A local perception-based distributed swarm shape formation system and implementation method thereof

Abstract

The technical scheme of the present application discloses a distributed cluster shape forming system based on local perception, characterized in that it comprises an input conversion module, an information perception module, a dynamic control module and a density adjustment module. Another aspect of the present application discloses an implementation method of the above-mentioned distributed cluster shape forming system based on local perception. The technical scheme disclosed by the present application provides adaptive formation, dynamic formation real-time transformation and autonomous obstacle avoidance functions through a completely distributed control architecture design, so as to meet the increasingly complex, dynamic and personalized requirements of unmanned aerial vehicle performances.

Description

Technical Field

[0001] This invention relates to a distributed cluster shape formation system based on local perception and its implementation method, which is particularly suitable for application scenarios such as drone swarms and robot formations that form complex two-dimensional or three-dimensional shapes, and belongs to the field of distributed intelligent agent collaborative control. Background Technology

[0002] Since Raffaello D'Andrea first demonstrated the flight of micro-drones in swarms at TED in 2016, this technology has gradually moved from laboratory demonstrations to commercialization, replacing traditional fireworks and becoming the preferred performance solution for large-scale events. With the rapid development of drone swarm performance technology, large-scale drone light shows have been widely used in celebrations, advertising, and art exhibitions. Especially driven by the upgrading of the "low-altitude economy," "nighttime economy," and "digital cultural tourism" industries, drone performances, as an emerging form of digital media art, are experiencing explosive market demand. However, as performance content becomes more complex and personalized, existing technologies still face significant challenges in achieving large-scale three-dimensional modeling, dynamic deformation, and real-time interaction.

[0003] Currently, mainstream drone performance technologies all employ a centralized architecture, relying on pre-simulated trajectory generation strategies. A typical process includes: generating 3D animation, calculating spatiotemporal constrained trajectories, online software simulation, deploying base stations and drones, drone information synchronization, and trajectory tracking control. The advantage of this architecture is its ability to pre-generate precise flight trajectories, enabling aesthetically pleasing formations even with a small number of drones.

[0004] However, centralized drone performance technology has the following inherent drawbacks:

[0005] (1) Single point of failure risk: The central base station is responsible for the information synchronization of all drones. Once the base station crashes or communication is interrupted, the entire cluster will be out of control.

[0006] (2) Animation trajectories require long-term pre-training and simulation, which makes it difficult to meet personalized and customized needs and limits flexibility;

[0007] (3) Scale bottleneck: On the one hand, the delay in drone information synchronization increases exponentially with the cluster size; on the other hand, each drone needs to store the complete pre-trained trajectory, and the onboard storage demand increases dramatically with the increase in animation complexity. Summary of the Invention

[0008] The purpose of this invention is to address the technical shortcomings of existing centralized architecture drone performances as pointed out in the background art.

[0009] To achieve the above objectives, the technical solution of the present invention discloses a distributed cluster shape formation system based on local perception, characterized in that it includes:

[0010] The input conversion module is used to convert 3D models or 2D images into point cloud data;

[0011] The information perception module adopts a distributed information processing mechanism to dynamically adjust the perception range based on the distance between the current UAV and the nearest target point, and at the same time, it is used to collect local environmental information in real time.

[0012] The dynamic control module is used to autonomously calculate the optimal motion trajectory of each UAV based on real-time acquired point cloud information and environmental data, and to control each UAV to achieve dynamic obstacle avoidance while maintaining the target formation. Specifically, the dynamic control module updates the UAV motion based on the composite potential field, and the formula for calculating the total resultant force of the composite potential field is:

[0013] F total =F coh +F sep +F align +F obs +F surf +F mean

[0014] In the formula: F coh F sep F align These are the cluster aggregation vector, the outlier vector, and the velocity alignment vector, respectively.

[0015] F obs For obstacle avoidance vectors, Among them, W obs It is the obstacle avoidance vector weight scalar, Obs i P is the set of obstacles observed by drone i. i and P j These are the three-dimensional coordinate position vectors of UAV i and UAV j, respectively;

[0016] F surf It is the surface tension vector. Among them, W surf It is the surface tension vector weight scalar, S i P is the set of nearest point clouds detected by the drone. k The three-dimensional coordinates of point cloud k, d i,k =‖P k -P i Let ‖ be the distance scalar from UAV i to point cloud k, and d ref It is the distance set by surface tension;

[0017] F mean For a density uniform vector, Among them, W mean N is a scalar of density equilibrium weighting coefficients. i Let K be the set of neighbors of UAV t, K(·) be the Epanechnikov kernel function, and h be the bandwidth parameter scalar;

[0018] The density adjustment module is used to achieve autonomous adjustment of cluster density through the Mean-Shift algorithm, enabling each UAV to intelligently move to areas with lower point cloud density based on the distribution of neighboring UAVs, thus ensuring formation uniformity.

[0019] Preferably, the input conversion module outputs a specified number of 3D point clouds through an intelligent sampling algorithm, while retaining the original input RGB color information.

[0020] Preferably, the local environmental information includes spatial distribution data of obstacles and the location information of the K nearest neighboring machines.

[0021] Preferably, the aggregation vector F coh Calculate using the following formula:

[0022] F coh =F coh +W coh ·(d i,j -r ref )·(P j -P i )

[0023] In the formula, d i,j r is the distance scalar between drone i and drone j. ref W is a scalar parameter representing the distance between drones. coh It is a clustered vector weight scalar.

[0024] Preferably, the outlier vector F sep Calculate using the following formula:

[0025] F sep =W sep ·(d i,j -r ref )·(P j -P i )

[0026] In the formula, W sep It is the outlier vector weight scalar, d i,j r is the distance scalar between drone i and drone j. ref It is a scalar parameter for the distance between drones.

[0027] Preferably, the outlier vector velocity alignment vector F align Calculate using the following formula:

[0028]

[0029] In the formula, W align It is the velocity vector weight scalar, N i V is the set of neighbors of drone i. j That is the velocity vector of drone j.

[0030] Another aspect of the present invention discloses a method for implementing the above-described distributed cluster shape forming system based on local awareness, characterized by comprising the following steps:

[0031] Step 1: Input and transform the target shape;

[0032] Input 3D model data or 2D images, and convert the input into a discrete point cloud set with RGB information through a point cloud sampling algorithm;

[0033] Step 2, Local Information Perception:

[0034] Each drone autonomously completes the following perception tasks:

[0035] 1) Surface point cloud detection: A dynamic sensing radius mechanism is adopted, and the radius is inversely proportional to the distance from the UAV to the nearest point cloud point;

[0036] 2) Neighbor location: Fast retrieval of nearest neighbor drones based on KD-Tree spatial index;

[0037] 3) Obstacle detection: Real-time scanning of the spatial coordinates of surrounding obstacles;

[0038] Step 3, Dynamic Behavior Control:

[0039] Comprehensive calculation of cohesion F coh Separation force F sep Alignment force F align Surface tension F surf Obstacle repulsive force F obs ;

[0040] Step 4: Uniform cluster density:

[0041] To further homogenize the spatial distribution of the cluster, the mean-shift algorithm is used to calculate the uniform force F. mean ;

[0042] Step 5, Coverage Status Assessment:

[0043] Based on the drone's entry into the shape range rate and point cloud coverage, a joint assessment is made as to whether the target shape has been formed;

[0044] Step 6: Dynamically switch shapes:

[0045] Continuous deformation can be achieved by sequentially switching pre-stored point cloud sequences, resulting in animation effects. Furthermore, it supports real-time uploading of new models or images, allowing the cluster to autonomously transition to the new target shape.

[0046] The technical solution disclosed in this invention provides adaptive formation, real-time dynamic formation changes, and autonomous obstacle avoidance functions through a fully distributed control architecture design, in order to meet the increasingly complex, dynamic, and personalized needs of drone performances, and can simultaneously meet the following key requirements:

[0047] (1) Distributed architecture: No need to rely on a central base station, avoiding the risk of single point of failure;

[0048] (2) Dynamic formation control: No pre-trained trajectories are required. It supports the autonomous formation and maintenance of specific shapes by the cluster and can dynamically switch formations.

[0049] (3) Large-scale cluster support: It has the ability to support ultra-large-scale clusters.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] (1) Fully distributed architecture design: The fully peer-to-peer distributed architecture is adopted, and each UAV only needs to share local location information. Cluster collaboration is achieved through unified point cloud data replicas and consistency algorithms, which completely eliminates the risk of single point of failure inherent in centralized architecture.

[0052] (2) Dynamic real-time reconstruction capability: Breaking through the limitations of the traditional pre-calculated trajectory mode, it supports real-time uploading and dynamic switching of target shapes, and can adapt to sudden changes in performance requirements.

[0053] (3) Linear scaling characteristic: Based on the current distributed computing architecture, the single-machine computing complexity is linearly related to the cluster size (O(n)). Simulation verification shows that it still maintains stable operation at a scale of tens of thousands of machines, breaking through the scale bottleneck of traditional centralized scheduling algorithms.

[0054] (4) End-to-end automated interaction: Provides a one-stop solution for "upload and perform": Users only need to submit models or images, and the system automatically completes point cloud conversion and subsequent shape composition and optimization. Personalized customization needs can be met through real-time shape modification function. Attached Figure Description

[0055] Figure 1 This is a schematic flowchart illustrating the implementation of the present invention.

[0056] Figure 2 This is an example of the first specific embodiment of the present invention.

[0057] Figure 3 This is a second specific embodiment example of the present invention.

[0058] Figure 4 This is a third specific embodiment example of the present invention.

[0059] Figure 5 This is an example of the fourth specific embodiment of the present invention. Detailed Implementation

[0060] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0061] One aspect of this invention discloses a distributed cluster shape forming system based on local perception. This system is a novel distributed cluster shape composition system and mainly includes several core modules described below.

[0062] 1) Input Conversion Module

[0063] It is responsible for converting 3D models (supporting formats such as .obj and .fbx) or 2D images (supporting formats such as .png and .jpg) into point cloud data. The input conversion module outputs a specified number of 3D point clouds through an intelligent sampling algorithm, while retaining the RGB color information of the original input.

[0064] (ii) Information Sensing Module

[0065] Employing a distributed information processing mechanism, each UAV stores the same global point cloud data, but through the information perception module, it can dynamically adjust its perception range based on its distance from the nearest target point (the closer the distance, the smaller the perception range). Simultaneously, the information perception module is also used to collect local environmental information in real time, mainly including obstacle spatial distribution data and the position information of the K nearest neighboring UAVs.

[0066] (iii) Dynamic Control Module

[0067] Based on real-time point cloud information and environmental data, each UAV uses a dynamic control module to autonomously calculate the optimal trajectory, achieving dynamic obstacle avoidance while maintaining the target formation.

[0068] In this embodiment of the invention, the dynamic control module updates the UAV motion based on the composite potential field. The formula for calculating the total resultant force of the composite potential field is:

[0069] F total =F coh +F sep +F align +F obs +Fsurf +F mean

[0070] In the formula: F coh F sep F align These are the cluster aggregation vector, outlier vector, and velocity alignment vector, respectively, which are part of the original basic Boids model.

[0071] When a single drone moves away from the cluster, it generates a clustering vector F. coh The direction points towards the cluster centroid, that is, if d i,j <r ref F coh =F coh +W coh ·(d i,j -r ref )·(P j -P i ), where d i,j r is the distance scalar between drone i and drone j. ref W is a scalar parameter representing the distance between drones. coh It is a clustered vector weight scalar, P i and P j These are the three-dimensional coordinate position vectors of UAV i and UAV j, respectively;

[0072] Conversely, when drones get too close to the swarm, they generate an outlier vector F. sep The direction points in the opposite direction to the cluster centroid, that is, F sep =W sep ·(d i,j -r ref )·(P j -P i ), where W sep It is an outlier vector weight scalar;

[0073] The velocity alignment vector, on the other hand, emphasizes the velocity alignment between the drone and the swarm, i.e. Among them, W align It is the velocity vector weight scalar, N i V is the set of neighbors of drone i. j That is the velocity vector of drone j;

[0074] Based on this, the present invention adds F obs F surf F mean It is used for dynamic obstacle avoidance, surface tension behavior, and uniform cluster density.

[0075] F obsThis is the obstacle avoidance vector, pointing in the opposite direction to the obstacle. The specific calculation formula is as follows: Among them, W obs It is the obstacle avoidance vector weight scalar, Obs i It is the set of obstacles observed by UAV i;

[0076] F surf The surface tension vector is used to simulate the "thin film effect" of drone swarms in space, enabling the swarm to "cover" the surface of the target graphic. The specific calculation formula is as follows: Among them, W surf It is the surface tension vector weight scalar, S i P is the set of nearest point clouds detected by the drone. k The three-dimensional coordinates of point cloud k, d i,k =‖P k -P i ‖ represents the distance scalar from UAV i to point cloud k, and d represents the distance d. ref The surface tension setting distance is used because the point cloud detection range decreases as the distance between the drone and the point cloud decreases, thus achieving a preliminary effect of uniform clustering.

[0077] F mean As a density uniform vector, the density gradient force of the particle system is simulated to drive the UAV to move towards a locally sparse region, thereby further homogenizing the spatial distribution of the cluster. Essentially, this is an application of the MeanShift algorithm based on kernel density estimation (KDE) in three-dimensional space. The specific calculation formula is as follows: Among them, W mean It is a scalar of density equilibrium weighting coefficient, N i K is the set of neighbors of UAV i (i.e., the distance between them is less than h), K(·) is the Epanechnikov kernel function, and h is the bandwidth parameter scalar.

[0078] IV) Density Adjustment Module

[0079] The Mean-Shift algorithm enables autonomous adjustment of cluster density, allowing each drone to intelligently move towards areas with lower point cloud density based on the distribution of neighboring drones, thus ensuring uniformity of formation.

[0080] Another aspect of this invention discloses a method for implementing the above-mentioned distributed cluster shape forming system based on local awareness, specifically including the following steps:

[0081] Step 1: Input and transform the target shape;

[0082] Input 3D model data or 2D images, and convert the input into a discrete point cloud set with RGB information through a point cloud sampling algorithm. In this embodiment of the invention, the supported formats for the model are obj and fbx, and the supported formats for the images are png and jpg.

[0083] Step 2, Local Information Perception:

[0084] Each drone autonomously completes the following perception tasks:

[0085] 1) Surface point cloud detection: A dynamic sensing radius mechanism is adopted, and the radius is inversely proportional to the distance from the UAV to the nearest point cloud point;

[0086] 2) Neighbor location: Fast retrieval of nearest neighbor drones based on KD-Tree spatial index;

[0087] 3) Obstacle detection: Real-time scanning of the spatial coordinates of surrounding obstacles;

[0088] Step 3, Dynamic Behavior Control:

[0089] Comprehensive calculation of multiple forces: basic Boids force (cohesive force F) coh / Separation force F sep Alignment Force F align ), surface tension F surf Obstacle repulsive force F obs .

[0090] Step 4: Uniform cluster density:

[0091] To further homogenize the spatial distribution of the cluster, the mean-shift algorithm is used to calculate the uniform force F. mean ;

[0092] Step 5, Coverage Status Assessment:

[0093] Based on the drone's entry into the shape range rate and point cloud coverage, a joint assessment is made as to whether the target shape has been formed;

[0094] Step 6: Dynamically switch shapes:

[0095] It can sequentially switch between pre-stored point cloud sequences to achieve continuous deformation and realize animation effects. In addition, it also supports real-time uploading of new models or images, and the cluster can autonomously transition to the new target shape.

[0096] Figure 2 and Figure 3 The invention demonstrates its practical application in two-dimensional graphic formation and verifies the algorithm's excellent adaptability to 2D clusters. The left side shows the target graphic input to the algorithm. Figure 2 The simulation results of 100 drones were presented. Figure 3This demonstrates the swarm performance of 500 drones. Experiments prove that this invention is not only applicable to drone swarm control, but also effectively applied to the control of two-dimensional swarm systems such as unmanned vehicles.

[0097] Figure 4 The simulation further demonstrates the algorithm's superior performance in 3D space through examples of 3D morphology formation. The simulation experiment used 500 drones; the left side shows the input target 3D model, and the right side shows the final 3D structure.

[0098] Figure 5 The adaptability of the invention was verified using a larger-scale cluster. Through a simulation experiment with 2000 UAVs, after inputting the target 3D model on the left, the expected three-dimensional configuration was successfully achieved on the right.

Claims

1. A distributed cluster shape formation system based on local perception, characterized in that, include: The input conversion module is used to convert 3D models or 2D images into point cloud data; The information perception module adopts a distributed information processing mechanism to dynamically adjust the perception range based on the distance between the current UAV and the nearest target point, and at the same time, it is used to collect local environmental information in real time. The dynamic control module is used to autonomously calculate the optimal motion trajectory of each UAV based on real-time acquired point cloud information and environmental data, and to control each UAV to achieve dynamic obstacle avoidance while maintaining the target formation. Specifically, the dynamic control module updates the UAV motion based on the composite potential field, and the formula for calculating the total resultant force of the composite potential field is: F total =F coh +F sep +F align +F obs +F surf +F mean In the formula: F coh F sep F align These are the cluster aggregation vector, the outlier vector, and the velocity alignment vector, respectively. F obs For obstacle avoidance vectors, Among them, W obs It is the obstacle avoidance vector weight scalar, Obs i P is the set of obstacles observed by drone i. i and P j These are the three-dimensional coordinate position vectors of UAV i and UAV j, respectively; F surf It is the surface tension vector. Among them, W surf It is the surface tension vector weight scalar, S i P is the set of nearest point clouds detected by the drone. k The three-dimensional coordinates of point cloud k, d i,k =‖P k -P i Let ‖ be the distance scalar from UAV i to point cloud k, and d ref It is the distance set by surface tension; F mean For a density uniform vector, Among them, W mean N is a scalar of density equilibrium weighting coefficients. i Let K be the set of neighbors of UAV t, K(·) be the Epanechnikov kernel function, and h be the bandwidth parameter scalar; The density adjustment module is used to achieve autonomous adjustment of cluster density through the Mean-Shift algorithm, enabling each UAV to intelligently move to areas with lower point cloud density based on the distribution of neighboring UAVs, thus ensuring formation uniformity.

2. The distributed cluster shape forming system based on local perception as described in claim 1, characterized in that, The input conversion module outputs a specified number of 3D point clouds through an intelligent sampling algorithm, while retaining the original input RGB color information.

3. The distributed cluster shape forming system based on local perception as described in claim 1, characterized in that, The local environmental information includes spatial distribution data of obstacles and the location information of the K nearest neighboring machines.

4. The distributed cluster shape forming system based on local perception as described in claim 1, characterized in that, The aggregation vector F coh Calculate using the following formula: F coh =F coh +W coh ·(d i,j -r ref )·(P j -P i ) In the formula, d i,j r is the distance scalar between drone i and drone j. ref W is a scalar parameter representing the distance between drones. coh It is a clustered vector weight scalar.

5. The distributed cluster shape forming system based on local perception as described in claim 1, characterized in that, The outlier vector F sep Calculate using the following formula: F sep =W sep ·(d i,j -r ref )·(P j -P i ) In the formula, W sep It is the outlier vector weight scalar, d i,j r is the distance scalar between drone i and drone j. ref It is a scalar parameter for the distance between drones.

6. The distributed cluster shape forming system based on local perception as described in claim 1, characterized in that, The outlier vector velocity alignment vector F align Calculate using the following formula: In the formula, W align It is the velocity vector weight scalar, N i V is the set of neighbors of drone i. j That is the velocity vector of drone j.

7. A method for implementing a distributed cluster shape forming system based on local perception as described in claim 1, characterized in that, Includes the following steps: Step 1: Input and transform the target shape; Input 3D model data or 2D images, and convert the input into a discrete point cloud set with RGB information through a point cloud sampling algorithm; Step 2, Local Information Perception: Each drone autonomously completes the following perception tasks: 1) Surface point cloud detection: A dynamic sensing radius mechanism is adopted, and the radius is inversely proportional to the distance from the UAV to the nearest point cloud point; 2) Neighbor location: Fast retrieval of nearest neighbor drones based on KD-Tree spatial index; 3) Obstacle detection: Real-time scanning of the spatial coordinates of surrounding obstacles; Step 3, Dynamic Behavior Control: Comprehensive calculation of cohesion F coh Separation force F sep Alignment force F align Surface tension F surf Obstacle repulsive force F obs ; Step 4: Uniform cluster density: To further homogenize the spatial distribution of the cluster, the mean-shift algorithm is used to calculate the uniform force F. mean ; Step 5, Coverage Status Assessment: Based on the drone's entry into the shape range rate and point cloud coverage, a joint assessment is made as to whether the target shape has been formed; Step 6: Dynamically switch shapes: Continuous deformation can be achieved by sequentially switching pre-stored point cloud sequences, resulting in animation effects. Furthermore, it supports real-time uploading of new models or images, allowing the cluster to autonomously transition to the new target shape.

Images