAV. Through segmented speed planning, refined speed management is implemented at the inflection points of the smoothed trajectory: Deceleration upon entry: Decelerating near the inflection point ensures a smooth entry into the curve, preventing deviation from the trajectory due to excessive centrifugal force. Constant speed passage: Maintaining stable low-speed flight in the core area of ​​the inflection point provides conditions for possible drop operations, if the drop point is located at the inflection point or for precise attitude control. Acceleration upon exit: Accelerating after passing the inflection point to quickly return to cruising speed and ensure overall operational efficiency. This strategy directly guarantees the UAV's tracking accuracy and flight stability on complex paths.

[0095] By monitoring environmental wind disturbances and sudden obstacles in real time, and predicting the drone's motion state based on extended Kalman filtering, the system dynamically adjusts flight trajectory parameters and delivery timing, forming a closed-loop control system with feedforward compensation. It senses external disturbances, such as gusts and airflow, and internal state changes, such as thrust changes caused by battery voltage drops, predicting the impact of these disturbances on the drone's future state, position, and attitude. Through dynamic adjustments, feedforward compensation is performed. This includes adjusting the flight trajectory; if wind is predicted to deflect the drone, it preemptively controls the drone to fly a slight offset in the opposite direction to counteract the wind's effect. It also adjusts the delivery timing; during high-speed flight, based on real-time position and speed, it accurately calculates and issues delivery commands in advance, ensuring the agent is released at the precise time and lands on the target point due to inertia. This forms a closed-loop feedback control, constructing a continuous cycle of perception-prediction-compensation, enabling the system to proactively adapt to environmental changes rather than merely responding passively. Therefore, even under various uncertain conditions, it can reliably complete the ultimate task of precise delivery.

[0096] These five steps are interconnected and have clearly defined functions. Together, they form a complete technical system that integrates enhanced perception, identification and matching, fusion positioning, intelligent decision-making, and precise execution. The system aims to systematically address the shortcomings of existing drone-based plant protection technologies in terms of accuracy, intelligence, and environmental adaptability.

[0097] Example 2 like Figure 2 As shown, this application provides an architecture diagram of a drone-based biological control system based on real-time visual perception and precise delivery decision-making, which is applied to the drone-based biological control system based on real-time visual perception and precise delivery decision-making as described in Embodiment 1. It includes a data acquisition and enhancement module 200, a target detection and matching module 210, a three-dimensional perception and fusion module 220, an intelligent decision-making and planning module 230, and a precise execution and control module 240.

[0098] The data acquisition and enhancement module 200 is used to synchronously acquire farmland environmental data through a multi-sensor system carried by the UAV, and to perform low-light enhancement processing on visible light images.

[0099] The target detection and matching module 210 is used to input the enhanced image into the target detection network, simultaneously identify pest targets and detect the placement location, and establish the correspondence between the target and the placement point.

[0100] The 3D perception and fusion module 220 is used to spatially align the 2D detection results with the point cloud data to achieve precise 3D positioning, and to fuse image features, thermal infrared features and 3D spatial features.

[0101] The intelligent decision-making and planning module 230 is used to introduce a visual language model to perform cross-modal reasoning on multimodal perception data, combine task instructions to perform target disambiguation and delivery priority analysis, and generate a collision-free delivery task sequence.

[0102] The precision execution and control module 240 is used to plan the optimal flight path according to the delivery task sequence, achieve precise drug delivery through trajectory optimization and speed control, and form closed-loop feedback control based on real-time environmental data.

[0103] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.

[0104] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.

[0105] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0106] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) to achieve the functions of the above-described embodiments of the present invention, enabling the computer to function. Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (ROM) and flash ROM, or hard disk drives (HDDs). It is then read, corrected, and written by the CPU as needed.

[0107] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0108] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0109] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0110] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly of multiple systems, a system group. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0111] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A method for biological control by a UAV based on real-time visual perception and accurate delivery decision, characterized in that, The method comprises the following steps: Synchronously collecting farmland environment data by a multi-sensor system carried by a UAV, and performing low-light enhancement processing on a visible light image; Inputting the enhanced image into a target detection network to synchronously identify pest targets and detect a release posture, and establishing a correspondence between the targets and release points; Aligning the two-dimensional detection results with point cloud data in space to achieve three-dimensional accurate positioning, and fusing image features, thermal infrared features and three-dimensional space features; Introducing a visual language model to perform cross-modal reasoning on multi-modal perception data, combining task instructions to perform target disambiguation and release priority analysis, and generating a collision-free release task sequence; Planning an optimal flight path according to the release task sequence, achieving accurate pesticide release through trajectory optimization and speed control, and forming a closed-loop feedback control based on real-time environment data. 2.The unmanned aerial vehicle biological control method based on real-time visual perception and accurate delivery decision of claim 1, wherein, The synchronous collection of farmland environment data and the low-light enhancement processing on the visible light image comprise: Synchronously collecting RGB images, thermal infrared images and six-degree-of-freedom pose data of the farmland environment by a visible light camera, an infrared thermal imager and a GPS / IMU combined navigation system carried by the UAV; Inputting the collected visible light image into a multi-scale Transformer enhancement network based on physical guidance for processing, wherein the enhancement network comprises: A decomposition network for decoupling the input image into illumination components and reflection components; A multi-scale Transformer encoder for performing illumination distribution estimation and brightness enhancement on the illumination components; A detail restoration module for performing texture enhancement and noise suppression on the reflection components; A reconstruction network for fusing the processed illumination components and reflection components to output the enhanced visible light image; Wherein, the enhancement network is optimized in the training stage by using an adaptive illumination perception loss function, which combines illumination smoothness constraint, color fidelity and structural similarity measure. 3.The unmanned aerial vehicle biological control method based on real-time visual perception and accurate delivery decision of claim 2, wherein, The adaptive illumination-aware loss function defined by the following equation: ， wherein, is a lighting smoothness constraint term defined as an enhanced gradient difference between the front and back lighting components; ， For the color fidelity term, the color difference is calculated in the Lab color space: ， For the structural similarity measure term, the structural information based on local image patches is computed: ， is the weight coefficient, is the enhanced image, is the original image, SSIM(·) is the structural similarity computation function, denotes the gradient operator of the image, denotes the square of the L2 norm over all pixels, : the three channels of the CIELAB color space, : the image obtained after enhancement of the corresponding channel in Lab space, : the original image of the corresponding channel in Lab space. 4.The unmanned aerial vehicle biological control method based on real-time visual perception and accurate delivery decision of claim 1, wherein, The inputting of the enhanced image into the target detection network, the synchronous identification of pest targets and the detection of release positions, and the establishment of the correspondence between the targets and release points comprise: Inputting the enhanced image into an improved small target detection network, which uses EfficientNet-B0 as a shared backbone network, realizes multi-scale feature fusion through a bidirectional feature pyramid network, and embeds a coordinate attention mechanism to enhance the feature expression ability of small pest targets; The coordinate attention mechanism realizes space information modeling in the following way: First, global pooling is performed along the height and width directions respectively to obtain direction-aware feature descriptors: ， ， wherein, , are the first and second global feature descriptors of the channel in the height and width directions, respectively, , are the first and second global feature descriptors of the channel in the height and width directions, respectively, , are the input feature maps of the channel at positions h, w, respectively, h, w are the height H and width W indices, respectively.​​ The pooled features are then concatenated and convoluted to generate attention weights through a sigmoid activation : ， wherein, is a Sigmoid activation function, is a convolution operation, is a concatenation of the feature descriptors along the height and width directions; Finally, the attention weights are multiplied with the original feature map to enhance the feature response of the target region; In establishing the correspondence between the target and the delivery point, an optimized matching method is adopted, and a weighted score function is defined as : ， wherein represents a target bounding box and a delivery position box , and the intersection over union of is a delivery quality confidence, and α is an adjustable balance parameter; Through a local optimal selection strategy, the best release position box is matched for each target bounding box in order from high to low weighted score, ensuring one-to-one accurate correspondence.

5. The method of claim 1, wherein the method further comprises: determining a target area for the UAV to fly to based on the real-time visual perception and the precision delivery decision. 5 The spatial alignment of the two-dimensional detection results with the point cloud data to achieve three-dimensional accurate positioning and the fusion of image features, thermal infrared features and three-dimensional space features comprise: Based on the real-time pose of the unmanned aerial vehicle and the sensor parameters, the two-dimensional recognition result obtained by target detection is spatially registered with the three-dimensional point cloud data to establish the coordinate correspondence between the image space and the three-dimensional space; A multi-scale voxelization processing strategy is adopted to construct a hierarchical spatial feature representation, including a coarse-grained voxel layer for extracting global spatial context information, a medium-grained voxel layer for retaining main geometric structure features, and a fine-grained voxel layer for capturing local detail features; Through a cross-modal attention fusion mechanism, visible light image features, thermal infrared image features, and three-dimensional space features are adaptively weighted and fused, wherein a multi-head attention architecture is used to process different modal feature information in parallel; Finally, a comprehensive perception feature map that integrates multi-source information is generated, which contains the visual appearance characteristics, thermal radiation characteristics, and three-dimensional geometric characteristics of the target, providing spatial context-aware perception data for subsequent intelligent decision-making.

6. The method of claim 1, wherein the method further comprises: The introduced visual language model performs cross-modal reasoning on multi-modal perception data, combines task instructions for target disambiguation and drop priority analysis, and generates a collision-free drop task sequence, including: The visual language model is introduced as the core reasoning engine to align multi-modal perception features with natural language task instructions, enabling semantic-level understanding of pest targets; A visual marking mechanism is used to assign unique identifiers to detected targets, and similarity calculations are performed in a unified embedding space based on visual features, spatial positions, and semantic descriptions, solving the recognition ambiguity problem of similar targets; Based on spatial constraints and task requirements, a drop priority evaluation model is established by considering the semantic importance, spatial accessibility, and operation safety of targets; A structured prompt template is used to integrate environmental information, task instructions, and operation constraints to generate a collision-free drop task sequence that complies with the rules of the physical world.

7. The method of claim 6, wherein the method further comprises: It also includes: A multi-modal task understanding and planning framework based on the visual language model is constructed, which includes the following processing mechanisms: Task semantic analysis and embodiment mapping mechanism for decomposing abstract natural language instructions into specific task targets related to the current perception scene; Cross-modal instance disambiguation mechanism based on visual marking, which assigns unique identifiers to detected targets to accurately distinguish appearance-similar targets in a unified semantic space; Spatial-semantic collaborative priority reasoning mechanism that generates a work sequence that complies with physical constraints by considering the semantic importance, spatial accessibility, and operation safety of targets; Structured prompt engineering-driven planning generation mechanism that outputs a standardized executable task sequence by integrating environmental context, task specifications, and skill knowledge through hierarchical prompt templates. 8.The unmanned aerial vehicle biological control method based on real-time visual perception and accurate delivery decision of claim 7, wherein, The mechanisms in the multi-modal task understanding and planning framework specifically include: The task semantic analysis and embodiment mapping mechanism analyzes key semantic elements in the instructions through a pre-trained language understanding module and maps abstract task descriptions to specific perception features and operation requirements; The cross-modal instance disambiguation mechanism based on visual marking calculates the similarity of visual features, text descriptions, and spatial positions in a unified embedding space to achieve accurate target recognition in complex scenes; The space-semantic collaborative priority reasoning mechanism comprehensively considers semantic importance, spatial accessibility, time sequence dependency and safety constraint factors to generate an optimal operation sequence. The structured prompt engineering driven planning generation mechanism ensures the generation of a task sequence that can be directly parsed and executed by downstream systems by designing a hierarchical template structure including an environmental context layer, a task specification layer, a skill knowledge layer and an output specification layer. 9.The unmanned aerial vehicle biological control method based on real-time visual perception and accurate delivery decision of claim 1, wherein, According to the delivery task sequence, the optimal flight path is planned, the precise pesticide delivery is realized through trajectory optimization and speed control, and the closed-loop feedback control is formed based on real-time environmental data, including: According to the delivery task sequence, the global optimal flight path is calculated by combining the current state parameters of the unmanned aerial vehicle and the operation environment constraints using a path search algorithm; The flight path is smoothed and optimized through a curve fitting method, and a transition curve is inserted at the path inflection point to ensure the continuity of the trajectory in the position, speed and acceleration layers; Based on the dynamic characteristics of the unmanned aerial vehicle, a speed adaptive control strategy is designed, and a segmented speed planning is used in the path inflection point area, including three stages of deceleration entry, constant speed through and acceleration exit; Through real-time monitoring of environmental disturbances and system states, the flight trajectory and delivery timing are dynamically adjusted based on a motion state prediction algorithm to form a closed-loop control system with feedforward compensation.

10. An unmanned aerial vehicle biological control system based on real-time visual perception and precise release decision, applied to the unmanned aerial vehicle biological control method based on real-time visual perception and precise release decision according to any one of claims 1 to 9, characterized in that, The system comprises: A data acquisition and enhancement module for synchronously acquiring farmland environment data through a multi-sensor system carried by the unmanned aerial vehicle and performing low-light enhancement processing on visible light images; A target detection and matching module for inputting the enhanced images into a target detection network to synchronously identify pest targets and detect delivery positions, and establishing a correspondence between the targets and the delivery points; A three-dimensional perception and fusion module for spatially aligning two-dimensional detection results with point cloud data to achieve three-dimensional accurate positioning, and fusing image features, thermal infrared features and three-dimensional space features; An intelligent decision and planning module for introducing a visual language model to perform cross-modal reasoning on multi-modal perception data, combining task instructions to perform target disambiguation and delivery priority analysis, and generating a collision-free delivery task sequence; A precise execution and control module for planning an optimal flight path according to the delivery task sequence, realizing precise pesticide delivery through trajectory optimization and speed control, and forming a closed-loop feedback control based on real-time environmental data.

Images