Real-time path optimization system for drones based on edge computing

Abstract

This invention discloses a real-time path optimization system for unmanned aerial vehicles (UAVs) based on edge computing, relating to the field of UAV path planning. The system includes: a perception data acquisition module deployed on various UAV platforms, which synchronously acquires multi-source sensor data and generates multi-source perception data packets; a global map fusion module, which receives data packets and generates a globally consistent probabilistic map with time-varying characteristics; a collaborative task planning module, which allocates tasks based on the probabilistic map using a market auction mechanism, generating a global path sequence that integrates path length, safety, and energy consumption; a local obstacle avoidance execution module, which receives planning schemes and, combined with real-time sensor data, generates control commands using model predictive control and replanning algorithms to achieve trajectory tracking and dynamic obstacle avoidance; and a closed-loop feedback update module, which analyzes execution logs, generates optimization commands through deviation evaluation and incremental learning, and feeds them back to the fusion and planning module.

Description

Technical Field

[0001] This invention relates to the technical field of unmanned aerial vehicle (UAV) path planning, and more specifically, to a real-time UAV path optimization system based on edge computing. Background Technology

[0002] Path optimization technology refers to computational methods that use algorithms to find the optimal or suboptimal path from the starting point to the destination for a moving vehicle. This technology faces severe challenges in fire rescue scenarios, as indoor and outdoor fire environments place extremely high demands on the collaborative operation capabilities of drone swarms. Due to the presence of numerous unknown obstacles, dense smoke interference, and complete lack of GPS signals inside fire scenes, drones must rely entirely on onboard sensors for environmental perception and autonomous positioning, and ensure the safety and mission efficiency of multi-drone collaboration through real-time path planning.

[0003] To meet the low-latency, high-reliability computing requirements of multi-drone collaboration in fire environments, existing technologies typically employ edge computing architectures, offloading computationally intensive tasks from the cloud to mobile edge servers deployed on fire trucks. These servers then fuse and process the perception data from multiple drones to generate flight paths.

[0004] However, edge computing architecture mainly improves data processing efficiency through distributed computing. It lacks a unified fusion mechanism for multi-source heterogeneous sensing data, resulting in poor consistency of the generated global environment model and failing to provide reliable conflict-free collaborative path planning for drone swarms. Summary of the Invention

[0005] The purpose of this invention is to achieve real-time path optimization for multiple UAVs in complex dynamic environments by unifying and fusing multi-source perception data, generating probabilistic maps, and planning distributed tasks under an edge computing architecture. This solves the problem of collaborative planning conflicts caused by inconsistent perception and improves the safety, operational efficiency, and robustness of UAV swarms.

[0006] The technical solution of the present invention is: to provide a real-time path optimization system for unmanned aerial vehicles based on edge computing, the system comprising: a perception data acquisition module, a global map fusion module, a collaborative task planning module, a local obstacle avoidance execution module, and a closed-loop feedback update module;

[0007] The perception data acquisition module is deployed on each UAV platform to synchronously acquire multi-source sensor data and generate multi-source perception data packages containing environmental point cloud datasets, environmental depth information matrices, and flight attitude quaternion sequences.

[0008] The global map fusion module receives multi-source sensing data packets and generates a globally consistent probabilistic map with time-varying characteristics through spatiotemporal alignment processing, voxel gridding processing, feature-level fusion, Bayesian update-based probabilistic fusion algorithm, and dynamic decay processing.

[0009] The collaborative task planning module is used to receive the global consistency probability map, and based on the global consistency probability map, it uses a market auction mechanism to allocate multiple UAV tasks, perform global path planning, and generate a multi-UAV collaborative task planning scheme that includes waypoint sequence, expected speed, task time window and emergency backup path.

[0010] The local obstacle avoidance execution module is used to receive multi-UAV collaborative mission planning schemes and combine them with real-time sensor data to generate real-time UAV control commands using model predictive control algorithms and replanning algorithms, thereby achieving local trajectory optimization and dynamic obstacle avoidance.

[0011] The closed-loop feedback update module receives the execution log data packets from the local obstacle avoidance execution module, and generates system optimization instructions through deviation analysis and incremental learning, which are then fed back to the global map fusion module and the collaborative task planning module.

[0012] In any of the above technical solutions, the sensing data acquisition module further adopts a master-slave acquisition architecture, using a high-precision time synchronization module as the master clock, and synchronizing the acquisition timing of the lidar, stereo vision camera and inertial measurement unit through hardware trigger signals to ensure that the multi-source data has a unified microsecond-level time reference.

[0013] The perception data acquisition module performs motion distortion compensation, coordinate system normalization, and voxel grid filtering on the lidar point cloud data to generate an environmental point cloud dataset.

[0014] The perception data acquisition module performs distortion correction, semi-global matching, and triangulation on the stereo vision images to generate an environmental depth information matrix.

[0015] The sensing data acquisition module performs zero bias compensation, temperature drift correction, scale factor calibration, and complementary filtering on the inertial measurement unit data to generate a flight attitude quaternion sequence.

[0016] In any of the above technical solutions, the perception data acquisition module further encapsulates the processed data into a multi-source perception data packet. The data packet includes the acquisition timestamp, UAV number, and sensor status flag, and is transmitted to the global map fusion module in real time via a gigabit Ethernet interface.

[0017] In any of the above technical solutions, the global map fusion module further employs an exponential decay model to perform confidence decay processing on grid cells that have not been updated for a long time, and the decay coefficient is adaptively adjusted according to the degree of dynamic change in the environment.

[0018] In any of the above technical solutions, the spatiotemporal alignment processing of the global map fusion module is further based on the flight attitude quaternion sequence, which transforms the environmental point cloud dataset and the environmental depth information matrix into the global coordinate system;

[0019] The voxel meshing process uses a 0.1-meter resolution voxel mesh to perform density statistics and feature extraction on the spatiotemporally aligned environmental point cloud dataset, generating a standardized point cloud distribution map.

[0020] Feature-level fusion converts the environmental depth information matrix into a point cloud format and supplements it with texture information;

[0021] The probabilistic fusion algorithm based on Bayesian updates uses the following recursive formula for adaptive fusion of multi-source data:

[0022] ;

[0023] In the formula, This represents the updated posterior occupancy probability; This represents the prior occupancy probability of a grid cell at the current moment, derived from the fusion results of historical observation data; This represents the occupancy probability calculated from the current sensor observations;

[0024] Dynamic decay processing uses an exponential decay model to reduce confidence:

[0025] ;

[0026] In the formula, This represents the new occupancy probability after attenuation processing; This represents the prior occupancy probability of a grid cell at the current time (ranging from 0 to 1). Indicator decay factor, It is the natural constant (approximately 2.71828); It is the decay rate constant ( >0), controls the rate of decay; It refers to the time interval or the elapsed time.

[0027] In any of the above technical solutions, the collaborative task planning module further considers path length, safety and energy consumption indicators when planning the path, and generates an emergency backup path for each path;

[0028] The multi-UAV collaborative mission planning scheme adopts a hierarchical storage structure, including mission allocation results, path sequences and waypoint parameters, and is smoothed by Bezier curves.

[0029] In any of the above technical solutions, the collaborative task planning module further transmits the multi-UAV collaborative task planning scheme to the local obstacle avoidance execution module and synchronizes the time scheduling information to the closed-loop feedback update module.

[0030] In any of the above technical solutions, the closed-loop feedback update module further receives the execution log data packet from the local obstacle avoidance execution module and generates system optimization instructions to be fed back to the global map fusion module.

[0031] The beneficial effects of this invention are:

[0032] This invention achieves a fundamental improvement in environmental cognition, moving from static snapshots to dynamic evolution. While traditional edge computing solutions can achieve multi-source data fusion, the resulting environmental models are essentially static and cannot effectively represent the dynamic evolution of the fire environment over time. This invention introduces a Bayesian update-based probabilistic fusion and spatiotemporal decay mechanism, making the global map a dynamic cognitive model with "memory" and "forgetting" capabilities. It not only reflects the instantaneous state of the environment but also intuitively represents the uncertainty and timeliness of environmental information through the time decay characteristics of occupancy probabilities. This allows path planning algorithms to weigh risks based on the reliability of information, thereby making safer and more forward-looking decisions in dynamically changing environments.

[0033] This invention constructs a collaborative paradigm shift from centralized control to distributed self-organization. Existing technologies largely rely on a central node for task allocation, resulting in single-point bottlenecks and insufficient flexibility. This invention transforms task allocation into a distributed decision-making process through a market auction mechanism. Each UAV autonomously bids based on its local state (location, battery level), and a globally near-optimal task allocation scheme naturally emerges through the simple and efficient medium of price signals. This paradigm not only reduces the system's communication and computational overhead but also endows the UAV swarm with strong self-organization capabilities and robustness. Even if some nodes fail or communication is interrupted, the remaining units can still continue to complete collaborative tasks through local interaction, greatly improving the system's survivability and operational capabilities in extreme and harsh environments. Attached Figure Description

[0034] The advantages of the above and additional aspects of the present invention will become apparent and readily understood in the description of the embodiments in conjunction with the following drawings, wherein:

[0035] Figure 1 This is a schematic diagram of the overall process structure of a real-time path optimization system for unmanned aerial vehicles based on edge computing according to an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the perception data acquisition module of a real-time path optimization system for unmanned aerial vehicles based on edge computing according to an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of the global map fusion module of a real-time path optimization system for unmanned aerial vehicles based on edge computing, according to an embodiment of the present invention.

[0038] Among them, 100-perception data acquisition module, 200-global map fusion module, 300-collaborative task planning module, 400-local obstacle avoidance execution module, and 500-closed-loop feedback update module. Detailed Implementation

[0039] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0040] In the following description, many specific details are set forth in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0041] This embodiment addresses the problem of collaborative planning conflicts caused by inconsistent perception data in complex environments during fire emergency rescue of drone swarms. It provides a drone real-time path optimization system based on edge computing. By establishing a unified fusion mechanism for multi-source heterogeneous perception data, a globally consistent environmental map with confidence assessment is generated. Based on this map, collaborative task allocation and real-time path planning for multiple drones are realized. At the same time, a closed-loop feedback mechanism is used to continuously optimize environmental perception and path schemes. Ultimately, the system enables drone swarms to operate safely, efficiently, and without conflict in complex fire environments with GPS rejection and unknown conditions.

[0042] like Figure 1 As shown, the system includes: a perception data acquisition module 100, a global map fusion module 200, a collaborative task planning module 300, a local obstacle avoidance execution module 400, and a closed-loop feedback update module 500.

[0043] The perception data acquisition module 100 is deployed on each UAV. Its main function is to synchronously acquire multi-source sensor data, including LiDAR point cloud, stereo vision depth information and IMU attitude quaternion, and generate standardized data packets after preprocessing. This module ensures microsecond-level time reference, supports motion compensation and filtering, and finally uploads the data packets to the global map fusion module 200 and receives map broadcasts to calibrate local perception.

[0044] The global map fusion module 200 serves as the core unit. Its main functions are to receive multi-source data packets, perform spatiotemporal alignment, voxel gridding, Bayesian probabilistic fusion, and exponential decay to generate a time-varying globally consistent probability map. This map includes occupancy probability, update timestamp, and dynamic identifiers. The map is then output to the collaborative task planning module 300 to broadcast updates to the UAV and receive closed-loop feedback to optimize fusion parameters, thereby achieving unified environmental perception.

[0045] The collaborative task planning module 300, as the decision-making unit, mainly performs the following tasks: decompose tasks based on probabilistic maps, allocate them through market auctions, generate global paths, consider length, safety, and energy consumption, reserve backup paths and time-sharing schedules, transmit the plan to the local obstacle avoidance execution module 400, synchronously schedule it to the closed-loop feedback module, and receive map update and optimization instructions to form a planning closed loop.

[0046] The local obstacle avoidance execution module 400 serves as the execution unit. Its main tasks include parsing the planning scheme, tracking the trajectory using the MPC algorithm, realizing dynamic obstacle avoidance, monitoring the status, and recording logs. This module drives the flight control execution, uploads logs to the closed-loop feedback update module 500, and receives planning updates and map broadcasts to ensure path cascading and conflict coordination.

[0047] The closed-loop feedback update module 500, acting as an optimization unit, primarily analyzes execution logs and change information, uses DTW to assess deviations, adjusts parameters through incremental learning, maintains the knowledge base and health monitoring, and triggers reassignment. In terms of collaboration, this module feeds back instructions to the fusion and planning modules, monitors execution status, and enables adaptive iteration across the entire system.

[0048] Specifically, such as Figure 2 As shown, the perception data acquisition module 100, as the front-end data acquisition unit of the system, is deployed on each UAV platform and is responsible for high-precision synchronous acquisition and preprocessing of multi-source heterogeneous sensor data. The module adopts a master-slave acquisition architecture, using a high-precision Beidou timing module as the master clock, and synchronizes the acquisition timing of each sensor through hardware trigger signals to ensure that all perception data have a unified microsecond-level time reference. The module has a built-in data acquisition controller, which connects to sensors such as LiDAR, stereo vision camera and inertial measurement unit through multiple high-speed data interfaces, and uses direct memory access (DMA) to achieve uninterrupted acquisition of high-speed data streams.

[0049] In terms of lidar data acquisition, this module controls a 32-line lidar to scan the environment at a frequency of 10Hz. Each frame acquires and generates corresponding point cloud data. The acquired raw lidar point cloud data is first compensated for motion distortion by using a linear interpolation algorithm to eliminate point cloud distortion caused by the movement of the UAV platform. Then, coordinate system unification processing is performed to transform the point cloud data from the radar coordinate system to the UAV body coordinate system. Finally, the data is downsampled using a voxel grid filtering algorithm to generate a structured environmental point cloud dataset.

[0050] In terms of visual data acquisition, the module controls the stereo vision camera to acquire image pairs with a resolution of 1280×1024 at a frequency of 20Hz. After the left and right images are exposed synchronously by hardware, the camera calibration parameters are loaded and distortion correction is performed first to eliminate the influence of lens distortion. Then, a semi-global matching algorithm is used to calculate the disparity map, and a floating-point depth map is generated through the principle of triangulation. Finally, the depth map is processed by median filtering and scale normalization to generate a high-precision environmental depth information matrix.

[0051] In terms of attitude data acquisition, the module acquires raw data from the inertial measurement unit at a frequency of 200Hz; performs zero-bias compensation and temperature drift correction on the gyroscope data, performs scale factor calibration and orthogonal error compensation on the accelerometer data, and uses a complementary filtering algorithm to fuse the gyroscope and accelerometer data to calculate the pitch, roll and yaw angles of the UAV in real time; finally, it generates a standardized flight attitude quaternion sequence through quaternion transformation.

[0052] The perception data acquisition module 100 collects and processes all the aforementioned data, aligns it with timestamps, and encapsulates it according to the prescribed data format. The environmental point cloud dataset is stored in compressed binary format, the environmental depth information matrix uses a lossy compression algorithm to reduce data volume, and the flight attitude quaternion sequence is stored directly in floating-point format. The module adds a data packet header containing information such as the acquisition timestamp, UAV number, and sensor status, and finally generates a multi-source perception data packet, which is transmitted in real time to the global map fusion module 200 for subsequent processing via a gigabit Ethernet interface.

[0053] Due to differences in the viewing angles of various UAVs, variations in sensor characteristics, and dynamic environmental changes, these data exhibit spatiotemporal inconsistencies and observational uncertainties, making them difficult to directly use for collaborative path planning. Therefore, this invention introduces a global map fusion module 200. For example... Figure 3 As shown, this module, as the core processing unit of the system, is responsible for fusing heterogeneous sensing data collected by multiple UAVs to generate a unified environmental map. This module employs a probabilistic fusion algorithm based on an improved Bayesian update rule, achieving adaptive fusion of multi-source data through the following recursive formula:

[0054] ;

[0055] In the formula, This represents the updated posterior occupancy probability; This represents the prior occupancy probability of the grid cell at the current moment (ranging from 0 to 1), derived from the fusion results of historical observation data; This represents the occupancy probability (range 0~1) calculated from the current sensor observation.

[0056] This formula ensures that the confidence level converges quickly when data from multiple sensors are consistent; when data conflict, a gradual adjustment strategy is adopted.

[0057] During the data reception phase, the module first parses and synchronizes the input multi-source sensing data packets. It adopts a sliding window-based time alignment algorithm to unify the environmental point cloud datasets, environmental depth information matrices, and flight attitude quaternion sequences collected by different UAVs at different times into the same time reference. Through the pose information provided by the flight attitude quaternion sequences, all sensing data are transformed into the global coordinate system.

[0058] In the point cloud data processing stage, the module uses the voxel grid method to spatially resample the environmental point cloud dataset. First, a voxel grid with a resolution of 0.1 meters is established, and density statistics and feature extraction are performed on the point cloud within each voxel. For high-density voxels, their surface normal vectors and reflection intensity are calculated; for low-density voxels, noise filtering is performed. At the same time, the environmental depth information matrix is ​​converted into point cloud form to supplement the texture information collected by the visual sensor.

[0059] During the data fusion phase, the module assigns different confidence weights to different sensors. The initial confidence weight for LiDAR data is set to 0.8, and the initial confidence weight for visual depth data is set to 0.6. These weights are dynamically adjusted based on the sensor's operating status and environmental conditions. For each grid cell, the occupancy probability provided by different sensors is calculated, and then iterative fusion is performed using Bayes' theorem. Sensor conflicts are handled specially during the fusion process: when two sensors provide contradictory observations, the principle of closest time is adopted, prioritizing the latest observation data.

[0060] This module also implements a dynamic environment update mechanism. Each grid cell, in addition to recording the occupancy probability, maintains the last update timestamp and observation count statistics. For areas that have not been updated for a long time, an exponential decay model is used to reduce their confidence level.

[0061] ;

[0062] In the formula, This represents the new occupancy probability after attenuation processing; This represents the prior occupancy probability of a grid cell at the current time (ranging from 0 to 1). Indicator decay factor, It is the natural constant (approximately 2.71828); It is the decay rate constant ( >0), controls the rate of decay; It refers to the time interval or the elapsed time.

[0063] This mechanism ensures that the system can adapt to changes in the environment and prevents outdated information from affecting decision-making.

[0064] During the output phase, the module generates a globally consistent probability map, which is stored using a multi-layered data structure. The bottom layer is an occupancy probability raster map with a resolution of 0.1 meters, the middle layer stores obstacle height information, and the top layer records the dynamic movement trajectories of obstacles. Each raster contains the following attributes: occupancy probability value (0-1), last update time, data source identifier, and dynamic characteristic flag.

[0065] This module ensures map consistency through the following design: First, a data quality assessment mechanism is established to evaluate the credibility of the input data from each sensor and calculate it using mean squared error; second, a multi-hypothesis testing method is used to cross-validate conflicting data; and finally, spatiotemporal consistency checks are performed to eliminate instantaneous errors and outliers.

[0066] The generated globally consistent probabilistic map is output to the collaborative task planning module 300 via a gigabit Ethernet interface, while map update information is broadcast to each UAV at a frequency of 10Hz. The map data uses differential compression transmission, sending only data from areas that have changed, significantly reducing communication bandwidth requirements. This design enables multiple UAVs to make collaborative decisions based on a unified environmental perception, effectively solving the planning conflict problem caused by differences in perception.

[0067] The collaborative task planning module 300 receives a globally consistent probabilistic map generated by the global map fusion module 200. Based on the occupancy probability information, dynamic characteristic markers, and obstacle height data contained in this map, it performs multi-UAV task allocation and path planning. This module first decomposes the rescue task, dividing the overall area into multiple sub-task areas according to the characteristics of the fire scene environment and the rescue objectives. Each sub-task area contains specific reconnaissance or search and rescue mission objectives. The collaborative task planning module 300 employs a collaborative task allocation algorithm based on a market auction mechanism. Each UAV participates in task bidding as an intelligent agent. The bidding price is calculated based on factors such as the UAV's current location, battery status, sensor configuration, and distance to the task point. Finally, through distributed negotiation, a suitable UAV execution unit is allocated to each task.

[0068] During the path planning phase, the collaborative task planning module 300 employs an improved path search algorithm for global path planning, comprehensively considering path length, safety, and energy consumption. Path cost calculation fully references the occupancy probability values ​​in the globally consistent probability map, assigning higher risk costs to high-probability occupancy areas to ensure that generated paths avoid uncertain regions. Simultaneously, this module also considers the predicted trajectories of dynamic obstacles, reserving safety intervals during path planning to avoid conflicts with moving obstacles. For path planning involving multiple UAVs, the module adopts a spatiotemporal joint planning method, ensuring not only the spatial safety of each path but also preventing UAVs from encountering each other in confined spaces through time scheduling.

[0069] This module generates a detailed mission execution plan for each UAV, including a waypoint sequence (a list of 3D coordinates), expected speed, and mission time window (start / end time). Each waypoint contains parameters such as its 3D coordinate position, expected arrival time, and allowable error range. Simultaneously, the module generates emergency backup paths, allowing for rapid switching when the main path is blocked. All path data is smoothed using Bezier curves to ensure stable flight for the UAV.

[0070] In the output phase, this module generates a multi-UAV collaborative task planning scheme, which includes the task allocation results for each UAV, the global path sequence, and the time scheduling plan. The scheme is stored in a hierarchical structure: the top layer is the overall task allocation scheme, the middle layer is the path sequence of each UAV, and the bottom layer contains detailed waypoint parameters and execution commands. The planning scheme also includes a version identifier and a generation timestamp to ensure consistency with the map data.

[0071] The collaborative task planning module 300 transmits the generated multi-UAV collaborative task planning scheme to the local obstacle avoidance execution module 400 via a data interface, providing global path guidance for each UAV. Simultaneously, the time scheduling information in the scheme is also sent to the closed-loop feedback update module 500 as a benchmark reference for execution status monitoring.

[0072] The local obstacle avoidance execution module 400 receives a multi-UAV collaborative task planning scheme from the collaborative task planning module 300. Based on the global path sequence and waypoint parameters in the scheme, combined with real-time UAV sensor data, it achieves real-time adjustment and precise execution of the local path. This module first parses and loads the global path, converting the preset waypoint sequence into an executable flight trajectory, while simultaneously monitoring the UAV's current status, including position, speed, attitude, and battery level. The module employs a model predictive control algorithm to optimize the trajectory at millisecond-level frequencies, ensuring the UAV can smoothly track the global path.

[0073] For local obstacle avoidance, the module processes environmental data collected by onboard sensors in real time, including LiDAR point clouds, visual images, and depth information. When a temporary or dynamic obstacle not marked in the global consistency probability map deviates from the predicted trajectory, the module immediately initiates a local replanning algorithm. This algorithm generates a local detour path while maintaining the overall direction of the original global path, avoiding collisions with obstacles. The replanning process fully considers the dynamic constraints of the UAV, ensuring that the generated local path conforms to the UAV's maneuverability.

[0074] The module also implements a distributed coordination mechanism among multiple drones, exchanging position and intent information through direct communication between them. When there is a risk of conflict in the local paths of multiple drones, a priority-based avoidance strategy is adopted to ensure the orderliness of air traffic. At the same time, the module continuously evaluates the performance of local paths, records the deviation between the actual flight trajectory and the planned trajectory, and any abnormal situations encountered.

[0075] The local obstacle avoidance execution module 400 ultimately generates real-time control commands for the UAV, including throttle, control surface, and attitude control parameters. These commands directly drive the UAV to execute through the flight control system. Simultaneously, the module records detailed execution process data, including the actual flight trajectory, information on encountered obstacles, and energy consumption. This data is packaged into an execution log data package.

[0076] The closed-loop feedback update module 500, as the system's terminal module, receives execution log data packets from the local obstacle avoidance execution module 400, while continuously monitoring environmental change information in the global map fusion module 200. This module first analyzes and processes the execution data, extracting key information such as path execution deviation, obstacle occurrence frequency, and environmental change characteristics. By comparing the planned path with the actual path, it evaluates the effectiveness and accuracy of the planning scheme.

[0077] The module employs an incremental learning mechanism, continuously optimizing system parameters based on execution feedback. For areas with frequent obstacle occurrences, it automatically adjusts the occupancy probability weights in the globally consistent probability map; for road segments with multiple path deviations, it corrects the risk assessment parameters in the path cost function. Simultaneously, the module maintains an anomaly knowledge base, recording handling experience and optimization strategies for various emergencies.

[0078] At the system level, the module implements performance monitoring and health management functions, assesses the mission execution status of each drone in real time, and detects system performance degradation trends. When an anomaly or performance degradation is detected in a drone, a mission reassignment mechanism is activated to transfer the mission to other drones, ensuring the successful completion of the overall mission.

[0079] The closed-loop feedback update module 500 ultimately generates system optimization instructions and map update information, which are fed back in real time to the global map fusion module 200 and the collaborative task planning module 300. Optimization instructions include parameter adjustment suggestions, algorithm improvement plans, and strategy optimization guidance; map update information includes newly discovered obstacle data, details of environmental changes, and adjustments to regional risk levels. Through this closed-loop feedback mechanism, the system can continuously learn and evolve, gradually improving its adaptability and reliability in changing environments.

[0080] Thus, the system has completed a full closed loop from environmental perception, global planning, local execution to feedback optimization, realizing the ability of multiple UAVs to work safely, efficiently, and adaptively in complex fire environments.

[0081] In summary, this invention proposes a real-time path optimization system for unmanned aerial vehicles (UAVs) based on edge computing, comprising: a perception data acquisition module 100, a global map fusion module 200, a collaborative task planning module 300, a local obstacle avoidance execution module 400, and a closed-loop feedback update module 500.

[0082] The perception data acquisition module 100 is deployed on various UAV platforms to synchronously acquire multi-source sensor data and generate multi-source perception data packages containing environmental point cloud datasets, environmental depth information matrices, and flight attitude quaternion sequences.

[0083] The global map fusion module 200 is used to receive multi-source sensing data packets and generate a globally consistent probabilistic map with time-varying characteristics through spatiotemporal alignment processing, voxel gridding processing, feature-level fusion, Bayesian update-based probabilistic fusion algorithm and dynamic decay processing.

[0084] The collaborative task planning module 300 is used to receive a globally consistent probability map, and based on the globally consistent probability map, it uses a market auction mechanism to allocate multiple UAV tasks, perform global path planning, and generate a multi-UAV collaborative task planning scheme that includes waypoint sequence, expected speed, task time window and emergency backup path.

[0085] The local obstacle avoidance execution module 400 is used to receive multi-UAV collaborative task planning schemes and combine real-time sensor data to generate real-time UAV control commands using model predictive control algorithms and replanning algorithms, thereby achieving local trajectory optimization and dynamic obstacle avoidance.

[0086] The closed-loop feedback update module 500 receives the execution log data packets from the local obstacle avoidance execution module 400, and generates system optimization instructions through deviation analysis and incremental learning, which are then fed back to the global map fusion module 200 and the collaborative task planning module 300.

[0087] The steps in this invention can be adjusted, combined, or deleted according to actual needs.

[0088] The units in the device of the present invention can be merged, divided, or reduced according to actual needs.

[0089] In this invention, the terms "installation," "connection," "linking," and "fixing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; "linking" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of these terms in this invention according to the specific circumstances.

[0090] The shapes of the components in the accompanying drawings are schematic and may differ from their actual shapes. The drawings are only used to illustrate the principles of the present invention and are not intended to limit the present invention.

[0091] Although the invention has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and not intended to limit the application of the invention. The scope of protection of the invention is defined by the appended claims and may include various modifications, alterations, and equivalents made to the invention without departing from the scope and spirit of the invention.

Claims

1. A real-time path optimization system for unmanned aerial vehicles (UAVs) based on edge computing, characterized in that, The system includes: a perception data acquisition module (100), a global map fusion module (200), a collaborative task planning module (300), a local obstacle avoidance execution module (400), and a closed-loop feedback update module (500). The perception data acquisition module (100) is deployed on each UAV platform to synchronously acquire multi-source sensor data and generate multi-source perception data packets containing environmental point cloud datasets, environmental depth information matrices and flight attitude quaternion sequences. The global map fusion module (200) is used to receive multi-source sensing data packets and generate a globally consistent probabilistic map with time-varying characteristics through spatiotemporal alignment processing, voxel gridding processing, feature-level fusion, Bayesian update-based probabilistic fusion algorithm and dynamic decay processing. The collaborative task planning module (300) is used to receive the global consistency probability map, and to use the market auction mechanism to allocate multiple UAV tasks based on the global consistency probability map, perform global path planning, and generate a multi-UAV collaborative task planning scheme that includes waypoint sequence, expected speed, task time window and emergency backup path. The local obstacle avoidance execution module (400) is used to receive multi-UAV collaborative task planning schemes and combine real-time sensor data to generate real-time UAV control commands using model predictive control algorithms and replanning algorithms, thereby achieving local trajectory optimization and dynamic obstacle avoidance. The closed-loop feedback update module (500) is used to receive the execution log data packet of the local obstacle avoidance execution module (400), and generate system optimization instructions through deviation analysis and incremental learning, and feed them back to the global map fusion module (200) and the collaborative task planning module (300). The spatiotemporal alignment processing of the global map fusion module (200) is based on the flight attitude quaternion sequence, which transforms the environmental point cloud dataset and the environmental depth information matrix into the global coordinate system; The voxel meshing process uses a 0.1-meter resolution voxel mesh to perform density statistics and feature extraction on the spatiotemporally aligned environmental point cloud dataset, generating a standardized point cloud distribution map. Feature-level fusion converts the environmental depth information matrix into a point cloud format and supplements it with texture information; The probabilistic fusion algorithm based on Bayesian updates uses the following recursive formula for adaptive fusion of multi-source data: ； In the formula, This represents the updated posterior occupancy probability; This represents the prior occupancy probability of a grid cell at the current moment, derived from the fusion results of historical observation data; This represents the occupancy probability calculated from the current sensor observations; Dynamic decay processing uses an exponential decay model to reduce confidence: ； In the formula, This represents the new occupancy probability after attenuation processing; This represents the prior occupancy probability of a grid cell at the current time. Indicator decay factor, It is a natural constant; It is the decay rate constant, which controls how fast the decay occurs; It refers to the time interval or the elapsed time.

2. The UAV real-time path optimization system based on edge computing as described in claim 1, characterized in that, The sensing data acquisition module (100) adopts a master-slave acquisition architecture, with a high-precision time synchronization module as the master clock. It synchronizes the acquisition timing of the lidar, stereo vision camera and inertial measurement unit through hardware trigger signals to ensure that the multi-source data has a unified microsecond-level time reference. The perception data acquisition module (100) performs motion distortion compensation, coordinate system normalization and voxel grid filtering on the lidar point cloud data to generate an environmental point cloud dataset. The perception data acquisition module (100) performs distortion correction, semi-global matching and triangulation processing on the stereo vision image to generate an environmental depth information matrix; The sensing data acquisition module (100) performs zero bias compensation, temperature drift correction, scale factor calibration and complementary filtering on the inertial measurement unit data to generate a flight attitude quaternion sequence.

3. The UAV real-time path optimization system based on edge computing as described in claim 1, characterized in that, The sensing data acquisition module (100) encapsulates the processed data into a multi-source sensing data packet. The data packet contains the acquisition timestamp, UAV number and sensor status flag, and is transmitted in real time to the global map fusion module (200) through a gigabit Ethernet interface.

4. The UAV real-time path optimization system based on edge computing as described in claim 1, characterized in that, The global map fusion module (200) uses an exponential decay model to perform confidence decay processing on grid cells that have not been updated for a long time, and the decay coefficient is adaptively adjusted according to the degree of dynamic change in the environment.

5. The UAV real-time path optimization system based on edge computing as described in claim 1, characterized in that, The collaborative task planning module (300) comprehensively considers path length, safety and energy consumption indicators when planning the path, and generates an emergency backup path for each path; The multi-UAV collaborative mission planning scheme adopts a hierarchical storage structure, including mission allocation results, path sequences and waypoint parameters, and is smoothed by Bezier curves.

6. The UAV real-time path optimization system based on edge computing as described in claim 1, characterized in that, The collaborative task planning module (300) transmits the multi-UAV collaborative task planning scheme to the local obstacle avoidance execution module (400) and synchronizes the time scheduling information to the closed-loop feedback update module (500).

7. The edge computing-based real-time path optimization system for unmanned aerial vehicles as described in claim 1, characterized in that, The closed-loop feedback update module (500) receives the execution log data packet from the local obstacle avoidance execution module (400) and generates system optimization instructions to be fed back to the global map fusion module (200).

Images