Unmanned aerial vehicle performance picture generation method and device, equipment and medium

Abstract

The application discloses a kind of unmanned plane performance picture generation method, device, equipment and medium, the unmanned plane performance picture generation method directly generates the line drawing image in accordance with unmanned plane performance specification by multi-modal generation model based on special data set training according to text or image input.The line drawing is analyzed, and the contour skeleton and internal filling area are separated.In the dot distribution stage, the contour key inflection point is innovatively identified and locked as a fixed anchor point, ensuring that the geometric features are not lost in subsequent processing;The optimal spacing is adaptively calculated by combining binary search, and the constrained space relaxation is applied in the filling area.Finally, through local safety evaluation, only the minimum amount of addition and deletion is performed on the point position on the safety line segment, and the target flight is accurately matched.The problem of low production efficiency caused by too much human operation is effectively solved, and the mass production demand of high-frequency commercial performance is fully met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cluster UAV control, and in particular to a UAV performance picture generation method, device, equipment and medium. BACKGROUND

[0002] In the field of UAV cluster performance, as the scale of the performance develops to "ten thousand", the design and arrangement of the performance picture has become a key bottleneck restricting the commercialization and scale development of the industry. The generation method under the existing technology has a serious man-machine imbalance problem, and the core contradiction is that the high-end hardware execution capability does not match the low-end manual design process, resulting in insufficient overall efficiency to meet the high-frequency and customized commercial performance needs. The generation of this bottleneck is due to the inherent defects of the existing technology in two core links.

[0003] Firstly, in the pattern generation link, the existing scheme mainly relies on art designers to manually draw using drawing software or to generate creative drafts using general AI text-to-image tools. However, manual drawing is extremely inefficient and cannot produce high-quality designs that meet customer requirements in a short time; while the general AI text-to-image tool can quickly generate images, but the generated content is usually realistic, oil painting, etc. artistic style, not the simple line drawing style required for UAV performance. More importantly, the training data of these general models is unrelated to the UAV performance scene, and there is no direct correlation between the complexity of the generated image and the specific number of UAVs specified by the customer. Designers must repeatedly adjust the prompt words and perform post-processing to make the visual weight of the generated pattern roughly match the fleet size. This process not only consumes time and effort, but also highly depends on human experience and cannot achieve standardization and automation.

[0004] Secondly, in the queue arrangement, i.e. UAV point arrangement link, the two-dimensional pattern needs to be discretized into specific coordinate points of thousands of UAVs, and this process also involves serious manual participation. Existing technologies can achieve basic equidistant point arrangement, but face two major challenges when dealing with complex patterns. First, geometric feature distortion, when applying a spatial relaxation algorithm to the point array after point arrangement to optimize the distance between points and improve flight safety, existing algorithms usually treat all points equally, resulting in the inward contraction of active points on the contour edge during the optimization process, thus smoothing the key geometric features such as sharp corners and turning points of the pattern, greatly reducing artistic expression, which must be manually repaired by engineers. Second, lack of quantity accuracy, the total number of points needs to be completely matched with the actual number of UAVs, but existing methods often fail to achieve accurate results due to improper parameter settings. When there is a slight difference between the point arrangement result and the target number of UAVs, the system lacks an intelligent fine-tuning mechanism, and engineers can only adjust the initial parameters repeatedly to re-arrange the points or perform tedious manual addition and subtraction operations, which is extremely inefficient.

[0005] In summary, due to the fact that the prior art relies on a large amount of inefficient and error-prone manual intervention at key stages, the production cycle of a performance plan takes as long as several days. Against this background, how to break down the barriers and build a fully automated method that can integrate AI creative generation and intelligent precise placement has become a technical problem that needs to be broken through in this field. SUMMARY

[0006] Embodiments of the present application provide a UAV performance picture generation method, device, equipment and medium, aiming to solve the technical problem that the pattern design and queue arrangement of the UAV performance picture generation method in the prior art need a large amount of manual operation, which cannot meet the existing commercial performance frequency requirements.

[0007] In a first aspect, embodiments of the present application provide a UAV performance picture generation method, which comprises: based on an input text description or a control image, outputting a line drawing image conforming to a UAV performance visual specification through a multi-modal generation model; performing structural analysis on the line drawing image to separate a contour skeleton layer and an internal filling layer; based on the geometric features of the contour skeleton layer, identifying and locking key inflection points as fixed anchor points, and calculating an optimal placement distance suitable for the overall image; according to the placement distance, performing equidistant sampling on the contour skeleton layer and retaining the fixed anchor points, generating an initial dot matrix on the internal filling layer and performing constraint space relaxation to form a preliminary dot graph; performing local safety evaluation on the preliminary dot graph, screening adjustable line segments, and based on a target performance number of times, adjusting the number of dot positions to output a final dot graph matching the target performance number of times.

[0008] In a second aspect, embodiments of the present application further provide a UAV performance picture generation device for executing the UAV performance picture generation method as described above.

[0009] In a third aspect, embodiments of the present application further provide a computer device, which comprises a memory and a processor connected to the memory; the memory is used to store a computer program; the processor is used to run the computer program stored in the memory to execute the steps of the above-mentioned UAV performance picture generation method.

[0010] In a fourth aspect, embodiments of the present application further provide a computer readable storage medium, which stores a computer program, and the computer program comprises program instructions, which, when executed by a processor, can implement the steps of the above-mentioned UAV performance picture generation method.

[0011] Compared with the prior art, the present application has the following advantages:

[0012] In the technical scheme of the present application, the unmanned aerial vehicle performance picture generation method directly generates a line drawing image conforming to the unmanned aerial vehicle performance specification according to text or image input through a multi-modal generation model trained based on a special data set. The line drawing is structurally analyzed to separate the contour skeleton and the internal filling area. In the dot placement stage, the contour key inflection points are innovatively identified and locked as fixed anchor points to ensure that the geometric features are not lost in subsequent processing; the optimal spacing is adaptively calculated in combination with binary search, and the constraint space relaxation is applied in the filling area. Finally, through local safety evaluation, only the minimum amount of addition and deletion of point positions is performed on the safe line segment to accurately match the target number of sorties. The pattern modification, feature repair and quantity adjustment links in the traditional process that rely on manual operation are replaced, effectively solving the problem of low production efficiency caused by too many human operations, and fully meeting the mass production demand of high-frequency commercial performances. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical scheme in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0014] Figure 1 The flowchart of the unmanned aerial vehicle performance picture generation method provided by the present application;

[0015] Figure 2 The first sub-flowchart of the unmanned aerial vehicle performance picture generation method provided by the present application;

[0016] Figure 3 The second sub-flowchart of the unmanned aerial vehicle performance picture generation method provided by the present application;

[0017] Figure 4 The third sub-flowchart of the unmanned aerial vehicle performance picture generation method provided by the present application;

[0018] Figure 5 The fourth sub-flowchart of the unmanned aerial vehicle performance picture generation method provided by the present application;

[0019] Figure 6 The fifth sub-flowchart of the unmanned aerial vehicle performance picture generation method provided by the present application;

[0020] Figure 7 The sixth sub-flowchart of the unmanned aerial vehicle performance picture generation method provided by the present application;

[0021] Figure 8 The seventh sub-flowchart of the unmanned aerial vehicle performance picture generation method provided by the present application;

[0022] Figure 9A schematic block diagram of the units of the UAV performance picture generation apparatus provided by the present application;

[0023] Figure 10 A schematic block diagram of the computer device provided by the embodiment of the present application;

[0024] Figure 11 The input graph (left) and the separated contour and fill graph (right) of the first example of the UAV performance picture generation method provided by the present application;

[0025] Figure 12 The line curvature heat map of the first example of the UAV performance picture generation method provided by the present application;

[0026] Figure 13 The feature point graph of the first example of the UAV performance picture generation method provided by the present application;

[0027] Figure 14 The dot placement graph before Voronoi relaxation of the first example of the UAV performance picture generation method provided by the present application;

[0028] Figure 15 The dot placement graph after Voronoi relaxation of the first example of the UAV performance picture generation method provided by the present application;

[0029] Figure 16 The input graph (left) and the extracted contour graph (right) of the second example of the UAV performance picture generation method provided by the present application;

[0030] Figure 17 The line curvature heat map of the second example of the UAV performance picture generation method provided by the present application;

[0031] Figure 18 The feature point graph of the second example of the UAV performance picture generation method provided by the present application;

[0032] Figure 19 The dot placement graph before Voronoi relaxation of the second example of the UAV performance picture generation method provided by the present application;

[0033] Figure 20 The dot placement graph after Voronoi relaxation of the second example of the UAV performance picture generation method provided by the present application. DETAILED DESCRIPTION

[0034] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts should fall within the scope of the present application.

[0035] It should be understood that the terms "comprise" and "include" as used in the specification and the appended claims indicate the presence of the described features, integers, steps, operations, elements, and / or components but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0036] It should also be understood that the terms used in the present application specification are only for the purpose of describing the embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0037] It should be further understood that the term "and / or" as used in the present application specification and the appended claims means one or more of the associated listed items as well as all possible combinations of the items.

[0038] The present application discloses a method for generating a performance picture of unmanned aerial vehicles to solve the technical problem that the pattern design and queue arrangement of the existing method for generating a performance picture of unmanned aerial vehicles require a large amount of manual operation, which cannot meet the existing commercial performance frequency requirements.

[0039] With reference to Figures 1 to 8 and Figures 11 to 20 , the method comprises the following steps:

[0040] S110, based on the input text description or control image, outputting a line drawing image conforming to the visual specification of the unmanned aerial vehicle performance through a multi-modal generation model;

[0041] S120, performing structural analysis on the line drawing image to separate a contour skeleton layer and an internal filling layer;

[0042] S130, based on the geometric features of the contour skeleton layer, identifying and locking key inflection points as fixed anchor points, and calculating an optimal point distribution distance suitable for the overall image;

[0043] S140, according to the point distribution distance, performing equidistant sampling on the contour skeleton layer and retaining the fixed anchor points, generating an initial dot matrix on the internal filling layer and performing constraint space relaxation, and forming a preliminary dot image;

[0044] S150, performing a local security evaluation on the preliminary point graph, screening adjustable line segments, and adjusting the number of point positions based on a target performance number of times, and outputting a final point graph matching the target performance number of times.

[0045] The entire process begins with the user providing creative input. When the user inputs a text description, the system calls a text-to-image multi-modal generation model trained on a dedicated data set; when the user uploads a brand logo image and inputs editing instructions, the system calls a image-to-image multi-modal generation model. Both models are trained using targeted data sets, with the text-to-image data set containing tens of thousands of line drawings of different complexity levels and their corresponding Chinese and English descriptions, ensuring that the model can associate the information in the text to the generation of high complexity patterns. The image-to-image data set contains paired data of control images and target line drawings, enabling the model to faithfully transfer the structure of the input image. After the multi-modal generation model runs, it outputs a standard format line drawing image that meets the visual specifications of the drone performance, which is a pure line drawing with black lines on a white background and does not contain any color or texture, and can be directly used for subsequent processing.

[0046] After the line drawing image is generated, the system analyzes its structure. First, the image is binarized, classifying all pixels as foreground (black lines) or background (white). Then, the Euclidean distance transform algorithm is used to process the binarized image, calculating the distance from each foreground pixel to the nearest background pixel and generating a distance map. Based on the gray scale distribution of the distance map, the system separates the image into two logical layers with a preset threshold. The area with smaller distance values constitutes the contour skeleton layer, representing the center path of the line. The area with larger distance values constitutes the internal filling layer, representing the solid part inside the line. This lays the foundation for subsequent differentiated processing.

[0047] In the deployment preparation stage, the system performs key feature protection and spacing calculation based on the separated contour skeleton layer. The system first traces the bitmap of the contour skeleton layer into a set of ordered vector polylines. Then, it performs Gaussian smoothing on these polylines to eliminate noise and calculates the curvature of each vertex. Points with curvature values significantly higher than their surrounding vertices are identified as key inflection points, such as the tips of graphics or sharp corners. The system immediately marks these key inflection points as fixed anchors and divides the original contour line into multiple independent line segment units using these points as boundaries, ensuring that the fixed anchors remain in place during subsequent processing and effectively preventing them from being smoothed. At the same time, the system initiates a binary search algorithm to calculate the optimal deployment spacing. In each iteration, the system simulates the number of equidistant sampling points for the contour skeleton layer and the number of hexagonal grid points for the internal fill layer based on the current spacing value, and then optimizes the fill point distribution using the constrained Lloyd relaxation algorithm to calculate the sum of the two. The system dynamically adjusts the search range based on the deviation of the total number from the target performance number of times until it finds the optimal spacing that makes the total number closest to the target number of times.

[0048] After obtaining the optimal deployment spacing and fixed anchors, the system generates the preliminary point map. For the contour skeleton layer, the system performs equidistant sampling along each independent line segment unit with the optimal deployment spacing, and the generated sampling point sequence explicitly retains all locked fixed anchors. For the internal fill layer, the system generates an initial point array of a hexagonal grid within its range and then applies constrained spatial relaxation. All sampling points of the contour skeleton layer, including fixed anchors, are included as fixed constraint points, allowing only the points of the internal fill layer to move towards the centroid of their Voronoi cells in iterations, thereby making the internal point distribution more uniform while maintaining the shape of the external contour, avoiding communication interference between drones. The point set formed in this way is the preliminary point map.

[0049] Finally, the system performs finishing touches on the preliminary point map. The system first performs local safety evaluation, using spatial indexing to analyze the point density of each line segment of the contour layer, and identifies line segments with adjacent point spacing greater than a preset safety threshold. These line segments are marked as adjustable line segments that can be resampled. The system calculates the difference between the total number of points in the current preliminary point map and the target performance number of times. If there is a difference, the system preferentially performs the minimum amount of point addition or deletion on the adjustable line segments. For example, if points need to be added, the longest adjustable line segment is resampled from n equal parts to n+1 equal parts. This operation can accurately achieve quantity matching and ensure that all new points meet the minimum flight safety spacing without damaging the overall image topology.

[0050] The final point graph after fine-tuning has a total number of points strictly equal to the target performance number of times, and can be directly used to generate UAV flight instructions. In specific implementation, a collision rehearsal module based on a physics engine is also added before the final output, which simulates the paths of all UAVs flying from standby positions to the corresponding point positions of the final point graph, automatically detects potential flight path conflicts, and performs slight offset optimization on the conflict point positions, thereby providing double safety protection.

[0051] In an embodiment, the step of S110 includes:

[0052] S111, constructing a plurality of line drawing data sets and a plurality of graph drawing data sets according to a plurality of performance level grades of the UAV cluster, the complexity of the line drawings in the line drawing data sets and the graph drawing data sets matching the corresponding performance level grades;

[0053] S112, training a line drawing model and a graph drawing model based on the line drawing data sets and the graph drawing data sets, respectively;

[0054] S113, when the user input is a text description generated from an image, invoking and outputting a line drawing image through the line drawing model;

[0055] S114, when the user input is a control image and a semantic editing text instruction, invoking and outputting a line drawing image through the graph drawing model.

[0056] To ensure that the patterns generated by AI are consistent with the actual performance scale, the system divides the UAV cluster into a plurality of performance level grades according to common scales. For the line drawing scenario, a line drawing data set is independently constructed for each performance level grade, each data set containing at least ten thousand line drawing images and their accurate Chinese and English descriptions. The key is that the complexity of the images is positively correlated with the number of UAVs in the corresponding level. The complexity includes parameters such as total line length, number of intersection points, and filled area. For example, the fourth level data set for training the "Happy Birthday" cake scenario has more detailed line drawings than the second level data set, including more decorative patterns and more complex spatial layout. For the graph drawing scenario, a unified graph drawing data set is constructed, which contains three hundred thousand image pairs and is equipped with semantic editing text instructions such as "convert to simple line drawing" or "enhance contour". This data set does not distinguish between performance levels and focuses on learning structure migration and instruction following ability.

[0057] After the dataset is constructed, the model training begins. Using the text-to-image dataset described above, the base multi-modal model such as Qwen-Image is fine-tuned to generate multiple specialized text-to-image models, each corresponding to a specific level of complexity, for example, model A is specialized for generating complex patterns for 1500-2000 drones, and model B is specialized for generating simple patterns for 300-800 drones. The training process follows the typical procedure of diffusion models, minimizing the mean square error of latent variable prediction, so that the model learns to internalize information such as "suitable for 1500-2000 drones" in the text description as control of the complexity of the generated pattern. At the same time, another base model such as Qwen-Image-Edit-2509 is trained using the image-to-image dataset, so that it can receive a control image and a set of text instructions to jointly generate a stylized line drawing.

[0058] In the online application stage, for example, when the user inputs a text description of a cartoon lion "a smiling cartoon lion, suitable for 1500-2000 drones", the system automatically parses the information "1500-2000 drones", selects the corresponding fourth-level specialized model from the trained text-to-image model, and calls the model to execute generation, ensuring that the output cartoon lion line drawing has rich details such as hair wrinkles and lively expressions, and the line density is sufficient to match high-altitude drones. For example, when the user uploads a real cake photo with "Happy Birthday" and inputs the instruction "convert to 2000 drone level minimalist line drawing", the system recognizes the presence of both image and text input, and calls the specially trained image-to-image model. The model first extracts the main structure from the cake photo, such as the round cake, the position of the candles, and the arrangement of the text, and then generates a line drawing image with clean lines, clear structure, and complexity that matches the "minimalist line drawing" and "2000 drone" instructions. Through this dual-model architecture driven by specialized datasets, the system realizes the automatic generation of line drawing images from user creativity with high matching degree, effectively solving the problem of content mismatch generated by general models.

[0059] In an embodiment, the step of S120 comprises:

[0060] S121, performing a binaryzation process on the line drawing image to obtain the pixel distribution of the foreground and background;

[0061] S122, calculating the distance from each foreground pixel to the nearest background edge by distance transform algorithm to form a distance map;

[0062] S123, based on a preset distance threshold, dividing the line drawing image into a contour skeleton layer and an internal filling layer.

[0063] When the multi-modal generative model outputs the labeled format line drawing image, first, an image preprocessing procedure is started. The system performs a binarization process on the line drawing image, and converts the image into a pure black and white dot array by setting a global threshold. For example, the global threshold sets the pixels with a gray value lower than 128 as foreground, with a value of 0, representing black lines; and sets the pixels with a value higher than 128 as background, with a value of 255, representing white areas. This operation eliminates the anti-aliasing gray edges that may exist in the original image, and obtains a clear foreground and background pixel distribution, laying a foundation for accurate distance calculation.

[0064] After binarization is completed, the system processes the foreground pixels using the Euclidean distance transform algorithm. The core of this algorithm is to calculate the Euclidean distance from each foreground pixel in the line drawing image, i.e., the point in the black line, to the nearest background pixel, i.e., the white area. After the calculation is completed, a gray-scale image with the same size as the original image is generated, which is called a distance map. In this distance map, the gray value of a pixel represents its distance to the background edge, and therefore presents a clear gradient change: the pixels on the central axis of the line have the farthest distance and the largest value, presenting the brightest gray; the pixels close to the edge of the line have smaller values and darker gray; and the background pixels have a value of 0, presenting pure black.

[0065] Based on this distance map, the system performs the final division operation. The system sets a preset distance threshold, which is defined as a certain intermediate proportion of the minimum and maximum values in the distance map. In implementation, the system traverses each pixel in the distance map. If the value of a foreground pixel in the distance map is less than or equal to the threshold, it is determined that the pixel is located in the edge region of the line, and the corresponding positions of these pixels in the original line drawing image are classified into the contour skeleton layer; if the value of the pixel is greater than the threshold, it is determined that the pixel is located in the central region of the line, and is classified into the internal filling layer. For example, when analyzing the circular ear of a cartoon lion, the narrow strip-shaped area close to the edge of the ear in the distance map is classified into the contour skeleton layer, which is responsible for defining the accurate path of the “lion ear”; and the wider central part of the ear is classified into the internal filling layer, which is responsible for generating the internal drone dot array in the subsequent process. Similarly, for the bold text in the “Happy Birthday” cake image, the hollow part of the strokes or the central region of the wider lines is also identified as the internal filling layer, and the outer boundary of the text is retained in the contour skeleton layer. Through this automatic division method based on distance transform, the system can intelligently and accurately separate the geometric skeleton and the filling region of the image, providing clear and reliable calculation basis for subsequent equidistant dot placement on the contour and uniform dot placement in the interior.

[0066] In an embodiment, the step S130 comprises:

[0067] S131, tracing the bitmap of the contour skeleton layer into an ordered vectorized line segment, and identifying the intersection nodes of the line segment;

[0068] S132. After performing Gaussian smoothing on each of the quantified line segments, calculate the curvature distribution and detect the curvature peak points;

[0069] S133. Determine the key inflection points as the curvature peak points, and break the line segment at the key inflection points to form independent line segment units;

[0070] S134. Set all the key inflection points as fixed anchor points.

[0071] Through a contour tracing algorithm based on pixel connectivity, such as Suzuki's connected component analysis algorithm, convert the bitmap of the contour skeleton layer composed of discrete pixel points into one or more sets of ordered vectorized polylines composed of vertex sequences. During this process, the algorithm will automatically analyze the topological structure of the lines, accurately identify the starting point, ending point of the line segment, and the "line segment intersection nodes" formed by the intersection of multiple line segments. For example, in the text "Happy Birthday", the intersection point of the two vertical lines and the horizontal line of the character "日", or the intersection point of the lion's hair and the lion's head in the cartoon lion. These intersection nodes are marked and retained in the vector data, providing a basis for subsequent precise segmentation.

[0072] After obtaining the vectorized line segments, the system enters the key feature protection process. To accurately calculate the geometric curvature, the system first performs Gaussian smoothing on each vector line segment. This process convolves a one-dimensional Gaussian kernel on the vertex sequence, effectively filtering out the tiny noise and jaggedness introduced by image quantization or tracing, making the subsequent curvature calculation results smoother and more reliable. After the smoothing process, the system calculates the discrete curvature of each vertex on each line segment. Commonly used curvature calculation methods are the chord length method or the tangent angle method: taking a certain vertex as V, and taking one adjacent vertex before and after as V- and V+ respectively, by calculating the angle θ between the vector V-V and the vector VV+, and dividing by the average value of the lengths of these two vectors, as the approximate curvature value of vertex V. The system traverses all vertices to form a curvature distribution curve that changes along the path.

[0073] The system detects the curvature peak points by setting a dynamic curvature threshold. In the line drawing of the cartoon lion, these peak points are precisely located at the tips of the lion's ears, the vertices of the nose, and the tips of the hair; in the "Happy Birthday" cake image, they appear at the curling points of the text strokes, the tips of the candle flames, etc. The system determines the points with curvatures significantly higher than the surrounding areas as key inflection points. Subsequently, the system performs a breaking operation on the original vector polyline at each key inflection point, that is, taking this inflection point as the boundary, dividing a long line segment into two independent line segment units. For example, the contour line of one ear of the lion is divided into two units, the upper edge line and the lower edge line.

[0074] Finally, the system will promote all the identified key inflection points from normal vertex data to fixed anchor objects with special properties. These fixed anchors are endowed with the property of unchangeable position and are marked in memory, ensuring that their three-dimensional coordinates will be forcibly locked in all subsequent iterations of isometric sampling and spatial relaxation, and any algorithmic operation attempting to move them will be ignored. Through this embodiment, the system protects the core artistic features of the pattern at the algorithmic level, providing fundamental technical guarantees to prevent fidelity loss.

[0075] Further, the step of S130 further comprises:

[0076] S135, estimating an initial distribution point spacing range according to the total length of the contour skeleton layer and the area of the internal filling layer;

[0077] S136, simulating distribution points through bisection search iteration in the initial distribution point spacing range, and calculating the sum of the contour sampling point number and the filling area grid point number as the total point number in each iteration;

[0078] S137, dynamically adjusting the search upper and lower bounds according to the deviation value of the current total point number and the target performance number of times, until the optimal distribution point spacing is searched.

[0079] The system estimates a reasonable initial distribution point spacing search range based on the geometric information of the image. The system calculates the total length of the contour skeleton layer and the total area of the internal filling layer respectively. The total length is obtained by calculating the cumulative distance of the vertices of all vectorized line segment units. The total area is determined by calculating the area of the internal filling layer polygon or counting the pixels of the internal filling layer and multiplying the physical area of the pixels. Based on these two values and the preset target performance number of times, the system uses an empirical estimation formula to determine the initial spacing range. For example, assuming that the target number of times is 1800, the system can estimate that the contour distribution points will occupy about 600 times, and the filling distribution points will occupy about 1200 times, then according to the density formula d ≈ sqrt(A / N*ρ) to estimate the initial spacing of the filling area, and combined with the contour length to estimate the initial spacing of the contour, finally take the dominant term in the two to determine a reasonable initial spacing minimum value and maximum value. Wherein, d represents the estimated initial distribution point spacing, unit: pixel, A: represents the total area of the internal filling layer, unit: pixel²; ρ represents the estimated distribution point density coefficient per unit area, which is an empirical value determined by the size of the unmanned aerial vehicle, the minimum safety spacing and other system parameters.

[0080] After determining the initial range, the system initiates a binary search algorithm to efficiently approach the optimal solution. In each iteration, the system takes the median value of the current range as the candidate point placement interval for this iteration. Subsequently, the system performs a full-flow point placement simulation: on the contour skeleton layer, the system samples along all segmented independent line segment units at the current candidate interval, and forces to keep all pre-set fixed anchors, and accumulates the contour sampling point number for this iteration. On the internal filling layer, the system generates a hexagonal honeycomb initial point lattice with the current candidate interval as the grid interval, and executes one or more rounds of constrained Voronoi relaxation on the initial point lattice with all fixed anchors in the region as constraint points, and counts the number of stable filling region grid points after relaxation. The system adds the contour sampling point number and the filling region grid point number to obtain the total point number for this iteration.

[0081] The system compares the total point number for this iteration with the pre-set target performance number of times. If the total point number is greater than the target number of times, it means that the interval is too small and the point placement is too dense, so the upper bound of the search range is reduced to the current median value; if the total point number is less than the target number of times, it means that the interval is too large and the point placement is too sparse, so the lower bound is raised to the current median value. Through this closed-loop feedback mechanism, the search range quickly converges after each iteration. The system continues to perform binary search until the width of the search range is less than a pre-set convergence threshold, or the absolute deviation of the total point number from the target number of times is less than an acceptable error. At this time, the system outputs the candidate interval used in the last iteration as the optimal point placement interval. The interval value is passed to the subsequent point placement module for actual generation of the preliminary point graph. Through this adaptive algorithm based on binary search, the system can intelligently and efficiently find a globally optimal point placement density, ensuring the accuracy of the total number of points from the source, and solving the problem of invalid regular interval calculation caused by the introduction of fixed anchors.

[0082] In an embodiment, the steps of S140 include:

[0083] S141, sampling along each line segment of the contour skeleton layer at the optimal point placement interval, and keeping all the fixed anchors;

[0084] S142, generating a polygon grid point lattice as an initial point position in the internal filling layer;

[0085] S143, setting the sampling points of the contour skeleton layer as constraint points, allowing only the internal filling points in the region of the internal filling layer as active points, and moving the centroids of the Voronoi cells based on the internal filling points iteratively to make the internal filling points perform constrained spatial relaxation, and obtaining a preliminary point graph.

[0086] The system performs accurate point placement on the contour skeleton layer. The system iterates through all the segmented independent line segment units in the contour skeleton layer. For example, the outer contour of a cartoon lion, the facial feature lines, or the text strokes and candle outlines in a "Happy Birthday" cake image. For each independent line segment, the system starts from the beginning of the line segment and samples along the vectorized path with the computed optimal point placement interval as the step size, generating a set of initial sampling points evenly distributed along the line segment. The most critical step is that the system explicitly preserves and inserts all the marked fixed anchor points in the sequence of equidistant sampling. For example, at the tip of the lion's hair or the turning vertex of the text "Happy", even if the equidistant sampling does not exactly fall on this point, the system will force the fixed anchor point to be added to the final set of position points, and adjust the distribution of nearby sampling points according to its position, ensuring that the key inflection points defining the core features of the image are accurately lit by the drone. After this operation, the point positions on all contour skeleton layers are finally determined.

[0087] Next, the system turns to the internal fill layer and constructs the formation of the internal drones. In each separated internal fill layer region, such as the eye circles, cheek solid areas of the cartoon lion, or the wide stroke inside the cake candle and decorative stripes, the system generates an initial point array with high density. The initial point array is arranged in a hexagonal grid, with an initial grid interval consistent with the optimal point placement interval, or slightly smaller to ensure coverage. This grid point array constitutes the initial positions of the internal fill points.

[0088] Subsequently, the system starts the constrained spatial relaxation engine to optimize the internal point array, prevent drone collisions, and improve visual aesthetics. The system first marks all the sampling points of the contour skeleton layer as constraint points, which will remain unchanged in subsequent iterations. The system also marks all the initial points of the internal fill layer as active points.

[0089] Next, the system begins the iterative calculation: in each iteration, it constructs a Voronoi diagram for all current points, finds the Voronoi cell corresponding to each active point, which is the polygon region enclosed by the perpendicular bisectors of the active point and all neighboring points, and calculates the geometric centroid of the cell.

[0090] Then, the system moves the active point one step towards its centroid position, with the moving step size controlled by the convergence factor. Since the positions of the constraint points are fixed, their Voronoi boundaries will not move, so when the active point moves towards the centroid, it will naturally form an even distance with the contour points, but will not shrink inward across the boundary, thus perfectly preserving the original shape of the contour. This iterative process continues until the movement distance of all active points is less than a preset convergence threshold, and the point array distribution reaches a stable state.

[0091] Finally, the complete set of all fixed contour skeleton layer sampling points and all optimized internal filling points constitutes a "preliminary point map". The preliminary point map not only accurately reproduces the original artistic form of the image to be performed to express, especially retains all the key visual features, but also has uniform internal point distribution, meets the minimum safety distance requirement of the UAV cluster flight, and is the direct input basis for subsequent quantity adjustment.

[0092] In an embodiment, the step of S150 includes:

[0093] S151, performing local point density analysis on each line segment of the contour skeleton layer, and identifying a line segment with a distance between adjacent points greater than a preset safety threshold as an adjustable line segment through spatial indexing;

[0094] S152, determining the number of points to be added or reduced according to the difference between the target performance episode and the total number of points in the preliminary point map;

[0095] S153, only on the adjustable line segment, resampling based on the path length of the adjustable line segment, performing minimum point addition and deletion, and ensuring that the newly added point position meets the minimum safety distance constraint and does not destroy the image topology.

[0096] The system performs local safety evaluation on the contour skeleton layer constituting the preliminary point map, which aims to identify line segments that meet the safety condition and have adjustment flexibility. The system traverses each independent line segment unit in the contour skeleton layer, which is the result of being broken at the key inflection points, such as each ear contour of the cartoon lion, the arm curve, or each stroke of "fast" in the "Happy Birthday" cake image, and the vertical line of the candle. For each line segment, the system analyzes the local density of all point positions distributed thereon. The system calculates the arc length distance between any two adjacent points on the line and compares it with a globally preset minimum safety distance threshold. This threshold is determined by the physical size of the UAV itself and the minimum safety interval that can be guaranteed by the flight control system.

[0097] The system efficiently identifies all line segments with a distance between adjacent points greater than the safety threshold by constructing a spatial indexing structure such as an R-tree. These line segments have a physical space for adding or deleting points due to their longer length and lower point density, and are selected and marked as adjustable line segments by the system. It is worth noting that those short line segments located at acute angles, tips, or bearing core features will be automatically excluded from the adjustable range because their point distance may be short due to fixed anchor points, thereby protecting the topology of the image.

[0098] After the system acquires the list of adjustable line segments, it enters the number fine-tuning stage. The system first counts the total number of points in the preliminary point graph, including the sampling points of the contour skeleton layer and the relaxed points of the internal filling layer, and calculates the difference ΔN between the total number and the user's preset "target performance number of times". If ΔN is zero, the preliminary point graph can be directly output as the final point graph. If ΔN is not zero, the system will start the fine-tuning program. If ΔN > 0, it means that the number of points is too small, and |ΔN| points need to be added. The system will select the longest |ΔN| line segments from the list of adjustable line segments. If ΔN < 0, it means that the number of points is too large, and |ΔN| points need to be deleted. In this case, the longest |ΔN| line segments are also selected for point deletion. The selection of the longest line segments is to minimize the adjustment range and minimize the impact on the overall shape.

[0099] The specific addition and deletion operations are realized by resampling the selected line segments. For an adjustable line segment selected for adding points, the system breaks the equal division relationship of its current n points, resamples according to n+1 equal division, and removes the original points, replacing them with n+1 new points. The newly added points are automatically distributed according to the equal distance rule. This resampling process ensures that the distance between the newly added points and the neighboring points is greater than or equal to the minimum safety distance constraint. The operation of deleting points is similar, changing n equal division to n-1 equal division. The system only allows this operation to be performed on paths marked as adjustable line segments, fundamentally avoiding modification in key feature areas and ensuring that the topological structure of the image is not damaged. After resampling, the total number of points in the preliminary point graph is exactly equal to the target performance number of times.

[0100] Finally, the system determines the fine-tuned point array as the final point graph. Each point in the point graph represents an exact three-dimensional coordinate position that a UAV needs to reach during performance, and the total number is exactly the same as the actual size of the UAV cluster. This scheme realizes zero-error matching of the number of points under the premise of ensuring high fidelity of the pattern and flight safety, providing a reliable technical foothold for one-key automatic generation.

[0101] In an embodiment, the UAV performance picture generation method further includes:

[0102] S160, mapping the two-dimensional coordinates of all points in the final point graph to three-dimensional space coordinates and outputting a control file recognizable by a cluster UAV flight control system.

[0103] The system first acquires the two-dimensional coordinate data of all points in the final point map. These coordinates are defined in the image processing stage, located in the two-dimensional Cartesian coordinate system (x, y), and the unit is usually pixels. In order to adapt to the real flight environment, the system needs to map these two-dimensional coordinates into the three-dimensional geographic coordinate system. The system reads the preset performance venue parameters, which include the central geographic coordinates of the virtual performance plane, the inclination of the performance plane relative to the ground, and the actual physical distance corresponding to a single pixel. Based on these parameters, the system performs coordinate conversion: the two-dimensional pixel coordinates (x, y) of each point are first converted to plane distance in meters according to the scale, and then superimposed on the central geographic coordinates of the performance plane to generate new longitude and latitude values, while the set flight height is taken as the Z-axis coordinate. After this mapping, each point in the final point map is assigned a unique three-dimensional spatial coordinate, accurately indicating the aerial position where each UAV should hover during the performance.

[0104] After completing the three-dimensional coordinate mapping, the system enters the file output stage. The system formats all UAV three-dimensional coordinate data according to the standard protocol required by the cluster UAV flight control system. Common output formats include plain text files (.txt) or comma-separated value files (.csv), where each row represents a UAV, arranged in a preset fixed order (or a preset ID order), containing its corresponding longitude, latitude, and altitude values.

[0105] In more advanced embodiments, the output file can also be a project file specific to the flight control software. In addition to the static point map of the final hovering points, this file can also be expanded to include time dimension information. For example, while generating the final point map, the system can calculate the smooth flight trajectories of all UAVs from the ground assembly point to their target point simultaneously, encapsulate these trajectory data with the point map data, and generate a dynamic control file containing the complete flight path. This control file can be transmitted to the master computer or base station of the UAV cluster through wired or wireless means. After the flight control system loads the file, it can decode and distribute instructions to guide each UAV to accurately fly to the designated location, and finally accurately present the complete pattern in the night sky.

[0106] To more clearly illustrate the UAV performance picture generation method of the present application, the following two examples are described respectively:

[0107] In the first example, refer to Figures 16 to 20The user selects the text-to-image mode in the graphical interface of the system and inputs the text description: "A beautiful birthday cake with the prominent Chinese characters 'Happy Birthday' in the middle, all the text parts are filled solid lines, suitable for 4000 drones performance." This text description not only defines the performance content and visual style, but more importantly, indicates the number of drone flights, providing the core parameters for subsequent automated processing.

[0108] After receiving the input, the system starts the text-to-image generation engine. The system first parses the information "suitable for 4000 drones" from the text, according to the preset flight level division rule, 4000 belongs to the high complexity level of "4000-4500 flights". The system immediately calls a Qwen-Image text-to-image multi-modal generation model specially trained for this level. The model has learned a large number of high-density line drawing samples and their corresponding text descriptions in the offline training stage, and can internalize the "4000 flights" information as accurate control of the complexity of the generated image. After the model runs, a PNG format line drawing image that meets the requirements is generated. The image presents a circular layered cake, with the Chinese characters "Happy Birthday" in the center of the top layer, all lines are black, the background is pure white, and the strokes of "Happy Birthday" are explicitly designed as solid filled areas to meet the user's visual demand of "internal filling color".

[0109] Subsequently, the system enters the image structure analysis stage. The algorithm first performs binaryzation processing on the generated PNG line drawing, accurately separating the foreground (black lines) and background (white). Then, the Euclidean distance transform algorithm is used to calculate the distance from each foreground pixel to the nearest background edge, generating a distance map. The system uses the Otsu algorithm to automatically determine a global threshold, and divides the image into a contour skeleton layer and an internal filling layer. The outer contour of "Happy Birthday" and the outer edge of the cake are classified as the contour skeleton layer, while the hollow rectangular area of the "birth" stroke and the interior of the "day" character are identified and labeled as the internal filling layer.

[0110] The next step is the key geometric feature recognition and optimal spacing calculation. The system traces the contour skeleton layer as a vector polyline, identifying line segment intersection nodes at the end of the "joy" character, the "fast" character corner, etc. After Gaussian smoothing of these line segments, the algorithm calculates the curvature and detects curvature peaks at the sharp corners and turning points of each character, which are determined as "key inflection points" and broken into independent line segment units. At the same time, the system estimates the initial point spacing range based on the total length of the contour and the total area of the filling region, and repeatedly simulates the point distribution through the binary search algorithm to dynamically adjust the spacing until the optimal point spacing is found that makes the total number of points closest to 4000.

[0111] Based on the optimal distance and fixed anchor points, the system generates a preliminary point graph. In the contour skeleton layer, the system performs equidistant sampling along each independent line segment unit and retains all fixed anchor points. In the internal filling layer, the system generates an initial point lattice of hexagonal grid and sets all contour sampling points as constraint points, allowing only internal points to move towards the centroid of their Voronoi cells, performing a constrained spatial relaxation to ensure uniform distribution of internal points and no shrinkage of contour shape.

[0112] Finally, the system performs precise quantity fine-tuning. The system performs point density analysis on the contour line segments and filters out line segments with adjacent distances greater than the safety threshold, such as the long horizontal line of the "day" character and the vertical line of the "fast" character, as "adjustable line segments". The system calculates the difference between the total number of the current preliminary point graph 4626 and the target 4000. The system then performs resampling on the longest two adjustable line segments, respectively, by changing n equal divisions to n+1 equal divisions, each adding one point position, ensuring that the newly added points meet the minimum safety distance, and finally obtaining a final point graph with a total number of 4000.

[0113] Finally, the system maps the two-dimensional pixel coordinates of all points in the final point graph to three-dimensional geographic coordinates according to the pre-set field center coordinates, flight height, and pixel physical size, and outputs them as a CSV format control file recognizable by the UAV swarm control system. After loading this file, 4000 UAVs can be precisely commanded to fly to the predetermined point positions, perfectly presenting the "Happy Birthday" three-dimensional light sculpture in the user's creative design in the night sky.

[0114] In the second example, referring to Figures 11 to 15 , the user selects the "graph-to-graph" mode in the graphical interactive interface of the system and uploads a self-designed color cartoon lion image. The image depicts a smiling lion, slightly turning its head, sitting on the ground and shaking its tail, with a distinctive overall style and high commercial recognition. The user also adds instructions in the text input box: "Convert to a simple black line drawing suitable for a performance of 1500 UAVs, maintain the lion's contour and hair details, and fill the nose, eyes, and tail end with solid areas".

[0115] After receiving the input, the system initiates the graph generation engine. The system recognizes the presence of both a control image and editing text instructions, and therefore calls a dedicated graph generation multi-modal model, such as Qwen-Image-Edit-2509. This model has learned the mapping relationship between hundreds of thousands of "real image / painting" and "target line drawing" in the training stage, and can understand the semantics of text instructions such as "convert to simple line drawing", "keep outline", "solid fill", etc. When the model runs, it first deeply analyzes the uploaded control image, extracts the complete contour of the lion, the shape of the lion's mane, the posture of the four legs, and other key structural information through image segmentation technology. Then, the model combines the text instructions to strip the complex color and texture information and generate a new PNG format line drawing image. The line drawing retains the general outline of the lion, the shape of the ears, the smiling expression, and accurately restores the edge of the hat, the collar and sleeve of the Tang suit, and other detailed lines. In particular, for the user's requirement of "solid fill" for the collar area, the model generates it as a solid closed black area, rather than a fine line, to meet the visual performance needs of high-density point array lighting in the UAV array, perfectly converting the user's two-dimensional design into a high-fidelity input line drawing that meets the UAV performance specifications.

[0116] Subsequently, the system automatically enters the subsequent automated processing flow. The algorithm performs binaryzation and distance transformation on the "smiling lion" line drawing, separating it into a "contour skeleton layer" and an "internal fill layer". Among them, the overall contour of the lion, the facial feature lines, and the edge lines of the clothing are classified as the contour skeleton layer; while the solid areas of the ears, eyes, and collar are identified as the internal fill layer. The system vectorizes the contour skeleton layer and performs Gaussian smoothing on the lines, calculating the curvature of each vertex. At the peak positions of significant curvature such as the turning points of the lion's hair tips, the algorithm detects key inflection points and breaks them to form independent line segment units, while locking all these inflection points as fixed anchor points to ensure that these features defining the brand image are not lost in subsequent processing.

[0117] In calculating the optimal point spacing, the system uses a binary search algorithm to iteratively simulate the optimal point spacing closest to 1500 based on the target number of 1500 and the geometric dimensions of the image. Using this spacing, the system performs equidistant sampling on each independent line segment of the contour skeleton layer, retaining all fixed anchor points, and generates an initial point array of hexagonal grids in the internal fill layer. Subsequently, the system sets all contour sampling points as constraint points, performs a constrained Voronoi relaxation, optimizes the distribution of internal point arrays, and forms a preliminary point graph.

[0118] In the final fine-tuning stage, the system performs a local safety evaluation on the preliminary point graph, identifying long line segments such as the lion's long hair line and the tail tip that can be safely added or removed points. After calculation, if the preliminary point number is 1497, the system performs resampling on the three longest adjustable line segments, each adding one point, accurately matching the target of 1500. In the final generated point graph, each point is accurately located on a two-dimensional plane.

[0119] At the end of the process, the system maps the two-dimensional coordinates of the final point graph into three-dimensional space coordinates according to the parameters of the flight site, and exports all point information as a.csv control file readable by the swarm flight control system. This file can command 1500 drones to accurately form a formation in the air, reproducing the "smiling lion" image designed by the customer, with clear and visible clothing details and sharp key features, achieving seamless, efficient, and automated conversion from user personalized design to large-scale aerial art display, fully demonstrating the strong ability of the invention in supporting customized commercial performances.

[0120] As can be seen from the above embodiments and examples, the UAV performance picture generation method of the present invention has broad and far-reaching application prospects and will become the core technology engine connecting personalized creativity and large-scale aerial art presentation. Based on the efficiency, accuracy, and automation level demonstrated by the above embodiments, the method of the present invention can completely revolutionize the traditional performance design process and give birth to a new service mode and business form.

[0121] In the field of personalized celebrations and brand marketing, this method will achieve customized services with a response time of minutes. In the past, if a customer wanted a UAV performance with a theme of their own pet, a parent-child image, or a specific logo, they had to go through a design, modification, and review cycle that took several days or even weeks, which was costly and limited the scope of application. However, with the method of the present invention, customers only need to upload a photo or input a few descriptions on a mobile app, and the system can generate a high-quality flight plan that meets the requirements within five minutes. For example, a newlywed couple can immediately convert the "new couple silhouette" in their wedding photo into the "heart surrounding two people" pattern in the air; a new store owner can instantly convert their hand-drawn store sign into a shining aerial advertisement in the night sky. The picture generation model can faithfully reproduce any visual element, and the text-to-picture model can quickly materialize imaginative ideas, making each performance unique and greatly enhancing the exclusivity and communication value of the event.

[0122] Secondly, in large theme parks and cultural tourism projects, the method supports dynamic content library and theme rotation. Theme parks can use the text-to-image model in advance to generate hundreds of performance patterns according to different festivals and IP themes, and store them as templates. During festivals, operators do not need to redesign, but only need to select templates, adjust the number of stages and flight parameters, and can publish them with one key. More advanced applications can combine real-time data, such as photos of lucky visitors entering the park on the same day, to quickly generate a special name lighting night sky program, and insert it into the performance on the same night, providing visitors with an unparalleled interactive experience. This not only reduces the cost of content production, but also greatly enhances the timeliness of the performance and the emotional connection with the visitors.

[0123] Finally, in terms of emergency response and information display to the public, the method also has potential. In large gatherings or disaster sites, management departments can use the rapid generation capability of the invention to quickly deploy simple symbols such as green arrows representing safe passages or text information into drone performances, achieving long-distance, high-altitude, and large-scale information transmission with a visual impact far exceeding traditional flat advertisements. Its fully automatic generation feature ensures that it can respond quickly even in emergency situations, making up for the slow deployment and narrow coverage of traditional methods.

[0124] In summary, the drone performance picture generation method of the present invention not only solves the low efficiency problem in the prior art, but also changes the drone performance from a high threshold and long cycle luxury product to a mass service product that can be scaled, personalized, and quickly deployed. It provides a strong technical infrastructure for the innovative development of smart cities, digital tourism, interactive entertainment, and other fields.

[0125] Figure 9 is a schematic block diagram of a drone performance picture generation device 600 provided by an embodiment of the present invention. As Figure 9 shown, corresponding to the above drone performance picture generation method, the present invention also provides a drone performance picture generation device 600. The drone performance picture generation device 600 includes units for executing the above drone performance picture generation method, and the device can be configured in a desktop computer, a tablet computer, a smartphone, and the like. terminal.

[0126] Specifically, please refer to Figure 9 , the drone performance picture generation device 600 includes:

[0127] The line drawing generation unit 610 is configured to output a line drawing image conforming to the visual specification of the drone performance based on the input text description or control image through the multi-modal generation model;

[0128] The structure separation unit 620 is configured to perform structural analysis on the line drawing image and separate the contour skeleton layer and the internal filling layer;

[0129] The anchor point recognition and interval calculation unit 630 is configured to recognize and lock key inflection points as fixed anchor points based on geometric features of the contour skeleton layer, and calculate an optimal point arrangement interval applicable to the whole image;

[0130] The preliminary point arrangement generation unit 640 is configured to perform equidistant sampling on the contour skeleton layer and retain the fixed anchor points according to the point arrangement interval, generate an initial point array on the internal filling layer and perform a constrained spatial relaxation, and form a preliminary point map.

[0131] The point map fine-tuning output unit 650 is configured to perform local safety evaluation on the preliminary point map, screen adjustable line segments, and perform quantity fine-tuning on point positions based on a target performance episode, and output a final point map matched with the target performance episode.

[0132] In an embodiment, the sketch generation unit 610 includes:

[0133] A special data set construction unit is configured to construct a plurality of text-to-image data sets and a plurality of image-to-image data sets according to a plurality of episode levels of the UAV cluster, and the complexity of line drawings in the text-to-image data sets and the image-to-image data sets matches the corresponding episode levels.

[0134] A model training unit is configured to train a text-to-image model and an image-to-image model based on the text-to-image data sets and the image-to-image data sets, respectively.

[0135] A text-driven generation unit is configured to output a sketch image by calling the text-to-image model when a user input is a text description of an image.

[0136] An image editing generation unit is configured to output a sketch image by calling the image-to-image model when a user input is a control image and a semantic editing text instruction.

[0137] In an embodiment, the structure separation unit 620 includes:

[0138] An image binarization unit is configured to perform binarization processing on the sketch image to obtain pixel distribution of foreground and background.

[0139] A distance transformation unit is configured to calculate the distance of each foreground pixel to the nearest background edge by a distance transformation algorithm to form a distance map.

[0140] A layer region division unit is configured to divide the sketch image into a contour skeleton layer and an internal filling layer based on a preset distance threshold.

[0141] In an embodiment, the anchor point recognition and interval calculation unit 630 includes:

[0142] a contour vectorization unit configured to trace the bitmap of the contour skeleton layer into ordered vectorized line segments and identify line segment intersection nodes;

[0143] a curvature calculation and detection unit configured to calculate curvature distribution after Gaussian smoothing of each of the quantized line segments and detect curvature peak points;

[0144] a line segment breaking and generation unit configured to determine the curvature peak points as key inflection points and break the line segments at the key inflection points to form independent line segment units;

[0145] an anchor point locking unit configured to set all the key inflection points as fixed anchor points.

[0146] Further, the anchor point identification and distance calculation unit 630 further comprises:

[0147] a dot spacing preliminary estimation unit configured to estimate an initial dot spacing range according to the total length of the contour skeleton layer and the area of the internal filling layer;

[0148] a dot simulation calculation unit configured to simulate dots in the initial dot spacing range through binary search iteration, and calculate the sum of contour sampling points and filling area grid points in each iteration, denoted as total points;

[0149] a distance search optimization unit configured to dynamically adjust the search upper and lower bounds according to the deviation of the current total points and the target performance number of times, until the optimal dot spacing is searched.

[0150] In an embodiment, the preliminary dot generation unit 640 comprises:

[0151] a contour equidistant sampling unit configured to equidistantly sample along each line segment of the contour skeleton layer at the optimal dot spacing, and retain all the fixed anchor points;

[0152] a filling grid generation unit configured to generate a polygon grid array as initial points in the internal filling layer;

[0153] a constraint relaxation optimization unit configured to set the sampling points of the contour skeleton layer as constraint points, allow only internal filling points in the area of the internal filling layer as active points, and perform iterative movement based on the centroid of the Voronoi cell of the internal filling points to make the internal filling points perform constraint space relaxation, and obtain a preliminary dot map.

[0154] In an embodiment, the dot map fine-tuning output unit 650 comprises:

[0155] a safety evaluation and screening unit configured to perform local point density analysis on each line segment of the contour skeleton layer, and identify line segments with adjacent point spacing greater than a preset safety threshold as adjustable line segments through spatial indexing;

[0156] The point difference calculation unit is used to determine the number of points to be added or reduced based on the difference between the target number of performances and the total number of points in the preliminary point map.

[0157] The point fine-tuning resampling unit is used to resample only on the adjustable line segment based on the path length of the adjustable line segment, to add or delete points with the minimum amount, and to ensure that the added points meet the minimum safe distance constraint without destroying the image topology.

[0158] In one embodiment, the drone performance image generation device 600 further includes:

[0159] The coordinate mapping output unit is used to map the two-dimensional coordinates of all points in the final point map to three-dimensional spatial coordinates and output them as a control file that can be recognized by the cluster UAV flight control system.

[0160] The aforementioned drone performance image generation device 600 can be implemented as a computer program, which can, for example... Figure 10 It runs on the computer device shown.

[0161] Please see Figure 10 , Figure 10 This is a schematic block diagram of a computer device 500 provided in an embodiment of this application. The computer device 500 can be a terminal or a server. The terminal can be an electronic device with communication functions, such as a desktop computer, tablet computer, or smartphone. The server can be a standalone server or a server cluster composed of multiple servers.

[0162] See Figure 10 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0163] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a method for generating drone performance footage.

[0164] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0165] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for generating drone performance footage.

[0166] The network interface 505 is configured to perform network communication with other devices. Those skilled in the art can understand that Figure 10 The structure shown in FIG. 5 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device 500 to which the scheme of the present application is applied. Specifically, the computer device 500 can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0167] The processor 502 is configured to run the computer program 5032 stored in the memory, so as to implement the steps of the above method.

[0168] It should be understood that, in the embodiments of the present application, the processor 502 can be a central processing unit (CPU), and the processor 502 can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0169] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned method embodiments can be completed by a computer program instructing related hardware. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the above-mentioned method embodiments.

[0170] Therefore, the present application also provides a storage medium. The storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. The program instructions are executed by a processor to make the processor perform the steps of the above-mentioned method.

[0171] The storage medium can be a U disk, a mobile hard disk, a read-only memory (ROM), a magnetic disk or an optical disk, and various computer-readable storage media that can store program codes.

[0172] Those skilled in the art can understand that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in general terms in the above description. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0173] In several embodiments provided by the present application, it should be understood that the disclosed apparatus and method can be implemented in other ways. For example, the apparatus embodiments described above are merely schematic. For example, the division of each unit is only a logical function division, and actual implementation can have another division manner. For example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed.

[0174] The steps in the method embodiments of the present application can be adjusted, combined and deleted in sequence according to actual needs. The units in the apparatus embodiments of the present application can be combined, divided and deleted according to actual needs. In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0175] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art, or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a terminal or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application.

[0176] The above description is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for generating footage of a drone performance, characterized in that, The method comprises: Based on the input text description or control image, output the line drawing image conforming to the unmanned aerial vehicle performance visual specification through the multi-modal generation model, comprising: constructing a plurality of text-to-image data sets and a plurality of image-to-image data sets according to a plurality of mission levels of a drone cluster, the complexity of line drawing in the text-to-image data set and the image-to-image data set matches the corresponding mission level; based on the text-to-image data set and the image-to-image data set, respectively training a text-to-image model and an image-to-image model; when the user input is an image generated text description, calling and outputting a line drawing image through the text-to-image model; when the user input is a control image and a semantic editing text instruction, calling and outputting a line drawing image through the image-to-image model; Structurally analyzing the line drawing image to separate the contour skeleton layer and the internal filling layer; Based on the geometric features of the contour skeleton layer, identify and lock the key inflection points as fixed anchor points, and calculate the optimal point distribution distance suitable for the overall image, comprising: tracking the bitmap of the contour skeleton layer into an ordered vectorized line segment, and identifying the line segment intersection nodes; after Gaussian smoothing of each quantized line segment, calculate the curvature distribution and detect the curvature peak points; the curvature peak points are determined as key inflection points, and the line segments are broken at the key inflection points to form independent line segment units; all key inflection points are set as fixed anchor points; According to the point distribution distance, equidistant sampling is performed on the contour skeleton layer while retaining the fixed anchor points, and an initial point array is generated on the internal filling layer and subjected to constrained space relaxation to form a preliminary point graph, comprising: along each line segment of the contour skeleton layer, equidistant sampling is performed at the optimal point distribution distance, and all fixed anchor points are retained; a polygon grid point array is generated in the internal filling layer as an initial point position; the sampling points of the contour skeleton layer are set as constraint points, only allowing internal filling points in the internal filling layer region as active points, and based on the centroid of the Voronoi cell of the internal filling point, the internal filling point is moved iteratively to perform constrained space relaxation to obtain a preliminary point graph; Perform local safety evaluation on the preliminary point graph, screen adjustable line segments, and adjust the number of point positions based on the target performance mission to output a final point graph matching the target performance mission. 2.The method of claim 1, wherein, The step of structurally analyzing the line drawing image to separate the contour skeleton layer and the internal filling layer comprises: Performing binary processing on the line drawing image to obtain the pixel distribution of the foreground and background; Calculate the distance of each foreground pixel to the nearest background edge through the distance transform algorithm to form a distance map; Based on the preset distance threshold, the line drawing image is divided into a contour skeleton layer and an internal filling layer. 3.The method of claim 1, wherein, The step of identifying and locking the key inflection points as fixed anchor points based on the geometric features of the contour skeleton layer, and calculating the optimal point distribution distance suitable for the overall image further comprises: According to the total length of the contour skeleton layer and the area of the internal filling layer, estimate the initial point distribution distance range; In the initial distribution point spacing range, the distribution points are simulated iteratively by dichotomy search, and the sum of the number of contour sampling points and the number of filling area grid points is calculated respectively in each iteration, and is recorded as total points; According to the deviation value of the current total points and the target performance times, the upper and lower bounds of the search are dynamically adjusted until the optimal distribution point spacing is searched. 4.The method of claim 1, wherein, The step of performing local security evaluation on the preliminary point graph, screening adjustable line segments, and adjusting the number of point positions based on the target performance times to output a final point graph matching the target performance times comprises: Performing local point density analysis on each line segment of the contour skeleton layer, and identifying line segments with a distance between adjacent points greater than a preset safety threshold as adjustable line segments through spatial indexing; According to the difference between the target performance times and the total number of points in the preliminary point graph, determine the number of points to be increased or decreased; Only on the adjustable line segments, resample based on the path length of the adjustable line segments, and perform minimum point position addition and deletion, and ensure that the newly added point positions meet the minimum safety distance constraint and do not destroy the image topology.

5. An apparatus for generating a drone performance picture, the apparatus comprising: a drone performance picture generator configured to generate a drone performance picture based on a drone performance picture generation request. A computer device for executing the method for generating a performance picture of a UAV according to any one of claims 1 to 4.

6. A computer device, comprising: The computer device comprises a memory and a processor connected to the memory; the memory is used to store a computer program; and the processor is used to run the computer program stored in the memory to execute the steps of the method according to any one of claims 1 to 4.

7. A computer readable storage medium characterized in that, The storage medium stores a computer program, and the computer program comprises program instructions which, when executed by a processor, can implement the steps of the method according to any one of claims 1 to 4.

Images