Obstacle avoidance method and system based on satellite internet chain unmanned aerial vehicle
By obtaining three-dimensional point cloud data on the drone and transmitting parameters to the central server using the satellite Internet link to optimize, and combining the multi-target pigeon flock algorithm to update the obstacle avoidance model, the problem of the drone avoidance in complex environments is solved, and the problem of obstacle avoidance is realized High precision and adaptive obstacle avoidance effect.
Patent Information
- Application Number
- CN202411665545.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-20
- Publication Date
- 2025-02-18
Smart Images
Figure CN119472735A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of unmanned aerial vehicle obstacle avoidance, and in particular to an obstacle avoidance method and system for unmanned aerial vehicles based on a satellite Internet link. Background Art
[0002] At present, with the widespread application of rotary-wing UAVs in various fields, the obstacle avoidance safety issues of UAVs have received increasing attention. Existing real-time obstacle avoidance decision-making methods still have limitations when facing complex dynamic environments, and it is difficult to effectively deal with unknown obstacles. The multi-objective pigeon flock learning algorithm shows good optimization ability and adaptability by simulating the behavior of individuals in a pigeon flock learning from pigeons that are superior to themselves. However, how to organically combine this algorithm with the autonomous learning system of the UAV to achieve dynamic obstacle avoidance strategy adjustment and improve the ability to avoid unknown obstacles is still a technical problem that needs to be solved urgently. Current research mainly focuses on single-objective obstacle avoidance optimization, and it is difficult to fully take into account multiple performance indicators such as safety and efficiency. In addition, existing methods often rely on preset obstacle avoidance rules, lack adaptive learning capabilities, and are difficult to cope with complex and changing flight environments. Summary of the invention
[0003] The present application provides an obstacle avoidance method and system based on satellite Internet chain UAVs, which is used to improve the high-precision obstacle avoidance solution for the formation of satellite Internet chain UAVs.
[0004] In a first aspect, an embodiment of the present application provides an obstacle avoidance method for a UAV based on a satellite internet link, the method comprising: Acquire three-dimensional point cloud data of the surrounding environment, and train a local obstacle avoidance model according to the three-dimensional point cloud data to obtain local gradient parameters; Reading the hardware capability parameters, transmitting the local gradient parameters and the hardware capability parameters to the central server through the satellite Internet link, so that the central server aggregates the local gradient parameters and the hardware capability parameters according to the preset multi-objective pigeon flock algorithm and obtains the optimized gradient parameters; Obtaining the optimized gradient parameters, performing a timeliness check on the optimized gradient parameters, and determining target gradient parameters; The local obstacle avoidance model is updated based on the target gradient parameter to obtain an optimized local obstacle avoidance model, and the optimized local obstacle avoidance model is used to guide obstacle avoidance flight.
[0005] In a second aspect, an embodiment of the present application provides an obstacle avoidance system based on a satellite internet link drone, the system comprising: A model training module is used to obtain three-dimensional point cloud data of the surrounding environment, train a local obstacle avoidance model according to the three-dimensional point cloud data, and obtain local gradient parameters; A data transmission module is used to read the hardware capability parameters, and transmit the local gradient parameters and the hardware capability parameters to the central server through the satellite Internet link, so that the central server aggregates the local gradient parameters and the hardware capability parameters according to the preset multi-objective pigeon flock algorithm and the hardware capability parameters to obtain the optimized gradient parameters; A parameter checking module is used to obtain the optimized gradient parameters, perform a timeliness check on the optimized gradient parameters, and determine target gradient parameters; A parameter updating module is used to update the local obstacle avoidance model based on the target gradient parameter to obtain an optimized local obstacle avoidance model, and the optimized local obstacle avoidance model is used to guide obstacle avoidance flight.
[0006] In a third aspect, an embodiment of the present application provides an electronic device, the electronic device comprising a memory and a processor; The memory is used to store computer programs; The processor is used to execute the computer program and implement the obstacle avoidance method based on the satellite Internet link UAV as described in any one of the embodiments of the present application when executing the computer program.
[0007] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor implements an obstacle avoidance method for a satellite Internet link drone as described in any one of the embodiments of the present application.
[0008] An embodiment of the present application provides an obstacle avoidance method for a UAV based on a satellite Internet link, the method comprising: acquiring three-dimensional point cloud data of the surrounding environment, training a local obstacle avoidance model according to the three-dimensional point cloud data, and obtaining local gradient parameters; reading hardware capability parameters, and transmitting the local gradient parameters and the hardware capability parameters to a central server through a satellite Internet link, so that the central server aggregates the local gradient parameters and obtains optimized gradient parameters according to a preset multi-target pigeon flock algorithm and the hardware capability parameters; acquiring the optimized gradient parameters, performing a timeliness check on the optimized gradient parameters, and determining target gradient parameters; updating the local obstacle avoidance model based on the target gradient parameters to obtain an optimized local obstacle avoidance model, and the optimized local obstacle avoidance model is used to guide obstacle avoidance flight. Through the above method, three-dimensional point cloud data is used for local model training, local gradients and hardware parameters are transmitted through the satellite Internet chain, and a multi-objective pigeon flock algorithm is used on the central server to perform parameter aggregation optimization. The optimized gradient parameters are then transmitted back to the local server for timeliness check and model update. The local computing resources and environmental perception capabilities of each UAV in a distributed environment are fully utilized, and the generalization performance of the model is improved through the global optimization of the central server. At the same time, the hardware capability differences and parameter timeliness are taken into account, thereby effectively balancing the communication overhead and computing resource consumption while improving the obstacle avoidance accuracy and adaptability, providing strong support for the safe and efficient flight of UAVs in complex and changing environments. BRIEF DESCRIPTION OF THE DRAWINGS
[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.
[0010] Figure 1 A schematic flow chart of an obstacle avoidance method for a satellite internet link drone provided in an embodiment of the present application; Figure 2 A schematic block diagram of an obstacle avoidance system for a satellite internet-linked drone provided in an embodiment of the present application. DETAILED DESCRIPTION
[0011] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0012] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.
[0013] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.
[0014] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0015] Satellite Internet Chain Description: The low-orbit Starlink satellite network consists of a large number of small communication satellites with an orbital altitude of about 500-1200 kilometers. Compared with traditional communication satellites, it has the advantages of wide coverage, low transmission delay (20-40 milliseconds), and high communication rate (above 100Mbps), which can provide high-speed and stable communication services for drones worldwide. Satellite Internet Chain UAVs can communicate with the Satellite Internet Chain (hereinafter referred to as drones).
[0016] See also Figure 1 , Figure 1 The embodiment of the present application also provides a schematic flow chart of an obstacle avoidance method for a UAV based on a satellite internet link. The specific steps of the obstacle avoidance method for a UAV based on a satellite internet link include: S101-S104.
[0017] S101, acquiring three-dimensional point cloud data of the surrounding environment, and training a local obstacle avoidance model according to the three-dimensional point cloud data to obtain local gradient parameters.
[0018] For example, drones use onboard sensors such as lidar and depth cameras to scan the surrounding environment and obtain detailed three-dimensional point cloud data. The raw data usually contains a lot of information and needs to be preprocessed, such as noise reduction, filtering, and data compression, to improve the efficiency of subsequent processing. Advanced algorithms such as deep learning are used to extract features from the processed point cloud data to identify key elements and potential obstacles in the environment. The extracted features are input into the local obstacle avoidance model for training. During the training process, the model optimizes its understanding of the environment and obstacle avoidance decision-making capabilities by continuously adjusting parameters. Through algorithms such as back propagation, local gradient parameters that reflect the model's learning results are obtained. These parameters contain the model's specific understanding and adaptability to the current environment.
[0019] S102, reading the hardware capability parameters, and transmitting the local gradient parameters and the hardware capability parameters to the central server through the satellite Internet link, so that the central server aggregates the local gradient parameters and the hardware capability parameters according to the preset multi-objective pigeon flock algorithm and the hardware capability parameters to obtain the optimized gradient parameters.
[0020] Exemplarily, the drone reads its own hardware capability parameters, including key indicators such as processor performance, memory capacity, and battery life. These parameters, together with the local gradient parameters obtained in the previous step, are transmitted to the central server through the low-orbit Starlink satellite network. The Starlink satellite network consists of a large number of small low-orbit communication satellites, providing communication services with wide coverage, low latency (usually 20-40 milliseconds), and high transmission rates (up to 100Mbps or more), ensuring the rapidity and reliability of data transmission. After receiving data from multiple drones, the central server uses a preset multi-target pigeon flock algorithm to aggregate and optimize the data. This algorithm takes into account the differences in hardware capabilities of different devices, intelligently weights and comprehensively analyzes the gradient parameters, generates optimized gradient parameters, and provides each drone with a more comprehensive and optimized obstacle avoidance strategy.
[0021] S103, obtaining the optimized gradient parameters, performing a timeliness check on the optimized gradient parameters, and determining target gradient parameters.
[0022] Exemplarily, the drone receives optimized gradient parameters transmitted back from the central server through the Starlink network. These parameters are strictly checked for timeliness to ensure the timeliness and relevance of the data. The timestamps of the parameters are verified to ensure that they reflect the latest environmental information and model optimization results. These parameters are compared with the version of the current local obstacle avoidance model to ensure compatibility and applicability. Taking into account the changing environment that the drone may face, the optimized gradient parameters are further screened and adjusted according to factors such as the current flight status and environmental conditions. This process may involve local fine-tuning or selective application of parameters to ensure that the final target gradient parameters can adapt to the current actual situation to the greatest extent, providing the best basis for subsequent model updates and obstacle avoidance decisions.
[0023] S104. Update the local obstacle avoidance model based on the target gradient parameter to obtain an optimized local obstacle avoidance model, where the optimized local obstacle avoidance model is used to guide obstacle avoidance flight.
[0024] Exemplarily, the determined target gradient parameters are imported into the local obstacle avoidance model. This process usually uses methods such as incremental learning or online learning to ensure that the new parameters can be smoothly integrated into the existing model without causing sudden changes or instability in model performance. After the update is completed, the new model is subjected to a series of verifications and tests, including simulating different obstacle scenarios and evaluating the model's response speed and decision accuracy. These tests help ensure that the updated model can effectively improve obstacle avoidance performance while maintaining stability. After verification, the optimized obstacle avoidance model is officially deployed to the drone's obstacle avoidance system. In actual flight, the updated model continues to guide the drone to make real-time obstacle avoidance decisions, enabling it to respond to various complex flight environments more intelligently and efficiently, greatly improving flight safety and mission completion efficiency.
[0025] An embodiment of the present application provides an obstacle avoidance method for a UAV based on a satellite Internet link, the method comprising: acquiring three-dimensional point cloud data of the surrounding environment, training a local obstacle avoidance model according to the three-dimensional point cloud data, and obtaining local gradient parameters; reading hardware capability parameters, and transmitting the local gradient parameters and the hardware capability parameters to a central server through a satellite Internet link, so that the central server aggregates the local gradient parameters and obtains optimized gradient parameters according to a preset multi-target pigeon flock algorithm and the hardware capability parameters; acquiring the optimized gradient parameters, performing a timeliness check on the optimized gradient parameters, and determining target gradient parameters; updating the local obstacle avoidance model based on the target gradient parameters to obtain an optimized local obstacle avoidance model, and the optimized local obstacle avoidance model is used to guide obstacle avoidance flight. Through the above method, three-dimensional point cloud data is used for local model training, local gradients and hardware parameters are transmitted through the satellite Internet chain, and a multi-objective pigeon flock algorithm is used on the central server to perform parameter aggregation optimization. The optimized gradient parameters are then transmitted back to the local server for timeliness check and model update. The local computing resources and environmental perception capabilities of each UAV in a distributed environment are fully utilized, and the generalization performance of the model is improved through the global optimization of the central server. At the same time, the hardware capability differences and parameter timeliness are taken into account, thereby effectively balancing the communication overhead and computing resource consumption while improving the obstacle avoidance accuracy and adaptability, providing strong support for the safe and efficient flight of UAVs in complex and changing environments.
[0026] In order to more clearly introduce the technical solution of the present application, the technical solution of the present application will be introduced through specific embodiments below. It should be noted that the specific embodiments are used to expand the technical solution of the present application, but are not intended to limit the present application.
[0027] In some embodiments, three-dimensional point cloud data of the surrounding environment is obtained, and a local obstacle avoidance model is trained according to the three-dimensional point cloud data to obtain local gradient parameters, including: adaptive histogram preprocessing of the left and right images collected by the binocular camera to obtain a high-precision depth map; based on the depth map, a quadtree decomposition algorithm is used to adaptively segment the scene, and an improved RANSAC algorithm is applied to each quadtree node for plane fitting to obtain a layered geometric primitive representation; a mixed probability occupancy grid map is constructed by combining the geometric primitives and the original depth information to obtain a multi-resolution representation of the environment; a wavelet transform is performed on the flight state parameters of the drone to obtain a multi-scale time-frequency feature sequence; the multi-resolution environment representation and the multi-scale flight state feature sequence are input into a hierarchical attention network to obtain context-aware obstacle avoidance features; based on the obstacle avoidance features, an initial trajectory is generated using a dynamic programming algorithm, and a spline interpolation algorithm is applied to obtain a smooth obstacle avoidance path, and the smooth obstacle avoidance path is time-optimized segmented to obtain a segmented flight parameter sequence considering dynamic constraints; the segmented flight parameter sequence and the multi-resolution environment representation are input into a reinforcement learning model for training to obtain local gradient parameters.
[0028] A high-precision three-dimensional environment representation is obtained by preprocessing the binocular camera images and generating depth maps. The scene is segmented and plane-fitted using quadtree decomposition and the improved RANSAC algorithm to construct a hierarchical geometric primitive representation. This representation combines the original depth information to form a mixed probability occupancy grid map, providing a multi-resolution representation of the environment. The flight state of the drone is wavelet transformed to obtain a multi-scale time-frequency feature sequence. These environmental representations and flight state features are input into a hierarchical attention network to generate context-aware obstacle avoidance features. Based on these features, dynamic programming and spline interpolation algorithms are used to generate a smooth obstacle avoidance path, and time-optimized segmentation is performed to obtain a segmented flight parameter sequence that considers dynamic constraints. These parameter sequences are input into a reinforcement learning model together with the environmental representation, and local gradient parameters are trained. The main advantages of this method are its comprehensiveness and adaptability. By combining a variety of advanced algorithms and techniques, it is possible to perceive and understand complex three-dimensional environments more accurately, and on this basis generate safe and efficient flight paths. Multi-resolution environmental representations and multi-scale flight state features enable information to be processed at different scales, improving the adaptability to various obstacles and flight conditions. The introduction of hierarchical attention networks further enhances the contextual understanding capability and helps to make smarter obstacle avoidance decisions. Path planning that considers dynamic constraints ensures that the generated flight path is executable. Fine-tuning through reinforcement learning can continuously optimize its performance and adapt to different flight environments and mission requirements. However, this complex approach may also face challenges such as high computational complexity and strict real-time requirements, and further optimization and trade-offs may be required in practical applications.
[0029] In some embodiments, a segmented flight parameter sequence and a multi-resolution environment representation are input into a reinforcement learning model for training to obtain local gradient parameters, including: encoding the segmented flight parameter sequence in time series to obtain a flight state embedding vector; using a graph convolutional network to extract features from the multi-resolution environment representation to obtain an environment space feature vector; performing feature fusion on the flight state embedding vector and the environment space feature vector to obtain an input of a state-action value function; constructing a dual Q network structure to estimate the state-action value function to obtain a Q-value prediction; generating an obstacle avoidance action sequence based on the Q-value prediction; using a priority experience replay mechanism to sample and store historical interaction data to construct training batch data; calculating a temporal difference error based on the training batch data, and performing gradient updates on the network parameters; regularizing and gradient clipping the updated network parameters to obtain local gradient parameters.
[0030] The segmented flight parameter sequence is encoded into a time series to obtain a flight state embedding vector, and the graph convolutional network is used to extract features from the multi-resolution environment representation to obtain the environment space feature vector. The two vectors are feature fused and used as the input of the state-action value function. A dual Q network structure is constructed to estimate the state-action value function, obtain the Q value prediction, and generate the obstacle avoidance action sequence based on it. The priority experience replay mechanism is used to sample and store the historical interaction data to construct the training batch data. The time difference error is calculated according to the training data, the network parameters are gradient updated, and then regularization and gradient clipping are performed to finally obtain the local gradient parameters. This method integrates time series information and spatial information, and can better capture the relationship between flight state and environmental characteristics. The dual Q network and priority experience replay help improve learning efficiency and stability. Regularization and gradient clipping can prevent overfitting and ensure the generalization ability of the model. Overall, this method realizes intelligent planning of flight trajectories in complex environments through deep reinforcement learning, and has strong environmental adaptability and obstacle avoidance capabilities. However, practical application issues such as computational complexity and real-time performance still need to be considered, and the algorithm should be further optimized to improve performance.
[0031] In some embodiments, the central server is used to execute the aggregation of local gradient parameters and according to the preset multi-objective pigeon flock algorithm and hardware capability parameters to obtain optimized gradient parameters, specifically for executing: S1021-S1027.
[0032] S1021. Based on the preset multi-objective pigeon flock algorithm, the local gradient parameters are divided into multiple learning gradient parameter groups according to the hardware capability parameters.
[0033] Exemplarily, the multi-objective pigeon swarm algorithm plays a key role in this process. By evaluating the impact of different hardware capability parameters on gradient learning, local gradient parameters are intelligently allocated to different parameter groups. This division strategy takes into account the efficiency of hardware resource utilization, ensuring that each group can achieve optimal performance within its hardware limitations, and also ensures that the parameters learned by the drone are in line with its current hardware conditions. For example, even if the example flight action is optimal, the current hardware conditions of one of the drones cannot perform the example flight action. If the drone learns the action, there will be risks of failure and collision. This method not only optimizes resource allocation, but also improves the flexibility of overall computing, so that each learning gradient parameter group can operate effectively under specific conditions.
[0034] S1022. Divide each local gradient parameter in the learning gradient parameter group into a gradient parameter pair sequence, where the gradient parameter pair sequence includes: a path gradient parameter and a corresponding action gradient parameter.
[0035] For example, each sequence consists of path gradient parameters and corresponding action gradient parameters. This division aims to transform complex gradient information into a more tractable pair structure, making subsequent analysis and optimization more efficient. By clarifying the association between paths and actions, the algorithm can better understand the interaction between different parameter pairs, thereby providing more detailed basic data for subsequent optimization steps.
[0036] S1023. Generate a blank parameter sequence, where the blank parameter sequence includes multiple blank parameter positions.
[0037] Exemplarily, this step provides a flexible structure for subsequent parameter filling and optimization. The design of the blank parameter sequence enables flexible adjustment and optimization of its position when processing complex gradient parameter pairs, thereby achieving the best parameter distribution in the entire sequence. Through this preset blank structure, parameter filling and adjustment can be performed more effectively in subsequent steps, ensuring that each filled parameter can play its role in the appropriate position.
[0038] S1024. Determine a first filling parameter according to a first gradient parameter pair of all gradient parameter pair sequences, where the first filling parameter is used to fill a blank parameter sequence.
[0039] For example, choosing the right filling parameters is crucial for subsequent parameter optimization. By analyzing the first gradient parameter pair, we can preliminarily determine the optimal filling method to ensure that the filled parameters can coordinate with other parameters in the blank parameter sequence. This step lays the foundation for subsequent path and action scoring and provides a starting optimization direction.
[0040] S1025. Fill in the blank parameter positions in turn according to each gradient parameter pair sequence, and perform path gradient scoring and action connection scoring.
[0041] For example, through this scoring mechanism, the performance of each parameter pair in path planning and action execution can be evaluated. The path gradient score evaluates the effectiveness of the filling parameters in path planning, while the action connection score examines its performance in action continuity. This scoring method helps identify the optimal parameter combination and ensures that the final parameter configuration can provide the best performance in practical applications.
[0042] S1026: Determine a target parameter pair from the gradient parameter pair sequence according to the path scoring result and the action connection scoring result.
[0043] For example, by comprehensively evaluating the scoring results, the most advantageous parameter pairs can be identified, making them the focus of subsequent optimization. The determination of the target parameter pair is a key step in the entire optimization process, which directly affects the quality and performance of the optimized gradient parameters generated in the end. Through accurate identification and selection, it can be ensured that the final parameter optimization not only meets current needs, but also has the flexibility to adapt to future changes.
[0044] S1027. Generate optimized gradient parameters according to the target parameter pair and the blank parameter sequence.
[0045] For example, after the fine processing of the above steps, these parameters have higher effectiveness and adaptability, and can provide more accurate guidance in the application. Through this optimization process, it is ensured that the obtained gradient parameters can maintain stable performance in different environments, providing solid technical support for the obstacle avoidance and path planning of drones.
[0046] In some embodiments, path gradient scoring and action connection scoring are performed, including: performing Fourier transform on the path gradient parameters in each gradient parameter pair sequence to obtain a frequency domain feature sequence; based on the frequency domain feature sequence, using an adaptive weighted filter to smooth the path gradient parameters to obtain a smoothed path gradient sequence; applying a dynamic time warping algorithm to the smoothed path gradient sequence, calculating the similarity matrix between sequences, and obtaining a path gradient score; performing principal component analysis on the action gradient parameters in each gradient parameter pair sequence, extracting the main feature vectors, and obtaining the action features after dimensionality reduction; constructing an action transfer probability matrix based on the action features after dimensionality reduction, and using a Markov chain model to calculate the coherence of the action sequence to obtain an initial action connection score; applying a nonlinear activation function to the initial action connection score for normalization to obtain a final action connection score.
[0047] Exemplarily, the process of path gradient scoring and action connection scoring begins with Fourier transform of path gradient parameters into frequency domain feature sequences. This conversion allows for a deeper frequency analysis of path features. The frequency domain feature sequences are smoothed by adaptive weighted filters to eliminate noise and irregular fluctuations, and a smoothed path gradient sequence is obtained. The dynamic time warping algorithm is used to calculate the similarity matrix between these smoothed sequences to generate path gradient scores. For action gradient parameters, principal component analysis is used to extract the main feature vectors and achieve dimensionality reduction to simplify the complexity of the data. Based on the action features after dimensionality reduction, the action transition probability matrix is constructed, and the coherence of the action sequence is analyzed using a Markov chain model to obtain the initial action connection score. The initial score is normalized by applying a nonlinear activation function to obtain the final action connection score. The effect analysis shows that this method significantly improves the accuracy and reliability of path planning by combining frequency domain analysis and smoothing technology. The action score enhances the rationality of the action sequence through a statistical model, ensuring the smooth operation of the drone in a complex environment. The overall scoring mechanism has significant advantages in improving the accuracy and coordination of obstacle avoidance decisions.
[0048] In some embodiments, before filling the blank parameter positions in turn according to each gradient parameter pair sequence, the method further includes: determining a gradient average parameter sequence according to the gradient parameter pair sequence; and removing some values in the gradient parameter pair sequence according to the gradient average parameter sequence to obtain an optimized gradient parameter pair sequence.
[0049] Reduce the difficulty of parameter scoring by removing inefficient values.
[0050] In some embodiments, the optimized gradient parameters are obtained, the optimized gradient parameters are checked for timeliness, and the target gradient parameters are determined, including: timestamp parsing of the optimized gradient parameters to obtain the parameter generation time; calculating the time difference according to the parameter generation time and the current system time, and comparing the time difference with a preset timeliness threshold to obtain a timeliness judgment result; according to the timeliness judgment result, adaptively attenuating the optimized gradient parameters to obtain attenuated gradient parameters; performing outlier detection on the attenuated gradient parameters to eliminate abnormal gradient values that exceed a preset range to obtain filtered gradient parameters; according to the filtered gradient parameters and the current environmental state, using the Bayesian inference method to calculate the posterior probability distribution of the gradient parameters to obtain probability-weighted gradient parameters; applying a dynamic threshold clipping algorithm to the probability-weighted gradient parameters to remove low-confidence gradient components to obtain target gradient parameters.
[0051] Exemplarily, after obtaining the optimized gradient parameters, a timeliness check is performed to determine the target gradient parameters. By parsing the timestamp, the parameter generation time is identified, and the difference with the current time is calculated. This difference is compared with the preset timeliness threshold to determine the timeliness of the parameters. Based on the timeliness judgment result, adaptive attenuation processing is applied to adjust the effectiveness of the optimized gradient parameters and generate attenuated gradient parameters. Outlier detection is performed on these parameters to remove abnormal gradient values that exceed the preset range to ensure the reliability of the data and obtain the filtered gradient parameters. Combined with the current environmental state, the posterior probability distribution of these parameters is calculated using the Bayesian reasoning method to form probability-weighted gradient parameters. The low-confidence gradient components are removed through the dynamic threshold clipping algorithm, and the target gradient parameters are finally obtained. The effect analysis shows that this process ensures the timeliness and accuracy of the parameters through multi-level data verification and adjustment. Adaptive attenuation and outlier detection improve the stability of the data, while Bayesian reasoning and dynamic threshold clipping enhance the adaptability and reliability of the parameters, enabling the drone to maintain a robust decision-making ability in environmental changes.
[0052] See also Figure 2 , Figure 2 The embodiment of the present application also provides a schematic block diagram of an obstacle avoidance system based on a satellite Internet link drone, and the obstacle avoidance system 200 based on a satellite Internet link drone is used to execute the aforementioned obstacle avoidance method based on a satellite Internet link drone. The obstacle avoidance system 200 based on a satellite Internet link drone can be configured in a server.
[0053] Among them, the server can be an independent server or a server cluster, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN) and big data and artificial intelligence platforms.
[0054] like Figure 2 As shown, the obstacle avoidance system 200 based on the satellite Internet chain UAV includes: a model training module 201, a data transmission module 202, a parameter checking module 203 and a parameter updating module 204.
[0055] The model training module 201 is used to obtain three-dimensional point cloud data of the surrounding environment, and train a local obstacle avoidance model according to the three-dimensional point cloud data to obtain local gradient parameters.
[0056] The data transmission module 202 is used to read the hardware capability parameters and transmit the local gradient parameters and the hardware capability parameters to the central server through the satellite Internet link, so that the central server aggregates the local gradient parameters and the hardware capability parameters according to the preset multi-objective pigeon flock algorithm and the hardware capability parameters to obtain the optimized gradient parameters.
[0057] The parameter checking module 203 is used to obtain the optimized gradient parameters, perform a timeliness check on the optimized gradient parameters, and determine target gradient parameters.
[0058] The parameter updating module 204 is used to update the local obstacle avoidance model based on the target gradient parameter to obtain an optimized local obstacle avoidance model, and the optimized local obstacle avoidance model is used to guide obstacle avoidance flight.
[0059] An embodiment of the present application provides an electronic device, which includes a memory and a processor; the memory is used to store computer programs; the processor is used to execute the computer program and implement an obstacle avoidance method for a satellite Internet link drone as described in any one of the embodiments of the present application when executing the computer program.
[0060] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor implements an obstacle avoidance method for a satellite Internet link drone as described in any one of the embodiments of the present application.
[0061] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.
Claims
1. An obstacle avoidance method for a satellite internet-based UAV, characterized in that: The method comprises: Acquire three-dimensional point cloud data of the surrounding environment, and train a local obstacle avoidance model according to the three-dimensional point cloud data to obtain local gradient parameters; Reading the hardware capability parameters, transmitting the local gradient parameters and the hardware capability parameters to the central server through the satellite Internet link, so that the central server aggregates the local gradient parameters and the hardware capability parameters according to the preset multi-objective pigeon flock algorithm and the hardware capability parameters to obtain the optimized gradient parameters; Obtaining the optimized gradient parameters, performing a timeliness check on the optimized gradient parameters, and determining target gradient parameters; The local obstacle avoidance model is updated based on the target gradient parameter to obtain an optimized local obstacle avoidance model, and the optimized local obstacle avoidance model is used to guide obstacle avoidance flight.
2. The obstacle avoidance method based on satellite internet link UAV as claimed in claim 1, characterized in that: The step of acquiring three-dimensional point cloud data of the surrounding environment and training a local obstacle avoidance model according to the three-dimensional point cloud data to obtain a local gradient parameter includes: Performing adaptive histogram preprocessing on the left and right images captured by the binocular camera to obtain a high-precision depth map; Based on the depth map, a quadtree decomposition algorithm is used to adaptively segment the scene, and an improved RANSAC algorithm is applied to each quadtree node to perform plane fitting to obtain a hierarchical geometric primitive representation; Combining the geometric primitives with the original depth information to construct a hybrid probabilistic occupancy grid map to obtain a multi-resolution representation of the environment; Performing wavelet transform on the flight state parameters of the UAV to obtain a multi-scale time-frequency feature sequence; Inputting the multi-resolution environment representation and the multi-scale flight state feature sequence into a hierarchical attention network to obtain context-aware obstacle avoidance features; Based on the obstacle avoidance feature, an initial trajectory is generated using a dynamic programming algorithm, and a smooth obstacle avoidance path is obtained by applying a spline interpolation algorithm, and the smooth obstacle avoidance path is segmented in a time-optimized manner to obtain a segmented flight parameter sequence that takes into account dynamic constraints; The segmented flight parameter sequence and the multi-resolution environment representation are input into a reinforcement learning model for training to obtain the local gradient parameter.
3. The obstacle avoidance method based on satellite internet link UAV as claimed in claim 2 is characterized in that: The step of inputting the segmented flight parameter sequence and the multi-resolution environment representation into a reinforcement learning model for training to obtain the local gradient parameter comprises: Performing time series encoding on the segmented flight parameter sequence to obtain a flight state embedding vector; Using a graph convolutional network to extract features from the multi-resolution environment representation to obtain an environment space feature vector; Performing feature fusion on the flight state embedding vector and the environment space feature vector to obtain an input of a state-action value function; Constructing a dual Q network structure, estimating the state-action value function, and obtaining a Q value prediction; Generate an obstacle avoidance action sequence based on the Q-value prediction; Use the priority experience replay mechanism to sample and store historical interaction data and construct training batch data; Calculating the temporal difference error according to the training batch data and performing gradient update on the network parameters; Regularization and gradient clipping are performed on the updated network parameters to obtain the local gradient parameters.
4. The obstacle avoidance method for a satellite internet link drone as claimed in claim 1, characterized in that: The central server is used to aggregate the local gradient parameters and obtain optimized gradient parameters according to the preset multi-objective pigeon flock algorithm and the hardware capability parameters, and is specifically used to execute: Based on the preset multi-objective pigeon flock algorithm, the local gradient parameter is divided into a plurality of learning gradient parameter groups according to the hardware capability parameter; Dividing each local gradient parameter in the learning gradient parameter group into a gradient parameter pair sequence, wherein the gradient parameter pair sequence includes: a path gradient parameter and a corresponding action gradient parameter; generating a blank parameter sequence, wherein the blank parameter sequence includes a plurality of blank parameter positions; Determine a first filling parameter according to a first gradient parameter pair of all the gradient parameter pair sequences, wherein the first filling parameter is used to fill the blank parameter sequence; According to each of the gradient parameter pair sequences, fill the blank parameter positions in turn, and perform path gradient scoring and action connection scoring; Determining a target parameter pair from the gradient parameter pair sequence according to a result of the path scoring and a result of the action connection scoring; The optimized gradient parameters are generated according to the target parameter pair and the blank parameter sequence.
5. The obstacle avoidance method for a satellite internet link drone as claimed in claim 4, characterized in that: The path gradient scoring and action connection scoring include: Perform Fourier transform on the path gradient parameters in each gradient parameter pair sequence to obtain a frequency domain feature sequence; According to the frequency domain feature sequence, a path gradient parameter is smoothed by using an adaptive weighted filter to obtain a smoothed path gradient sequence; Applying a dynamic time warping algorithm to the smoothed path gradient sequence, calculating a similarity matrix between sequences, and obtaining a path gradient score; Perform principal component analysis on the action gradient parameters in each gradient parameter sequence, extract the main feature vector, and obtain the action features after dimensionality reduction; According to the reduced-dimensional motion features, an motion transition probability matrix is constructed, and the coherence of the motion sequence is calculated using a Markov chain model to obtain an initial motion connection score; The initial action connection score is normalized by applying a nonlinear activation function to obtain a final action connection score.
6. The obstacle avoidance method based on satellite internet link UAV as claimed in claim 4, characterized in that: Before filling the blank parameter positions in turn according to each of the gradient parameter pair sequences, the method further comprises: Determining a gradient average parameter sequence according to the gradient parameter pair sequence; According to the gradient average parameter sequence, some values in the gradient parameter pair sequence are removed to obtain an optimized gradient parameter pair sequence.
7. The obstacle avoidance method for a satellite internet link drone as claimed in claim 1, characterized in that: The obtaining of the optimized gradient parameters, performing a timeliness check on the optimized gradient parameters, and determining target gradient parameters include: Performing timestamp analysis on the optimized gradient parameters to obtain parameter generation time; Calculate the time difference according to the parameter generation time and the current system time, and compare the time difference with the preset timeliness threshold to obtain a timeliness judgment result; According to the timeliness judgment result, the optimized gradient parameter is adaptively attenuated to obtain the attenuated gradient parameter; Performing outlier detection on the attenuated gradient parameters, removing abnormal gradient values beyond a preset range, and obtaining filtered gradient parameters; According to the filtered gradient parameters and the current environmental state, the posterior probability distribution of the gradient parameters is calculated using the Bayesian inference method to obtain the probability-weighted gradient parameters; A dynamic threshold clipping algorithm is applied to the probability-weighted gradient parameters to remove low-confidence gradient components to obtain the target gradient parameters.
8. An obstacle avoidance system for UAVs based on satellite internet links, characterized in that: The obstacle avoidance system based on the satellite internet link drone includes: A model training module is used to obtain three-dimensional point cloud data of the surrounding environment, train a local obstacle avoidance model according to the three-dimensional point cloud data, and obtain local gradient parameters; A data transmission module is used to read the hardware capability parameters, and transmit the local gradient parameters and the hardware capability parameters to the central server through the satellite Internet link, so that the central server aggregates the local gradient parameters and the hardware capability parameters according to the preset multi-objective pigeon flock algorithm and the hardware capability parameters to obtain the optimized gradient parameters; A parameter checking module is used to obtain the optimized gradient parameters, perform a timeliness check on the optimized gradient parameters, and determine target gradient parameters; A parameter updating module is used to update the local obstacle avoidance model based on the target gradient parameter to obtain an optimized local obstacle avoidance model, and the optimized local obstacle avoidance model is used to guide obstacle avoidance flight.