Unmanned aerial vehicle cluster point generation method and device based on SVG formation pattern

Abstract

The application discloses a kind of unmanned aerial vehicle cluster point generation method and device based on SVG formation figure, and the method comprises the following steps: analyzing scalable vector graphics format formation figure path, extracting geometric primitives, geometric feature points and topological relations, generating structured topological skeleton, reference adjacency matrix and normalized rendering graph;Feature coding is carried out on the normalized rendering graph, and the initial Gaussian noise point set is combined with each time step denoising, and the clean point position estimation set is obtained by Tweedie estimation;According to the signal-to-noise ratio of constraint signal, selectively introduce safety distance loss and connectivity maintenance loss to correct the denoising direction, iteratively generate point set, and then detect distance, repulsive displacement and coordinate mapping to obtain unmanned aerial vehicle cluster point coordinates.The technical scheme of the application can generate results considering both graphic features and topological integrity, reduce point conflicts through distance correction, and improve formation generation efficiency, safety and deployment reliability.

Description

Technical Field

[0001] This invention relates to the technical field of unmanned aerial vehicle (UAV) swarm control, and in particular to a method and apparatus for generating UAV swarm locations based on SVG formation graphics. Background Technology

[0002] Drone swarm formation requires generating point coordinates that match the number of drones according to a predetermined graphic, while considering graphic recognizability, safe spacing between drones, and topological connectivity between points. Scalable vector graphics describe two-dimensional graphics using mathematical parameters and path instructions, featuring distortion-free scaling, clear path semantics, and ease of parsing, making them suitable for representing target formation graphics of drone swarms. Existing point generation methods typically employ uniform arc length sampling or geometric discretization, dividing the graphic path into several discrete coordinate points and directly outputting them. While these methods are simple to calculate and generate quickly, they struggle to adjust sampling density based on local geometric features such as curvature changes, endpoints, inflection points, and intersections. When the number of drones is limited, this can easily lead to insufficient points in complex curve areas and redundant points in simple path areas, making it difficult to balance point utilization efficiency with the geometric fidelity of the formation graphics.

[0003] Diffusion probability models generate target samples through progressive noise addition and denoising, which can be used for generative modeling of UAV formation points and improve the expressive power of complex point distributions. However, the optimization objectives of existing diffusion probability models and point cloud generation methods usually focus on the statistical distribution fitting between generated data and training samples, lacking a direct expression of the actual constraints of UAV formations. If safety distances and topological connectivity constraints are not embedded in the training or inference process, generated points may exhibit problems such as excessively small local distances, broken connections, and overall structural degradation, still requiring manual inspection or post-processing adjustments, making it difficult to simultaneously guarantee the recognizability, topological integrity, and deployment security of the formation graphics. How to properly solve these problems has become an urgent issue for the industry. Summary of the Invention

[0004] This invention provides a method and apparatus for generating drone cluster locations based on SVG formation graphics, which improves the automation level of formation location generation while taking into account graphic fidelity, topological connectivity, and drone deployment security.

[0005] According to a first aspect of the present invention, a method for generating UAV swarm locations based on SVG formation graphics is provided, the method comprising: S1. Obtain the SVG formation graphic, parse the path instructions in the SVG formation graphic to obtain the geometric primitives that make up the SVG formation graphic, extract geometric feature points based on the geometric primitives and parse the topological relationship between the geometric primitives to obtain the structured topological skeleton and the reference adjacency matrix corresponding to the preset number of drones, and perform rasterization rendering and pixel normalization processing on the SVG formation graphic to obtain the normalized SVG rendering image. S2. Input the normalized SVG rendering image into the conditional encoder for multi-layer feature encoding to obtain a conditional feature sequence for characterizing the local spatial features of the SVG formation graphic, and generate an initial Gaussian noise point set according to the preset number of drones. S3. For each denoising time step, input the current noise point set, the denoising time step, and the conditional feature sequence into the point cloud denoiser to obtain the noise prediction result corresponding to the current noise point set. Then, perform Tweedie estimation based on the current noise point set, the noise prediction result, and the noise scheduling parameters corresponding to the denoising time step to obtain the current clean point estimation set. S4. Obtain the constraint signal signal-to-noise ratio (SNR) of the current denoising time step. When the constraint signal SNR reaches a preset activation threshold, calculate the safety distance loss based on the distance between each pair of points in the current clean point estimation set, calculate the connectivity preservation loss based on the nearest neighbor relationship of the current clean point estimation set and the reference adjacency matrix, perform weighted fusion of the safety distance loss and the connectivity preservation loss to obtain the topology constraint loss, and correct the denoising update direction of the current noisy point set based on the topology constraint loss. When the constraint signal SNR does not reach the preset activation threshold, perform basic denoising update on the current noisy point set based on the noise prediction result, and use the corrected denoising update result or the basic denoising update result as the current noise point set for the next denoising time step for iterative processing to obtain the generated point set. S5. Detect the distance between each pair of points in the generated point set, apply displacement to the point pairs whose distance is less than a preset safety threshold to make the point pairs move away from each other, and repeat the distance detection and displacement processing on the point set after displacement to obtain a safe point set that meets the preset safety threshold. Map the safe point set from normalized canvas coordinates to UAV deployment coordinates to obtain the UAV cluster point coordinates.

[0006] According to a second aspect of the present invention, a device for generating drone swarm locations based on SVG formation graphics is provided, comprising: The acquisition module is used to acquire SVG formation graphics, parse the path instructions in the SVG formation graphics to obtain the geometric primitives that make up the SVG formation graphics, extract geometric feature points based on the geometric primitives and parse the topological relationships between the geometric primitives to obtain a structured topological skeleton and a reference adjacency matrix corresponding to a preset number of drones, and perform rasterization rendering and pixel normalization processing on the SVG formation graphics to obtain a normalized SVG rendering image. The generation module is used to input the normalized SVG rendering image into the conditional encoder for multi-layer feature encoding to obtain a conditional feature sequence for characterizing the local spatial features of the SVG formation graphic, and to generate an initial Gaussian noise point set based on the preset number of drones. The estimation module is used to input the current set of noise points, the denoising time step and the conditional feature sequence into the point cloud denoiser for each denoising time step to obtain the noise prediction result corresponding to the current set of noise points, and to perform Tweedie estimation based on the current set of noise points, the noise prediction result and the noise scheduling parameters corresponding to the denoising time step to obtain the current clean point estimation set. The update module is used to obtain the constraint signal signal-to-noise ratio (SNR) of the current denoising time step. When the constraint signal SNR reaches a preset activation threshold, the module calculates the safety distance loss based on the distance between each pair of points in the current clean point estimation set, calculates the connectivity preservation loss based on the nearest neighbor relationship of the current clean point estimation set and the reference adjacency matrix, performs weighted fusion of the safety distance loss and the connectivity preservation loss to obtain the topological constraint loss, and corrects the denoising update direction of the current noisy point set based on the topological constraint loss. When the constraint signal SNR does not reach the preset activation threshold, the module performs basic denoising update on the current noisy point set based on the noise prediction result, and uses the corrected denoising update result or the basic denoising update result as the current noisy point set for the next denoising time step for iterative processing to obtain the generated point set. The mapping module is used to detect the distance between each pair of points in the generated point set, apply displacement to the point pairs whose distance is less than a preset safety threshold to make the point pairs move away from each other, and repeat the distance detection and displacement processing on the point set after displacement to obtain a safe point set that meets the preset safety threshold. The safe point set is then mapped from normalized canvas coordinates to UAV deployment coordinates to obtain the UAV cluster point coordinates.

[0007] According to a third aspect of the present invention, an electronic device is provided, comprising: a communication interface, a processor, and a memory; The memory is used to store program instructions, which, when executed by the processor connected to the memory via the communication interface, implement any of the above-described methods for generating UAV cluster locations based on SVG formation graphics.

[0008] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, which, when executed by a computer (e.g., a processor in the computer), implement any of the above-described methods for generating UAV cluster locations based on SVG formation graphics.

[0009] In summary, this invention provides a method and apparatus for generating UAV swarm location data based on SVG formation graphics. The method includes: parsing the formation graphic path in a scalable vector graphics format, extracting geometric primitives, geometric feature points, and topological relationships, and generating a structured topological skeleton, a reference adjacency matrix, and a normalized rendering graph; performing feature encoding on the normalized rendering graph, denoising it step-by-step using an initial Gaussian noise location set, and obtaining a clean location estimation set through Tweedie estimation; selectively introducing safety distance loss and connectivity preservation loss based on the constraint signal-to-noise ratio to correct the denoising direction, iteratively generating the location set, and then obtaining the UAV swarm location coordinates through distance detection, rejection displacement, and coordinate mapping. The technical solution of this invention can automatically convert SVG formation graphics into UAV swarm location coordinates, and introduce safety distance and connectivity preservation constraints during the denoising process, ensuring that the generated result balances graphic features and topological integrity; reducing location conflicts through distance correction improves formation generation efficiency, security, and deployment reliability.

[0010] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and drawings.

[0011] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0013] Figure 1 A flowchart of a method for generating UAV cluster locations based on SVG formation graphics, provided as an embodiment of the present invention; Figure 2 A flowchart of step S1 of a method for generating UAV cluster locations based on SVG formation graphics, provided for an embodiment of the present invention; Figure 3 A flowchart of step S2 of a method for generating UAV cluster locations based on SVG formation graphics, provided as an embodiment of the present invention; Figure 4 A flowchart of step S3 of a method for generating UAV cluster locations based on SVG formation graphics, provided as an embodiment of the present invention; Figure 5 A flowchart of step S4 of a method for generating UAV cluster locations based on SVG formation graphics, provided for an embodiment of the present invention; Figure 6 A flowchart of step S5 of a method for generating UAV cluster locations based on SVG formation graphics, provided as an embodiment of the present invention; Figure 7 A structural diagram of a drone cluster location generation device based on SVG formation graphics is provided for an embodiment of the present invention; Figure 8 A structural diagram of an electronic device provided as an embodiment of the present invention; Figure 9 A schematic diagram of multiple types of geometric feature points provided for embodiments of the present invention; Figure 10 A flowchart illustrating the adaptive curvature geometry sampling method provided for embodiments of the present invention; Figure 11 A flowchart illustrating the topology-constrained guided diffusion model method provided in an embodiment of the present invention. Detailed Implementation

[0014] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0015] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0016] Definitions: SVG, short for Scalable Vector Graphics, uses an extensible markup language to describe two-dimensional graphics. It saves graphics using mathematical parameters such as coordinates, paths, line segments, curves, and arcs, rather than recording them pixel-by-pixel. It maintains quality after scaling, has a clear path structure, and is easy to edit and parse. In this invention, SVG is used to describe UAV formation graphics. Path instructions can be further parsed into geometric primitives such as straight line segments, Bézier curve segments, and arc segments, and the coordinates of the points required for formation can be extracted accordingly.

[0017] Diffusion probability models are a type of generative model whose core idea is to learn the generation patterns of data through "progressive noise addition" and "progressive denoising." During training, the model continuously adds Gaussian noise to real samples, gradually transforming them into data that approximates random noise; simultaneously, a denoising network is trained to predict the noise added at each time step. During generation, the model starts with Gaussian noise and progressively removes noise over multiple denoising time steps, ultimately obtaining new samples that conform to the distribution of the training data.

[0018] Denoising diffusion implicit model: This is an accelerated generation method for diffusion probability models. It utilizes a pre-trained noise prediction network, starting from initial Gaussian noise, and progressively removes noise according to preset inference time steps to finally generate the target sample. Compared to conventional denoising diffusion probability models, the denoising diffusion implicit model does not need to infer step-by-step from all time steps of the training phase; it can select a portion of the training time steps for denoising, thus increasing generation speed while reducing the number of inferences. When using deterministic denoising, the same generation results can be obtained with the same model parameters, initial Gaussian noise, and conditional inputs, demonstrating good reproducibility.

[0019] Tweedie estimation is a method for estimating the conditional mean of original noise-free data based on noisy observations, noise statistics, and probability distribution information. Its core function is to establish a connection between "noisy data" and "estimated original data," enabling the model to infer relatively clean data from the noise-infused data. In diffusion probability models, the set of points at the current denoising time step contains a certain proportion of Gaussian noise. Tweedie estimation uses the current noisy point set, the noise prediction results output by the point cloud denoiser, and the noise scheduling parameters corresponding to this denoising time step to calculate the estimated coordinates of each point in a noise-free state, forming the current estimated set of clean points. This estimate is typically used to evaluate the shape, spacing, or topological relationships of the current generated results.

[0020] Feature-wise Linear Modulation (FiLM): FiLM is a method that dynamically adjusts intermediate features using conditional information. It generates scaling and offset parameters based on time steps, image features, and other conditions, scaling and offsetting the input features channel-by-channel, allowing the same network to change its feature response for different conditions. In this invention, FiLM adjusts the point features according to the denoising time step, enabling the point cloud denoiser to adapt to different noise stages and improve noise prediction accuracy.

[0021] Cross-attention: A correspondence is established between two sets of different features, where one set serves as query information and the other as key and value information. The value information is weighted and aggregated by calculating the correlation between the query information and the key information. In this invention, point-related features serve as query information, and the conditional feature sequence of the SVG formation graphic serves as key and value information. This allows each noisy point to acquire the local spatial features of the corresponding graphic region, thereby guiding the points to gradually complete denoising and position adjustment along the target formation graphic.

[0022] Inverse transform sampling: Samples are generated according to the target probability distribution. First, the weights of each candidate point are normalized and a cumulative distribution function is constructed. Then, random values ​​are drawn from a uniform distribution between zero and one, and the corresponding candidate point is searched within the cumulative distribution function. Candidate points with larger weights occupy a wider cumulative probability interval and have a higher probability of being selected. In this invention, this method can give higher sampling probabilities to candidate points with high curvature, near inflection points, or located at endpoints, thereby improving the preservation of key geometric features.

[0023] Force-oriented algorithm: This method simulates the constraint relationships between points as virtual forces and iteratively updates the point coordinates. For points that are too close together, the algorithm applies mutually repulsive forces along the line connecting the point pairs and combines this with a damping coefficient to limit the amplitude of single displacements, gradually achieving a stable distribution of points. In this invention, this algorithm is used to detect and eliminate spacing conflicts between UAV points, ensuring that the distance between each pair of points meets a preset safety threshold, serving as a safety correction measure for the generated points.

[0024] like Figure 1 As shown, this invention provides a method for generating drone swarm locations based on SVG formation graphics. This method includes steps S1-S5: In step S1, an SVG formation graphic is obtained, the path instructions in the SVG formation graphic are parsed to obtain the geometric primitives that make up the SVG formation graphic, geometric feature points are extracted based on the geometric primitives and the topological relationship between the geometric primitives is parsed to obtain a structured topological skeleton and a reference adjacency matrix corresponding to a preset number of UAVs, and the SVG formation graphic is rasterized and pixel normalized to obtain a normalized SVG rendering image. In step S2, the normalized SVG rendering image is input into the conditional encoder for multi-layer feature encoding to obtain a conditional feature sequence for characterizing the local spatial features of the SVG formation graphic, and an initial Gaussian noise point set is generated according to the preset number of drones. In step S3, for each denoising time step, the current noise point set, the denoising time step, and the conditional feature sequence are input into the point cloud denoiser to obtain the noise prediction result corresponding to the current noise point set. Tweedie estimation is performed based on the current noise point set, the noise prediction result, and the noise scheduling parameters corresponding to the denoising time step to obtain the current clean point estimation set. In step S4, the constraint signal signal-to-noise ratio (SNR) of the current denoising time step is obtained. When the constraint signal SNR reaches a preset activation threshold, the safety distance loss is calculated based on the distance between each pair of points in the current clean point estimation set. The connectivity preservation loss is calculated based on the nearest neighbor relationship of the current clean point estimation set and the reference adjacency matrix. The safety distance loss and the connectivity preservation loss are weighted and fused to obtain the topological constraint loss. The denoising update direction of the current noisy point set is corrected based on the topological constraint loss. When the constraint signal SNR does not reach the preset activation threshold, the current noisy point set is updated based on the noise prediction result. The corrected denoising update result or the basic denoising update result is used as the current noisy point set for the next denoising time step for iterative processing to obtain the generated point set. In step S5, the distance between each pair of points in the generated point set is detected. For point pairs whose distance is less than a preset safety threshold, a displacement is applied to make the point pairs move away from each other. The distance detection and displacement processing are repeated on the point set after the displacement to obtain a safe point set that meets the preset safety threshold. The safe point set is mapped from normalized canvas coordinates to UAV deployment coordinates to obtain the UAV cluster point coordinates.

[0025] In one embodiment, a drone swarm formation in a two-dimensional plane is taken as the object, with a preset number of drones of N. The input is an SVG formation graphic describing the shape of the target formation. Each point is calculated within a normalized canvas. In this embodiment, the coordinates of the drone swarm points are generated based on the SVG formation graphic, the preset number of drones, the preset safety threshold, the preset activation threshold, and noise scheduling parameters.

[0026] The system reads movement commands, line commands, quadratic Bézier curve commands, cubic Bézier curve commands, arc commands, and closed path commands from the SVG formation graphic, converting the corresponding paths into line segments, quadratic Bézier curve segments, cubic Bézier curve segments, and arc segments. Geometric feature points are extracted based on the endpoint positions, curvature changes, angles, intersection relationships, and arc length positions of each geometric primitive. A structured topological skeleton is obtained by combining the connection relationships between geometric primitives. A reference adjacency matrix corresponding to the preset number of UAVs is generated based on the structured topological skeleton. The SVG formation graphic is then rasterized and pixel normalized to obtain a normalized SVG rendering image.

[0027] The normalized SVG rendering image is input into the conditional encoder, and multi-layer downsampling is performed through a convolutional coding network to obtain a spatial feature map. The spatial feature map is then converted into a spatial token sequence, forming a conditional feature sequence used to characterize the local spatial features of the SVG formation graphic. Based on the preset number of drones, a corresponding number of two-dimensional random coordinates are extracted from a Gaussian distribution to obtain an initial set of Gaussian noise points, which is then used as the current noise point set for the first denoising time step.

[0028] The current set of noisy points is embedded with coordinates, and the denoising time step is sinusoidally time-coded. Feature linear modulation parameters are generated based on the time coding results. The point cloud denoiser extracts inter-point correlation features through self-attention operations, fuses conditional feature sequences into the inter-point correlation features through cross-attention operations, and obtains the noise prediction result through multi-layer feature linear modulation and feedforward operations. Based on the current set of noisy points, the noise prediction result, and the noise scheduling parameters, Tweedie estimation is performed to determine the estimated coordinates of each UAV point in a noise-free state, resulting in the current clean point estimation set.

[0029] The constraint signal-to-noise ratio (SNR) is determined based on the signal and noise components at the current denoising time step. When the constraint signal SNR reaches a preset activation threshold, the Euclidean distance between each pair of points in the current clean point estimation set is calculated. For point pairs with distances less than a preset safety threshold, a distance penalty is determined to obtain the safety distance loss. Continuous adjacency weights are determined based on the distances between each point and its neighbors, and a current nearest neighbor weight matrix is ​​constructed. The connectivity preservation loss is obtained based on the structural differences between the current nearest neighbor weight matrix and the reference adjacency matrix. The safety distance loss and connectivity preservation loss are weighted to form the topological constraint loss, and the corresponding constraint gradient is used to correct the basic denoising update direction. When the constraint signal SNR does not reach the preset activation threshold, basic denoising updates are performed based on the noise prediction results. The updated point set is used as the current noise point set for the next denoising time step, and the generated point set is obtained iteratively.

[0030] The system detects the distance between each pair of points in the generated point set. For point pairs with a distance less than a preset safety threshold, it calculates the repulsion displacement and updates the corresponding point coordinates along the opposite direction of the line connecting the point pairs. This distance detection and displacement processing is repeated until a preset maximum number of iterations is reached. If spacing conflicts still exist, a compensation displacement is applied to the corresponding points to obtain a safe point set. Based on the scale ratio between the normalized canvas area and the UAV deployment area, the origin offset, and the coordinate axis direction, the safe point set is scaled, translated, and transformed to obtain the coordinates of the UAV cluster points.

[0031] Multiple types of geometric feature points were detected from the SVG path geometry primitive list, including but not limited to: Type 1 - Curve anchor points (the beginning and end points of Bézier curves and the start and end points of arc segments), Type 2 - Curve peaks (points with maximum curvature within curve segments), Type 3 - Endpoints of non-closed paths (the start and end points of non-closed path strokes), Type 4 - Acute angle vertices (vertices of straight line segments with an included angle less than a preset acute angle threshold), Type 5 - Obtuse angle vertices (vertices with an included angle greater than a preset obtuse angle threshold), Type 6 - High curvature points (this tag is added when the curvature value corresponding to a vertex exceeds a preset curvature threshold), Type 7 - Intersection points (the intersection point of two primitives that do not belong to the same stroke and do not share endpoints), Type 8 - Midpoints (the geometric midpoint between the ends of a line segment), and Type 9 - Trisection points (points at the 1 / 3 and 2 / 3 positions of a line segment), as shown in the appendix. Figure 9 As shown.

[0032] The specific implementation steps of the two embodiments are described in detail below.

[0033] Example 1 The UAV swarm location is generated using an adaptive curvature geometry sampling method, as shown in the attached figure. Figure 10As shown, candidate points are arranged along the arc length direction on each geometric primitive corresponding to the SVG formation graphic according to a preset oversampling factor, forming a candidate point set. This ensures that the geometric primitive segments with large curvature changes have a sufficient number of candidate points. For each candidate point in the candidate point set, adjacent candidate points located on both sides of the candidate point along the arc length direction of the corresponding geometric primitive are obtained. The discrete curvature of the candidate point is determined based on the ratio of the change in the direction angle between the two adjacent direction vectors to the change in the corresponding arc length. The discrete curvature corresponding to candidate points on straight line segments approaches zero, while the discrete curvature corresponding to geometric primitive segments with larger curvature increases accordingly.

[0034] The importance weight of each candidate point is determined based on its discrete curvature, distance to the nearest curve inflection point, and endpoint attributes. The preferred weighting coefficients are 55% for discrete curvature, 30% for distance to curve inflection points, and 15% for endpoint attributes. These coefficients can be adjusted based on training data or formation graphic features. The importance weights of each candidate point are normalized to sampling probabilities, and a cumulative distribution function is constructed according to the order of candidate points along the geometric primitives. Sampling parameters are extracted from a uniform distribution of zero to one. Candidate points whose cumulative probability is not less than the sampling parameter for the first time are selected as sampling points until the number of sampling points reaches the preset number of drones, resulting in an initial point set. Candidate points with higher importance weights correspond to wider cumulative probability intervals, thus having a higher selection probability. This allows candidate points with larger curvatures, closer to curve inflection points, or located at path endpoints to be preferentially included in the initial point set.

[0035] The algorithm iterates through various geometric feature points, calculating the nearest distance between each geometric feature point and the initial point set. When none of the geometric feature points in a certain category are covered by the initial point set, a target feature point is selected from that category. The coordinates of the sampling point in the initial point set closest to the target feature point are replaced with the coordinates of the target feature point. Simultaneously, the sampling point whose coordinates have been replaced is marked as a protected point, resulting in a point set to be corrected. The distance between each pair of points in the point set to be corrected is detected. For point pairs with a distance less than a preset safety threshold and containing unprotected points, one of the unprotected points is deleted based on the protection attribute and the preset geometric feature category priority. Points are then added from the unselected candidate points according to their importance weight. The distance detection, point deletion, and point addition are repeated until there are no point pairs containing unprotected points and with a distance less than the preset safety threshold. For protected points that still have spacing conflicts, a displacement is applied to move the corresponding point pairs away from each other. The corrected points are then mapped from normalized canvas coordinates to UAV deployment coordinates to obtain the UAV cluster point coordinates.

[0036] Example 2 The topology-constrained guided diffusion model method is used to generate UAV swarm locations, as shown in the attached figure. Figure 11 As shown, a training dataset containing multiple shape categories is constructed, and the control parameters corresponding to each shape category are varied within a preset range to improve the coverage of different SVG formation graphics by the training data. The standard point set corresponding to each training sample can be generated by an adaptive curvature geometry sampling method, ensuring that the standard point set meets a preset safety threshold and a preset topological relationship. The same network parameters are used in the training and inference phases; when there is a difference between the distribution of the SVG formation graphics to be processed and the training data, the ability of a point cloud denoiser to extract local features and the topological constraint correction during the inference process are used to reduce the impact of the distribution difference on the generation result.

[0037] The SVG formation image is rasterized and pixel-normalized to obtain a normalized SVG rendering image. This normalized SVG rendering image is input into a conditional encoder, and undergoes multi-layer downsampling processing through a convolutional coding network to obtain a spatial feature map. This spatial feature map is then converted into a spatial token sequence, forming a conditional feature sequence. An initial Gaussian noise point set is generated based on a preset number of drones, and this initial Gaussian noise point set is used as the current noise point set for the first denoising time step. The point cloud denoiser employs a multi-layer modulation transform network to embed the coordinates of the points in the current noise point set, performs sinusoidal time encoding on the current denoising time step, and generates feature linear modulation parameters based on the time encoding results. The multi-layer modulation transform network uses the feature linear modulation parameters to adjust the input features of the current layer, extracts inter-point correlation features through self-attention operations, fuses the conditional feature sequence into the inter-point correlation features through cross-attention operations, and obtains the noise prediction result based on the output features of the final layer.

[0038] Based on the current set of noisy points, the noise prediction results, and the noise scheduling parameters corresponding to the current denoising time step, Tweedie estimation is performed to obtain the current clean point estimation set. The Euclidean distance between each pair of points in the current clean point estimation set is calculated. For point pairs with an Euclidean distance less than a preset safety threshold, a spacing penalty is determined to obtain the safety spacing loss. For each point in the current clean point estimation set, a preset number of nearest neighbor points are determined. Based on the distance between each point and its corresponding neighbor, continuous adjacency weights are determined, and a current nearest neighbor weight matrix is ​​constructed. The connectivity preservation loss is obtained based on the structural difference between the current nearest neighbor weight matrix and the reference adjacency matrix. The safety spacing loss and connectivity preservation loss are weighted and fused according to preset spacing constraint weights and preset connectivity constraint weights to obtain the topological constraint loss. The constraint gradient of the topological constraint loss relative to the coordinates of each point in the current clean point estimation set is calculated.

[0039] The constraint signal signal-to-noise ratio (SNR) for the current denoising time step is obtained and compared with a preset activation threshold. When the constraint signal SNR does not reach the preset activation threshold, no constraint gradient is introduced, and basic denoising updates are performed based on the noise prediction results. When the constraint signal SNR reaches the preset activation threshold, a topological constraint correction is generated based on the constraint gradient, and the direction of the basic denoising update is corrected using this correction. This ensures that the denoising stage with a high noise ratio mainly forms point distribution, while the denoising stage with a low noise ratio introduces safety distance constraints and topological connectivity constraints. During inference, a deterministic denoising method using a denoising diffusion implicit model is adopted. Starting from the initial Gaussian noise point set, denoising updates are performed according to the preset inference time step sequence, and the updated point set corresponding to the last denoising time step is determined as the generated point set. Each point pair in the generated point set is traversed, and point pairs with a distance less than the preset safety threshold are subjected to mutually repulsive displacements, and distance detection and coordinate updates are repeatedly performed. When distance conflicts still exist after reaching the preset maximum number of iterations, compensation displacements are applied to the corresponding points to obtain a safe point set that meets the preset safety threshold.

[0040] This invention employs a technical approach encompassing scalable vector graphics semantic extraction, point generation, constraint embedding, and secure post-processing. Each processing stage exhibits strong independence and replaceability, allowing for the selection of different implementation methods based on application scenarios, computing resources, and deployment requirements. The conditional encoder can be configured among convolutional neural networks, visual transformers, and path sequence encoders, enabling feature extraction from formation graphics using raster images or directly utilizing the geometric semantics of path sequences, thus improving adaptability to different input formats. The point cloud denoiser can be replaced with other network structures capable of point set modeling, ensuring the point generation process is compatible with UAV formation tasks of varying scales and complexities. Connectivity constraints can be selected from nearest neighbor adjacency graphs, Delaunay triangulation graphs, or neighborhood graphs based on point density and formation shape, facilitating accurate representation of local connectivity relationships and maintaining formation topological integrity. The sampling process can choose between inverse transform sampling, importance resampling, and hierarchical sampling, and weight coefficients can be adjusted through grid search or Bayesian optimization to adapt point distribution to different curvatures, endpoints, and inflection point features. Post-processing of the spacing can employ force-oriented algorithms, Lagrange penalty constraint optimization, or spring-damped physical simulation to improve the flexibility and stability of point correction without altering the target safe spacing. The point generation framework can also be replaced by flow matching or fractional matching from the diffusion probability model, and different accelerated sampling methods can be configured to balance generation accuracy, inference speed, and computational overhead. The two-dimensional point generation method can be applied to three-dimensional space by expanding the coordinate dimensions and distance calculation methods without changing the fundamental principles of nearest neighbor relationship construction and safe post-processing, thereby expanding the applicability of this invention in scenarios such as UAV light formation, three-dimensional inspection, emergency search and rescue, and spatial collaborative deployment.

[0041] The technical solutions in the above embodiments, by constructing an end-to-endpoint position generation process for SVG formation graphics, can directly complete the automatic conversion from graphic path description to UAV cluster point coordinates, reducing the time and operational costs required for manual point-by-point calibration and repeated correction, and improving formation planning efficiency. The front-end parsing process can convert SVG path instructions into computable geometric primitives, and identify geometric feature points such as curve anchor points, curvature maxima, curve inflection points, endpoints, and intersections based on curvature, angles, intersection relationships, and arc length positions, thereby forming a structured topological skeleton, providing clear geometric semantics and topological basis for point generation. The adaptive curvature geometry sampling method determines importance weights based on local curvature, curve inflection point distance, and endpoint attributes, prioritizing the distribution of a limited number of UAV points in geometrically dense areas. Combined with inverse transformation sampling, geometric feature point protection, and spacing conflict resolution, it reduces the omission of key geometric positions and redundancy of points in simple path areas, which is beneficial to improving the recognizability and geometric fidelity of formation graphics. The topology-constrained guided diffusion model generates points through conditional feature encoding and time-step denoising. It then uses Tweedie estimation to obtain the current clean point estimate set. When the constraint signal-to-noise ratio (SNR) reaches a preset activation threshold, it introduces safety distance loss and connectivity preservation loss, ensuring that the point generation process is simultaneously guided by graphical conditions, point spacing, and topological connectivity, reducing excessive point clustering, local connection breaks, and formation degradation. The selective activation method based on constraint signal SNR reduces the interference of unreliable constraints on the point distribution formation process during high-noise stages and strengthens safety distance and topological connectivity constraints during medium- and low-noise stages, improving the stability of denoising updates. After force-guided distance correction and displacement compensation, the generated points further eliminate residual spacing conflicts, ensuring the final points meet the preset safety threshold. Both point generation methods share SVG parsing and safety post-processing, allowing selection of the appropriate generation method based on real-time performance, graphical complexity, and safety requirements, thus balancing generation efficiency, formation quality, topological integrity, deployment security, and application flexibility.

[0042] In one embodiment, such as Figure 2 As shown, parsing the path instructions in the SVG formation graph to obtain the structured topological skeleton and the reference adjacency matrix includes the following steps S11-S13: In step S11, the movement instructions, straight line instructions, quadratic Bézier curve instructions, cubic Bézier curve instructions, circular arc instructions and closed path instructions in the SVG formation graphics are read one by one. The path instructions are converted into at least one of straight line segments, quadratic Bézier curve segments, cubic Bézier curve segments and circular arc segments. The coordinates of each geometric primitive obtained by conversion are normalized to the preset canvas range. In step S12, based on the endpoint positions, curvature extrema, curvature sign changes, angles between adjacent line segments, intersection relationships between different path strokes, and arc length positions of each geometric element, at least one of the following is identified: curve anchor point, curvature maximum point, curve inflection point, non-closed path endpoint, acute angle vertex, obtuse angle vertex, high curvature point, intersection point, midpoint, and trisection point, to obtain the geometric feature points. In step S13, geometric feature points whose distance is not greater than a preset deduplication threshold are merged, and the feature category labels corresponding to each geometric feature point before merging are retained. Based on the merged geometric feature points, the geometric primitives to which each geometric feature point belongs, and the connection relationships between each geometric primitive, the structured topological skeleton is obtained, and a reference adjacency matrix corresponding to the preset number of UAVs is generated based on the structured topological skeleton.

[0043] In one embodiment, after obtaining the SVG formation graphics, the path instructions are parsed one by one according to their order in the file. Move instructions determine the starting position of the current path stroke; straight line instructions generate straight line segments based on the starting and ending coordinates; quadratic Bézier curve instructions generate quadratic Bézier curve segments based on the starting, control point, and ending coordinates; cubic Bézier curve instructions generate cubic Bézier curve segments based on the starting coordinates, two control point coordinates, and the ending coordinates; arc instructions generate arc segments based on arc parameters; and closed path instructions connect the ending and starting points of the current path stroke. The parsed straight line segments, quadratic Bézier curve segments, cubic Bézier curve segments, and arc segments are then identified as geometric primitives, and the coordinates of each geometric primitive are transformed to a preset canvas area, ensuring that SVG formation graphics of different sizes use a unified coordinate scale for subsequent processing.

[0044] Geometric feature points are identified based on the geometric parameters and connectivity of each geometric primitive. The start and end positions of quadratic and cubic Bézier curve segments and circular arc segments are identified as curve anchor points. Curve maxima are identified based on the curvature changes along the arc length of the geometric primitives, and the locations where the curvature sign changes are identified as curve inflection points. The start and end points of non-closed path strokes are identified as non-closed path endpoints. For vertices formed by adjacent straight line segments, the included angle between adjacent line segments is calculated. Vertices with included angles less than a preset acute angle threshold are identified as acute vertices, and vertices with included angles greater than a preset obtuse angle threshold are identified as obtuse vertices. When the curvature corresponding to a vertex exceeds a preset curvature threshold, the vertex is simultaneously identified as a high curvature point. The intersection positions between geometric primitives corresponding to different path strokes are detected. Intersection positions that do not belong to shared endpoints are identified as intersection points, and the midpoints and trisection points are identified based on the arc length positions of the geometric primitives to obtain geometric feature points.

[0045] Calculate the Euclidean distance between any two geometric feature points. Merge geometric feature points whose Euclidean distance is not greater than a preset deduplication threshold, and retain the feature category labels corresponding to each geometric feature point before merging, so that the same location can simultaneously have multiple feature category labels such as curve anchor point, high curvature point, or intersection point. Associate the merged geometric feature points with their respective geometric primitives, recording the arrangement, endpoint connection, closure, and intersection relationships of each geometric primitive in the corresponding path strokes, to obtain a structured topology skeleton. Configure a corresponding number of reference nodes on the structured topology skeleton according to the preset number of drones. Determine whether reference nodes are adjacent based on the adjacency relationship of reference nodes along the same path strokes and the connection relationship between different geometric primitives, and write the adjacency relationship between each reference node into a reference adjacency matrix, thus obtaining a reference adjacency matrix corresponding to the preset number of drones.

[0046] In one embodiment, such as Figure 3 As shown, obtaining the conditional feature sequence and the noise prediction result includes the following steps S21-S23: In step S21, the normalized SVG rendering image is subjected to multi-layer downsampling processing through a convolutional coding network to obtain a spatial feature map containing the features of different local regions of the SVG formation graphic, and the spatial feature map is converted into a spatial token sequence to obtain the conditional feature sequence. In step S22, coordinate embedding is performed on the coordinates of the points in the current noise point set, sinusoidal time encoding is performed on the denoising time step, and characteristic linear modulation parameters are generated based on the time encoding result to obtain the point embedding features and the modulation parameters corresponding to the denoising time step; In step S23, the point embedding features are used as the first layer input features to the multilayer modulation transform network. In each layer, the modulation parameters are used to perform feature linear modulation on the current layer input features. The inter-point correlation features are extracted through self-attention operation. The conditional feature sequence is fused to the inter-point correlation features through cross-attention operation. The current layer output features are input to the next layer. The noise prediction result is obtained based on the final layer output features.

[0047] In one embodiment, a normalized SVG rendering image is input into a conditional encoder, which uses a convolutional coding network to perform multi-layer downsampling processing on the normalized SVG rendering image. Each convolutional layer extracts edge morphology, path direction, local curvature, and spatial relationships between different path regions from the normalized SVG rendering image. The downsampling process progressively expands the feature perception range, resulting in a spatial feature map containing local spatial features at different scales. Based on the arrangement of spatial positions in the spatial feature map, the feature vectors corresponding to each spatial position are converted into spatial tokens, and all spatial tokens constitute a conditional feature sequence, allowing the conditional feature sequence to retain feature information from different local regions of the SVG formation graphic.

[0048] Coordinate embedding is performed on the coordinates of each point in the current noise point set, mapping the two-dimensional point coordinates to a preset feature dimension to obtain point embedding features that correspond one-to-one with each point in the current noise point set. Sine time encoding is performed on the current denoising time step, using sine and cosine functions of different frequencies to characterize the position of the current denoising time step in the denoising process. Based on the time encoding results, feature linear modulation parameters corresponding to the current denoising time step are generated. The feature linear modulation parameters include scaling parameters for adjusting the feature response amplitude and offset parameters for adjusting the feature offset.

[0049] The point embedding features are used as the first-layer input features to the multilayer modulation transform network. Each layer uses feature linear modulation parameters to scale and shift the current layer's input features, adapting them to the current denoising time step. Self-attention operations are used to calculate the correlation between different points in the current noise point set, obtaining inter-point correlation features reflecting the relative distribution and inter-point dependencies. Through cross-attention operations, using the inter-point correlation features as query information and the conditional feature sequence as key and value information, the local spatial features of the SVG formation graphic are fused into the inter-point correlation features. The current layer's output features are then fed forward into the next layer. The multilayer modulation transform network predicts the noise components contained in the current noise point set based on the final layer's output features, obtaining the noise prediction result corresponding to the current noise point set.

[0050] In one embodiment, such as Figure 4 As shown, the topological constraint loss is obtained through the following steps S31-S34: Based on the current set of noise points, the noise prediction results, and the noise scheduling parameters corresponding to the denoising time step, the estimated coordinates of each UAV point in a noise-free state are determined by Tweedie estimation, thus obtaining the current set of clean point estimates. In step S32, the Euclidean distance between each pair of points in the current clean point estimation set is calculated. When the Euclidean distance between the pair of points is less than the preset safety threshold, the spacing penalty of the corresponding pair of points is determined according to the difference between the preset safety threshold and the Euclidean distance. When the Euclidean distance between the pair of points is not less than the preset safety threshold, the spacing penalty of the corresponding pair of points is determined to be zero, and the safety spacing loss is obtained according to the spacing penalty of each pair of points. In step S33, for each point in the current clean point estimation set, a preset number of nearest neighboring points are determined, and the continuous adjacency weights are determined based on the distance between each point and its corresponding neighboring points. A current nearest neighbor weight matrix is ​​constructed, and the structural difference between the current nearest neighbor weight matrix and the reference adjacency matrix is ​​calculated to obtain the connectivity preservation loss. In step S34, the safety distance loss and the connectivity preservation loss are weighted and fused according to the preset spacing constraint weight and the preset connectivity constraint weight to obtain the topology constraint loss, and the constraint gradient of the topology constraint loss relative to the coordinates of each point in the current clean point estimation set is calculated.

[0051] In one embodiment, Tweedie estimation is performed based on the current set of noisy points, the noise prediction results, and the noise scheduling parameters corresponding to the current denoising time step. Tweedie estimation determines the estimated coordinates in a noise-free state from the coordinates of noisy points and the corresponding noise prediction results, and aggregates these estimated coordinates according to the point arrangement in the current set of noisy points to obtain the current set of clean point estimates. Topological constraint loss uses the current set of clean point estimates as the calculation object, ensuring that safety distance constraints and topological connectivity constraints apply to a point distribution close to a noise-free state, avoiding distortion of constraint direction caused by directly calculating constraints based on the coordinates of points with a high proportion of noise.

[0052] For the safety distance loss, each point pair in the current clean point estimation set is traversed, and the Euclidean distance between the coordinates of the two points in each point pair is calculated. This Euclidean distance is then compared to a preset safety threshold. When the Euclidean distance is less than the preset safety threshold, a distance penalty is determined based on the difference between the preset safety threshold and the Euclidean distance; the larger the difference, the larger the corresponding distance penalty. When the Euclidean distance is not less than the preset safety threshold, the distance penalty for the corresponding point pair is set to zero. The distance penalties for all point pairs are aggregated to obtain the safety distance loss. The gradient corresponding to the safety distance loss pushes the two points away from each other along the line connecting point pairs with insufficient distance, while point pairs that meet the preset safety threshold do not have a distance constraint effect.

[0053] For the connectivity preservation loss, for each point in the current clean point estimation set, a preset number of nearest neighbor points are determined according to Euclidean distance. Based on the distance between the point and each neighbor point, a preset distance mapping relationship is used to determine the continuous adjacency weights, and all continuous adjacency weights are written into the current nearest neighbor weight matrix according to the point correspondence relationship. The current nearest neighbor weight matrix is ​​compared with the reference adjacency matrix generated based on the SVG formation graph, and the structural differences of corresponding elements in the two matrices are aggregated to obtain the connectivity preservation loss. According to the preset spacing constraint weights and preset connectivity constraint weights, the safety spacing loss and connectivity preservation loss are weighted and fused to obtain the topology constraint loss, and the constraint gradient of the topology constraint loss relative to the coordinates of each point in the current clean point estimation set is calculated. The constraint gradient is used to push away points with insufficient spacing in subsequent denoising updates and reduce the difference between the current nearest neighbor connection relationship and the reference connection relationship to maintain the safety spacing and topological integrity of the UAV cluster points.

[0054] In one embodiment, such as Figure 5 As shown, the denoising update direction is corrected based on the constraint signal-to-noise ratio, including the following steps S41-S43: In step S41, the signal component and noise component corresponding to the current denoising time step are obtained, the constraint signal-to-noise ratio is determined according to the ratio between the signal component and the noise component, and the constraint signal-to-noise ratio is compared with the preset activation threshold. In step S42, when the signal-to-noise ratio of the constraint signal does not reach the preset activation threshold, the constraint gradient is not introduced in the current denoising time step, and basic denoising update is performed according to the noise prediction result to obtain the updated point set without topological constraint correction. In step S43, when the signal-to-noise ratio of the constraint signal reaches the preset activation threshold, a topology constraint correction amount is generated according to the constraint gradient, and the basic denoising update direction of the current noise point set is corrected using the topology constraint correction amount to obtain the updated point set after topology constraint correction. The updated point set without topology constraint correction or the updated point set after topology constraint correction is used as the current noise point set for the next denoising time step.

[0055] In one embodiment, signal components and noise components are obtained based on noise scheduling parameters corresponding to the current denoising time step, and the constraint signal-to-noise ratio (SNR) is determined based on the ratio between the signal components and the noise components. The constraint SNR characterizes the reliability of the current clean location estimation set. When the proportion of noise components in the current denoising time step is high, the constraint SNR is low, and the location distribution in the current clean location estimation set is not yet stable; when the proportion of signal components in the current denoising time step is high, the constraint SNR is high, and the current clean location estimation set can reliably reflect the distribution status of UAV locations. The constraint SNR is compared with a preset activation threshold, and based on the comparison result, it is determined whether the constraint gradient participates in the denoising update of the current denoising time step.

[0056] When the signal-to-noise ratio of the constraint signal does not reach the preset activation threshold, no constraint gradient is introduced in the current denoising time step. Basic denoising updates are performed on the current set of noise points based on the noise prediction results. The basic denoising update determines the change in point coordinates based on the current set of noise points, the noise prediction results, and the noise scheduling parameters corresponding to the current denoising time step. The current set of noise points is then updated according to these changes, resulting in an updated set of points without topological constraint correction. This process ensures that the denoising stage with a high proportion of noise forms a point distribution according to the noise prediction results, avoiding interference from the current clean point estimation set with lower reliability on the denoising update direction.

[0057] When the signal-to-noise ratio of the constraint signal reaches a preset activation threshold, a topology constraint correction amount is generated based on the constraint gradient of the topology constraint loss relative to the coordinates of each point in the current clean point estimation set. This topology constraint correction amount is then applied to the basic denoising update direction in the direction that reduces the topology constraint loss. The basic denoising update direction of the current noisy point set is corrected using this topology constraint correction amount, resulting in a topology-corrected updated point set. The updated point set without topology constraint correction or the updated point set with topology constraint correction is used as the current noisy point set for the next denoising time step. This ensures that basic denoising updates are maintained during high-noise stages, while safety distance constraints and topology connectivity constraints are introduced during medium- and low-noise stages, until the preset denoising time step sequence is completed.

[0058] In one embodiment, such as Figure 6 As shown, the denoising update direction is corrected based on the constraint signal-to-noise ratio, including the following steps S51-S54: In step S51, a deterministic denoising method based on the denoising diffusion implicit model is adopted. Starting from the initial Gaussian noise point set, denoising updates are performed according to the preset inference time step sequence. Based on the comparison results between the signal-to-noise ratio of the constraint signal corresponding to each denoising time step and the preset activation threshold, topological constraint correction is selectively performed, and the updated point set corresponding to the last denoising time step is determined as the generated point set. In step S52, each pair of points in the generated point set is traversed. When the distance between the point pairs is less than the preset safety threshold, the repulsion displacement is determined according to the difference between the preset safety threshold and the distance. Two opposite repulsion displacement directions are determined according to the direction of the line connecting the point pairs. The coordinates of the two points in the point pairs are updated according to the preset damping coefficient. In step S53, point pair traversal and coordinate update are repeatedly performed on the point set after coordinate update. When there are no point pairs with a distance less than the preset safety threshold, the coordinate update is terminated. When there are still point pairs with a distance less than the preset safety threshold after reaching the preset maximum number of iterations, a compensation displacement along the direction of the line connecting the point pairs is applied to at least one point in the corresponding point pair so that the distance between the point pairs reaches the preset safety threshold, thereby obtaining the safe point set. In step S54, based on the scale ratio between the normalized canvas range and the UAV deployment area, the coordinate origin offset, and the coordinate axis direction, the coordinates of each point in the set of safe points are scaled, translated, and transformed to obtain the coordinates of the UAV cluster points.

[0059] In one embodiment, a deterministic denoising method using a denoising diffusion implicit model is employed to process the initial Gaussian noise point set. Following a preset inference time step sequence, the initial Gaussian noise point set is determined as the current noise point set for the first denoising time step, and denoising updates are performed at each denoising time step based on noise prediction results. The constraint signal-to-noise ratio (SNR) for each denoising time step is obtained and compared with a preset activation threshold. When the constraint signal SNR does not reach the preset activation threshold, no constraint gradient is introduced, and basic denoising updates are performed based on noise prediction results. When the constraint signal SNR reaches the preset activation threshold, the basic denoising update direction is corrected using topological constraint correction, allowing safety distance constraints and topological connectivity constraints to participate in the point coordinate update for the current denoising time step. The updated point set obtained in the current denoising time step is used as the current noise point set for the next denoising time step, until the preset inference time step sequence is completed, and the updated point set corresponding to the last denoising time step is determined as the generated point set.

[0060] The process iterates through each pair of points in the generated point set, calculates the distance between the two points in each pair, and compares this distance with a preset safety threshold. When the distance is less than the preset safety threshold, a repulsion displacement is determined based on the difference between the preset safety threshold and the distance; the larger the difference, the larger the repulsion displacement. Two opposite repulsion displacement directions are determined based on the direction of the line connecting the two points in the pair. The repulsion displacement is scaled according to a preset damping coefficient, and the coordinates of the two points in the pair are updated along the two opposite repulsion displacement directions, causing the two points to move away from each other. The point pair traversal, distance comparison, and coordinate update are repeated for the point set with updated coordinates. The coordinate update is terminated when there are no point pairs with a distance less than the preset safety threshold. If, after reaching a preset maximum number of iterations, there are still point pairs with a distance less than the preset safety threshold, a compensation displacement is applied along the direction of the line connecting the point pairs to at least one point in the corresponding point pair, so that the distance between the corresponding point pairs reaches the preset safety threshold, thus obtaining a safe point set. This process serves as a safety guarantee after topological constraint correction and can eliminate local spacing conflicts that still exist during the denoising generation process.

[0061] Obtain the size parameters of the normalized canvas area and the drone deployment area. Determine the scale ratio based on the size relationship between the two, and obtain the coordinate origin offset and coordinate axis direction corresponding to the drone deployment area. Scale the coordinates of each point in the safety point set according to the scale ratio to adapt the point distribution in the normalized canvas to the actual size of the drone deployment area. Translate the scaled point coordinates according to the coordinate origin offset to move the point distribution to the target position in the drone deployment area. Perform coordinate axis transformation on the translated point coordinates according to the coordinate axis direction to ensure that the coordinate direction used by the normalized canvas is consistent with the coordinate direction used by the drone deployment area. Output the coordinates of each point after scaling, translation, and coordinate axis transformation according to the corresponding drone point number to obtain the drone swarm point coordinates suitable for drone formation deployment.

[0062] In one embodiment, Figure 7 This is a block diagram illustrating a drone swarm location generation device based on SVG formation graphics, according to an exemplary embodiment. Figure 7 As shown, the UAV cluster location generation device based on SVG formation graphics includes an acquisition module 71, a generation module 72, an estimation module 73, an update module 74, and a mapping module 75.

[0063] The acquisition module 71 is used to acquire SVG formation graphics, parse the path instructions in the SVG formation graphics to obtain the geometric primitives that make up the SVG formation graphics, extract geometric feature points based on the geometric primitives and parse the topological relationships between the geometric primitives to obtain a structured topological skeleton and a reference adjacency matrix corresponding to a preset number of UAVs, and perform rasterization rendering and pixel normalization processing on the SVG formation graphics to obtain a normalized SVG rendering image. The generation module 72 is used to input the normalized SVG rendering image into the conditional encoder for multi-layer feature encoding to obtain a conditional feature sequence for characterizing the local spatial features of the SVG formation graphic, and to generate an initial Gaussian noise point set based on the preset number of drones. The estimation module 73 is used to input the current set of noise points, the denoising time step and the conditional feature sequence into the point cloud denoiser for each denoising time step to obtain the noise prediction result corresponding to the current set of noise points, and to perform Tweedie estimation based on the current set of noise points, the noise prediction result and the noise scheduling parameters corresponding to the denoising time step to obtain the current clean point estimation set. The update module 74 is used to obtain the constraint signal signal-to-noise ratio (SNR) of the current denoising time step. When the constraint signal SNR reaches a preset activation threshold, it calculates the safety distance loss based on the distance between each pair of points in the current clean point estimation set, calculates the connectivity preservation loss based on the nearest neighbor relationship of the current clean point estimation set and the reference adjacency matrix, performs weighted fusion of the safety distance loss and the connectivity preservation loss to obtain the topology constraint loss, and corrects the denoising update direction of the current noisy point set based on the topology constraint loss. When the constraint signal SNR does not reach the preset activation threshold, it performs basic denoising update on the current noisy point set based on the noise prediction result, and uses the corrected denoising update result or the basic denoising update result as the current noisy point set for the next denoising time step for iterative processing to obtain the generated point set. The mapping module 75 is used to detect the distance between each pair of points in the generated point set, apply displacement to the point pairs whose distance is less than a preset safety threshold to make the point pairs move away from each other, and repeat the distance detection and displacement processing on the point set after displacement to obtain a safe point set that meets the preset safety threshold. The safe point set is then mapped from normalized canvas coordinates to UAV deployment coordinates to obtain the UAV cluster point coordinates.

[0064] The acquisition module 71, generation module 72, estimation module 73, update module 74, and mapping module 75 included in the block diagram of the UAV cluster location generation device based on SVG formation graphics are controlled to execute the UAV cluster location generation method based on SVG formation graphics described in any of the above embodiments.

[0065] like Figure 8 As shown, the present invention provides an electronic device 800, which includes: a communication interface, a processor 801, and a memory 802; The memory 802 stores program instructions. When executed by the processor 801, which is connected to the memory 802 via the communication interface, the program instructions parse the formation graphic path in the scalable vector graphics format, extract geometric primitives, geometric feature points, and topological relationships, and generate a structured topological skeleton, a reference adjacency matrix, and a normalized rendering map. The normalized rendering map is feature-encoded, and denoised step-by-step using an initial Gaussian noise point set. A clean point estimation set is obtained through Tweedie estimation. The denoising direction is corrected by selectively introducing safety distance loss and connectivity preservation loss based on the constraint signal-to-noise ratio. The point set is iteratively generated, and then the coordinates of the UAV cluster points are obtained through distance detection, rejection displacement, and coordinate mapping.

[0066] This invention provides a computer-readable storage medium storing computer program instructions. When executed by a processor, the computer program instructions parse the formation graphic path in a scalable vector graphics format, extract geometric primitives, geometric feature points, and topological relationships, and generate a structured topological skeleton, a reference adjacency matrix, and a normalized rendering map. The normalized rendering map is then feature-encoded, and denoised step-by-step using an initial Gaussian noise point set. A clean point estimation set is obtained through Tweedie estimation. Based on the constraint signal-to-noise ratio, a safety distance loss and connectivity preservation loss are selectively introduced to correct the denoising direction. The point set is iteratively generated, and then, through distance detection, rejection displacement, and coordinate mapping, the coordinates of the UAV cluster points are obtained.

[0067] It should be understood that the specific features, operations, and details described above regarding the method of the present invention can also be similarly applied to the apparatus and system of the present invention, or vice versa. Furthermore, each step of the method of the present invention described above can be performed by a corresponding component or unit of the apparatus or system of the present invention.

[0068] It should be understood that the various modules / units of the device of the present invention can be implemented wholly or partially through software, hardware, firmware, or a combination thereof. Each module / unit can be embedded in the processor of a computer device in hardware or firmware form or independent of the processor, or it can be stored in the memory of a computer device in software form for the processor to call to execute the operation of each module / unit. Each module / unit can be implemented as an independent component or module, or two or more modules / units can be implemented as a single component or module.

[0069] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores computer instructions executable by the processor, which, when executed by the processor, instruct the processor to perform steps of the methods of embodiments of the present invention. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the methods of the present invention.

[0070] This invention can be implemented as a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0071] It will be understood by those skilled in the art that the method steps of the present invention can be performed by a computer program instructing related hardware, such as a computer device or processor. The computer program may be stored in a non-transitory computer-readable storage medium, and its execution causes the steps of the present invention to be performed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0072] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating UAV swarm locations based on SVG formation graphics, characterized in that, include: S1. Obtain the SVG formation graphic, parse the path instructions in the SVG formation graphic to obtain the geometric primitives that make up the SVG formation graphic, extract geometric feature points based on the geometric primitives and parse the topological relationship between the geometric primitives to obtain the structured topological skeleton and the reference adjacency matrix corresponding to the preset number of drones, and perform rasterization rendering and pixel normalization processing on the SVG formation graphic to obtain the normalized SVG rendering image. S2. Input the normalized SVG rendering image into the conditional encoder for multi-layer feature encoding to obtain a conditional feature sequence for characterizing the local spatial features of the SVG formation graphic, and generate an initial Gaussian noise point set according to the preset number of drones. S3. For each denoising time step, input the current noise point set, the denoising time step, and the conditional feature sequence into the point cloud denoiser to obtain the noise prediction result corresponding to the current noise point set. Then, perform Tweedie estimation based on the current noise point set, the noise prediction result, and the noise scheduling parameters corresponding to the denoising time step to obtain the current clean point estimation set. S4. Obtain the constraint signal signal-to-noise ratio (SNR) of the current denoising time step. When the constraint signal SNR reaches a preset activation threshold, calculate the safety distance loss based on the distance between each pair of points in the current clean point estimation set, calculate the connectivity preservation loss based on the nearest neighbor relationship of the current clean point estimation set and the reference adjacency matrix, perform weighted fusion of the safety distance loss and the connectivity preservation loss to obtain the topology constraint loss, and correct the denoising update direction of the current noisy point set based on the topology constraint loss. When the constraint signal SNR does not reach the preset activation threshold, perform basic denoising update on the current noisy point set based on the noise prediction result, and use the corrected denoising update result or the basic denoising update result as the current noise point set for the next denoising time step for iterative processing to obtain the generated point set. S5. Detect the distance between each pair of points in the generated point set, apply displacement to the point pairs whose distance is less than a preset safety threshold to make the point pairs move away from each other, and repeat the distance detection and displacement processing on the point set after displacement to obtain a safe point set that meets the preset safety threshold. Map the safe point set from normalized canvas coordinates to UAV deployment coordinates to obtain the UAV cluster point coordinates.

2. The method for generating UAV cluster locations based on SVG formation graphics as described in claim 1, characterized in that, Parsing the path instructions in the SVG formation graph to obtain the structured topological skeleton and the reference adjacency matrix includes: S11. Read the movement instructions, straight line instructions, quadratic Bézier curve instructions, cubic Bézier curve instructions, circular arc instructions and closed path instructions in the SVG formation graphics one by one, convert the path instructions into at least one of straight line segments, quadratic Bézier curve segments, cubic Bézier curve segments and circular arc segments, and normalize the coordinates of each geometric primitive obtained by conversion to a preset canvas range. S12. Based on the endpoint positions, curvature extrema, curvature sign changes, angles between adjacent line segments, intersection relationships between different path strokes, and arc length positions of geometric primitives, identify at least one of the following: curve anchor points, curvature maximum points, curve inflection points, non-closed path endpoints, acute angle vertices, obtuse angle vertices, high curvature points, intersection points, midpoints, and trisection points, to obtain the geometric feature points. S13. Merge geometric feature points whose distance is not greater than a preset deduplication threshold, and retain the feature category labels corresponding to each geometric feature point before merging. Based on the merged geometric feature points, the geometric primitives to which each geometric feature point belongs, and the connection relationships between each geometric primitive, obtain the structured topological skeleton, and generate a reference adjacency matrix corresponding to the preset number of UAVs based on the structured topological skeleton.

3. The method for generating UAV cluster locations based on SVG formation graphics as described in claim 2, characterized in that, Obtaining the conditional feature sequence and the noise prediction result includes: S21. Perform multi-layer downsampling processing on the normalized SVG rendering image through a convolutional coding network to obtain a spatial feature map containing the features of different local regions of the SVG formation graphic, and convert the spatial feature map into a spatial token sequence to obtain the conditional feature sequence. S22. The coordinates of the points in the current noise point set are embedded, the denoising time step is sinusoidally time-coded, and the characteristic linear modulation parameters are generated according to the time coding results to obtain the point embedding features and the modulation parameters corresponding to the denoising time step. S23. The point embedding features are used as the first layer input features and input to the multilayer modulation transform network. In each layer, the modulation parameters are used to perform feature linear modulation on the current layer input features. The point correlation features are extracted through self-attention operation. The conditional feature sequence is fused to the point correlation features through cross-attention operation. The current layer output features are input to the next layer. The noise prediction result is obtained based on the final layer output features.

4. The method for generating UAV cluster locations based on SVG formation graphics as described in claim 1, characterized in that, The topological constraint loss is obtained by: S31. Based on the current set of noise points, the noise prediction results, and the noise scheduling parameters corresponding to the denoising time step, the estimated coordinates of each UAV point in a noise-free state are determined by Tweedie estimation, and the current set of clean point estimates is obtained. S32. Calculate the Euclidean distance between each pair of points in the current clean point estimation set. When the Euclidean distance between the pairs of points is less than the preset safety threshold, determine the spacing penalty of the corresponding pair of points based on the difference between the preset safety threshold and the Euclidean distance. When the Euclidean distance between the pairs of points is not less than the preset safety threshold, determine the spacing penalty of the corresponding pair of points to be zero, and obtain the safety spacing loss based on the spacing penalty of each pair of points. S33. For each point in the current clean point estimation set, determine the nearest preset number of neighboring points, determine the continuous adjacency weights based on the distance between each point and its corresponding neighboring points, construct the current nearest neighbor weight matrix, calculate the structural difference between the current nearest neighbor weight matrix and the reference adjacency matrix, and obtain the connectivity preservation loss. S34. According to the preset spacing constraint weight and the preset connectivity constraint weight, the safety spacing loss and the connectivity preservation loss are weighted and fused to obtain the topological constraint loss, and the constraint gradient of the topological constraint loss relative to the coordinates of each point in the current clean point estimation set is calculated.

5. The method for generating UAV cluster locations based on SVG formation graphics as described in claim 4, characterized in that, The denoising update direction is corrected based on the constraint signal-to-noise ratio, including: S41. Obtain the signal component and noise component corresponding to the current denoising time step, determine the constraint signal-to-noise ratio according to the ratio between the signal component and the noise component, and compare the constraint signal-to-noise ratio with the preset activation threshold. S42. When the signal-to-noise ratio of the constraint signal does not reach the preset activation threshold, the constraint gradient is not introduced in the current denoising time step, and basic denoising update is performed according to the noise prediction result to obtain the updated point set without topological constraint correction. S43. When the signal-to-noise ratio of the constraint signal reaches the preset activation threshold, a topology constraint correction amount is generated according to the constraint gradient, and the basic denoising update direction of the current noise point set is corrected using the topology constraint correction amount to obtain the updated point set after topology constraint correction. The updated point set without topology constraint correction or the updated point set after topology constraint correction is used as the current noise point set for the next denoising time step.

6. The method for generating UAV cluster locations based on SVG formation graphics as described in claim 5, characterized in that, Obtaining the set of safe locations and the coordinates of the drone cluster locations includes: S51. Using a deterministic denoising method with a denoising diffusion implicit model, starting from the initial Gaussian noise point set, denoising updates are performed according to a preset inference time step sequence. Based on the comparison result between the signal-to-noise ratio of the constraint signal corresponding to each denoising time step and the preset activation threshold, topological constraint correction is selectively performed, and the updated point set corresponding to the last denoising time step is determined as the generated point set. S52. Traverse each pair of points in the generated point set. When the distance between the pairs of points is less than the preset safety threshold, determine the repulsion displacement based on the difference between the preset safety threshold and the distance. Determine two opposite repulsion displacement directions based on the direction of the line connecting the pairs of points, and update the coordinates of the two points in the pairs of points according to the preset damping coefficient. S53. Repeat the point pair traversal and coordinate update for the point set after coordinate update. Terminate coordinate update when there are no point pairs with a distance less than the preset safety threshold. If there are still point pairs with a distance less than the preset safety threshold after reaching the preset maximum number of iterations, apply a compensation displacement along the direction of the line connecting the point pairs to at least one point in the corresponding point pair so that the distance between the point pairs reaches the preset safety threshold, thereby obtaining the safe point set. S54. Based on the scale ratio between the normalized canvas range and the UAV deployment area, the coordinate origin offset, and the coordinate axis direction, perform scale transformation, translation transformation, and coordinate axis transformation on the coordinates of each point in the set of safe points to obtain the coordinates of the UAV cluster points.

7. A device for generating drone swarm locations based on SVG formation graphics, characterized in that, include: The acquisition module is used to acquire SVG formation graphics, parse the path instructions in the SVG formation graphics to obtain the geometric primitives that make up the SVG formation graphics, extract geometric feature points based on the geometric primitives and parse the topological relationships between the geometric primitives to obtain a structured topological skeleton and a reference adjacency matrix corresponding to a preset number of drones, and perform rasterization rendering and pixel normalization processing on the SVG formation graphics to obtain a normalized SVG rendering image. The generation module is used to input the normalized SVG rendering image into the conditional encoder for multi-layer feature encoding to obtain a conditional feature sequence for characterizing the local spatial features of the SVG formation graphic, and to generate an initial Gaussian noise point set based on the preset number of drones. The estimation module is used to input the current set of noise points, the denoising time step and the conditional feature sequence into the point cloud denoiser for each denoising time step to obtain the noise prediction result corresponding to the current set of noise points, and to perform Tweedie estimation based on the current set of noise points, the noise prediction result and the noise scheduling parameters corresponding to the denoising time step to obtain the current clean point estimation set. The update module is used to obtain the constraint signal signal-to-noise ratio (SNR) of the current denoising time step. When the constraint signal SNR reaches a preset activation threshold, the module calculates the safety distance loss based on the distance between each pair of points in the current clean point estimation set, calculates the connectivity preservation loss based on the nearest neighbor relationship of the current clean point estimation set and the reference adjacency matrix, performs weighted fusion of the safety distance loss and the connectivity preservation loss to obtain the topological constraint loss, and corrects the denoising update direction of the current noisy point set based on the topological constraint loss. When the constraint signal SNR does not reach the preset activation threshold, the module performs basic denoising update on the current noisy point set based on the noise prediction result, and uses the corrected denoising update result or the basic denoising update result as the current noisy point set for the next denoising time step for iterative processing to obtain the generated point set. The mapping module is used to detect the distance between each pair of points in the generated point set, apply displacement to the point pairs whose distance is less than a preset safety threshold to make the point pairs move away from each other, and repeat the distance detection and displacement processing on the point set after displacement to obtain a safe point set that meets the preset safety threshold. The safe point set is then mapped from normalized canvas coordinates to UAV deployment coordinates to obtain the UAV cluster point coordinates.

8. The UAV cluster location generation device based on SVG formation graphics as described in claim 7, characterized in that: The acquisition module, the generation module, the estimation module, the update module, and the mapping module are controlled to execute the UAV cluster location generation method based on SVG formation graphics as described in any one of claims 2 to 6.

9. An electronic device, characterized in that, include: Communication interface, processor, memory; The memory is used to store program instructions, which, when executed by the processor connected to the memory via the communication interface, cause the electronic device to implement the UAV cluster location generation method based on SVG formation graphics as described in any one of claims 1 to 6.

10. A computer-readable storage medium having program instructions stored thereon, characterized in that, When the program instructions are executed by a computer, the computer implements the method for generating drone cluster locations based on SVG formation graphics as described in any one of claims 1 to 6.

Images