An unmanned aerial vehicle automatic driving processing system based on a service scene distributed GPU computing architecture and an unmanned aerial vehicle light NPU architecture

Abstract

The application relates to the technical field of unmanned aerial vehicle automatic driving, in particular to an unmanned aerial vehicle automatic driving processing system based on a distributed GPU computing architecture and an unmanned aerial vehicle lightweight NPU architecture, wherein an emergency judgment engine is used to identify high real-time requirement emergency tasks, the emergency tasks are directly distributed to an airborne component for processing through an NPU computing architecture, non-emergency tasks are forwarded to an edge side GPU computing service module for preprocessing, uploaded to a central GPU computing aggregation server for batch processing, global path planning instructions are generated and pushed to the airborne component, real-time sensor data are fused by a decision system algorithm, a dynamic optimal flight path is generated and executed. Through the collaborative mode of "airborne NPU architecture implementation decision + edge side distributed GPU computing architecture implementation global optimization", the technical bottleneck of the unmanned aerial vehicle in computing power, endurance and environmental adaptability is broken.

Description

Technical Field

[0001] This invention relates to the field of drone autopilot technology, and in particular to a drone autopilot processing system based on a service scenario-based distributed GPU computing architecture and a lightweight drone NPU architecture. Background Technology

[0002] Drone autopilot is the core technology system that enables drones to autonomously complete the entire process of takeoff, navigation, obstacle avoidance, mission execution and landing. In essence, it enables drones to adapt to complex environments and complete preset tasks without real-time human intervention through a closed-loop link of "perception-decision-control-execution".

[0003] Currently, no drone on the market truly achieves semantic visual positioning technology (specifically, millimeter-level positioning). Analysis shows that consumer-grade drones currently achieve positioning accuracy at the meter level, primarily using satellite signals such as GPS / BeiDou for positioning assistance. Industrial-grade drones achieve RTK positioning accuracy at the centimeter level. Both of these positioning technologies rely on satellite and base station signals, making them susceptible to weather and electromagnetic interference, leading to problems such as inability to locate or low positioning accuracy.

[0004] For drones to achieve truly fast and efficient autonomous driving, their image recognition and processing efficiency needs to be improved. Traditional architectures rely on a single GPU module, which requires data to be moved back and forth between the processor and cache for each calculation. Furthermore, GPU chips have high power consumption, hindering long-distance drone flight endurance. This solution adds an NPU module. The NPU's computing architecture simplifies the design of storing and retrieving data for each calculation, improving processing efficiency and ensuring more efficient drone operation. Simultaneously, the NPU chip's low power consumption enhances the drone's long-distance flight endurance. Additionally, deploying GPU computing power service modules in service scenarios reduces the drone's processing load and lowers latency. Summary of the Invention

[0005] To address the technical challenges of traditional architectures that rely on a single GPU for processing, where data needs to be moved back and forth between the calculator and the cache, resulting in high GPU chip power consumption and hindering long-distance drone flight, this invention provides a drone autopilot processing system based on a service scenario-based distributed GPU computing architecture and a lightweight NPU architecture for the drone. The aim is to fully leverage the technical advantages of the scenario-based distributed GPU computing architecture and the airborne lightweight NPU architecture to provide technical support for drone autopilot.

[0006] A drone autopilot processing system based on a service scenario-based distributed GPU computing architecture and a lightweight NPU architecture for drones, the drone autopilot processing system consisting of airborne components, edge computing components and a drone-end GPU computing power aggregation server;

[0007] When a drone enters the service area, the emergency judgment engine identifies urgent tasks with high real-time requirements and directly assigns them to the onboard components for processing via the NPU computing architecture. Non-urgent tasks are forwarded to the edge GPU computing power service module for preprocessing. The preprocessed data is then uploaded to the central GPU computing power aggregation server via Ethernet for batch processing.

[0008] After the GPU computing power aggregation server completes data processing, it generates global path planning instructions and pushes them to the onboard components. The decision system algorithm then integrates real-time sensor data to generate a dynamically optimal flight path and executes it.

[0009] Preferably, the airborne components include an emergency judgment module, an image perception algorithm module, a path planning optimization algorithm module, and a decision system algorithm module.

[0010] Preferably, the image perception algorithm module is supported by a neural network processor (NPU).

[0011] Preferably, the service scenario side deploys computing power nodes according to different service scenarios. Its components are distributed GPU computing power architecture, including several GPU computing power service modules. In the drone flight hotspot area, multiple GPU computing power nodes are deployed to form a dense computing power network. In remote areas or specific operation areas, computing power nodes are reasonably set according to the operation range and task requirements to ensure coverage of the drone's operation area.

[0012] Preferably, the emergency judgment module identifies urgent tasks with high real-time requirements through the emergency judgment engine and directly assigns them to the airborne image perception algorithm module for processing.

[0013] Preferably, the image perception algorithm module extracts the semantic segmentation results and visual features of image data through a deep feature aggregation network, constructs a 3D semantic map, and integrates multi-sensor data to output the UAV pose.

[0014] Preferably, the path planning optimization algorithm module is based on visual features and semantic information, integrates dynamic obstacle prediction to update the dynamic cost map, uses an improved A* algorithm to search for global paths, and generates time-parameterized trajectories through B-spline optimization.

[0015] Preferably, the decision-making system algorithm module integrates UAV pose, time-parameterized trajectory and real-time scene information, and makes decisions collaboratively with the rule engine and deep learning model.

[0016] The beneficial effects of this invention are as follows:

[0017] First, by adopting a collaborative model of "onboard NPU architecture for decision-making + edge-side distributed GPU computing architecture for global optimization", we can overcome the technical bottlenecks of drones in terms of computing power, endurance, and environmental adaptability. Specifically, the lightweight NPU can process sensor data in real time, ensuring that the drone can respond to changes in the surrounding environment in a timely manner, while the distributed GPU computing architecture can provide powerful computing support for complex tasks.

[0018] Secondly, the present invention can adapt to different service scenarios and task requirements. By dynamically adjusting the architecture and task allocation, it can meet the autonomous driving requirements of drones in various complex environments. The distributed architecture makes it easier to expand computing nodes. As the scale of drone applications expands, it is easier to increase computing resources. Attached Figure Description

[0019] Figure 1 This is an architecture diagram of an autonomous drone processing system based on a service-scenario distributed GPU computing architecture and a lightweight NPU architecture for drones. Detailed Implementation

[0020] Example 1:

[0021] In the field of drone autopilot, efficient computing power and real-time processing capabilities are key to ensuring its safe and stable operation. Drone autopilot processing systems that combine a distributed GPU computing architecture based on service scenarios with a lightweight NPU architecture for drones can fully leverage the advantages of both architectures and provide strong technical support for drone autopilot.

[0022] The drone autopilot processing system of the present invention consists of two parts: a service scenario-based distributed GPU computing architecture and a drone lightweight NPU architecture, which work together through an efficient communication mechanism.

[0023] This service-scenario-based distributed GPU computing architecture aims to provide powerful backend computing support for drone autopilot processing, capable of handling complex computing tasks. The architecture deploys computing nodes according to different service scenarios. In urban areas, multiple GPU computing nodes can be deployed around major drone flight hotspots, such as logistics parks and urban commercial districts, forming a dense computing network. In remote areas or specific operational areas, such as farmland and mines, computing nodes are rationally configured based on the operational scope and task requirements to ensure coverage of the drone's operational area.

[0024] The lightweight NPU architecture for drones is integrated on the drone itself. It is responsible for real-time processing of perception data and simple control decisions during drone flight, meeting real-time requirements. It uses a low-power, high-performance lightweight NPU chip as the core processing unit, along with corresponding sensor interfaces, communication modules, and other hardware components. The NPU chip has a dedicated architecture for neural network computing, enabling efficient execution of deep learning algorithms and real-time processing of sensor data such as images and LiDAR. Furthermore, the NPU can process data collected by sensors on the drone in real time, such as images captured by cameras and data scanned by LiDAR. Through real-time analysis and processing of this data, it can promptly perceive changes in the surrounding environment, providing a basis for the drone's autopilot decisions.

[0025] The distributed GPU computing architecture and the lightweight NPU architecture for drones work closely together. During flight, for simple tasks with high real-time requirements, such as obstacle avoidance and attitude adjustment, the lightweight NPU completes the task independently. For computationally complex tasks that require a large amount of computing power, such as 3D reconstruction of large-scale scenes and path optimization for multi-drone collaborative operations, the NPU transmits the relevant data to the distributed GPU computing architecture for processing and returns the processing results to the NPU to guide the drone's flight.

[0026] The system can dynamically adjust the working mode and task allocation of both the drone and the NPU based on factors such as the drone's flight status, mission changes, and environmental conditions. For example, when the drone approaches a complex environment, the distributed GPU computing architecture will perform relevant calculations and analyses in advance, providing the NPU with more detailed environmental information and decision support; when network communication quality is poor, the NPU will take on more processing tasks to ensure that the drone's basic autopilot function is not affected.

[0027] The specific implementation scheme of the system of the present invention is as follows:

[0028] An autonomous drone processing system based on a service-scenario distributed GPU computing architecture and a lightweight NPU architecture for drones, such as... Figure 1 As shown, it consists of airborne components, edge computing components, and a hub-and-spoke GPU computing power aggregation server;

[0029] The airborne components include sensor hardware, emergency judgment module, image perception algorithm module, path planning optimization algorithm module, decision system algorithm module, and 5G-A / 6G communication module;

[0030] Among them, the emergency judgment module and the decision system algorithm module are both supported by the central processing unit (CPU), while the image perception algorithm module and the path planning optimization algorithm module are both supported by the neural network processor (NPU).

[0031] The service scenario components include GPU computing power service module 1 deployed in service scenario 1, GPU computing power service module 2 deployed in service scenario 2, ..., GPU computing power service module N deployed in service scenario N;

[0032] The 5G-A / 6G communication module integrates an RF-ID radio frequency module, which can quickly identify the GPU computing power service module in the edge computing terminal component and quickly realize data interaction;

[0033] The interaction flow of each component in the drone autopilot processing system is as follows:

[0034] The sensor hardware senses environmental conditions and transmits the collected environmental perception data (including image data of roads, objects, obstacles, etc.) to the emergency judgment module. The emergency judgment module runs on the CPU and uses the emergency judgment engine to identify emergency tasks with high real-time requirements, such as obstacle avoidance and target recognition, and directly assigns them to the airborne image perception algorithm module for processing.

[0035] The image perception algorithm module runs on the NPU. It extracts semantic segmentation results and visual features from image data such as roads, objects, and obstacles through a deep feature aggregation network, constructs a 3D semantic map, and outputs the UAV pose by fusing data from multiple sensors such as IMU and Beidou. The semantic information and visual features are transmitted to the path planning and optimization algorithm module, and the UAV pose is simultaneously transmitted to the decision system algorithm module.

[0036] The path planning optimization algorithm module runs on the NPU. Based on visual features and semantic information, it integrates dynamic obstacle prediction to update the dynamic cost map, uses an improved A* algorithm to search for global paths, generates time-parameterized trajectories through B-spline optimization, and outputs them to the decision system algorithm module.

[0037] The decision-making system algorithm module runs on the CPU, integrating the drone's pose, time-parameterized trajectory, and real-time scene information. It makes decisions collaboratively with the rule engine and deep learning model, and outputs driving behavior commands, which are then sent to the CPU module for control execution.

[0038] When the drone enters service scenario area 1, the onboard emergency judgment module runs on the CPU and identifies non-emergency tasks through the emergency judgment engine. The non-emergency tasks are forwarded to the edge GPU computing power service module 1 through the onboard 5G-A / 6G communication module. The GPU computing power service module 1 performs data decompression, format conversion and data labeling preprocessing. The preprocessed data is uploaded to the central GPU computing power aggregation server through Ethernet EtherNet1 for batch and non-real-time processing.

[0039] When the drone enters service scenario area 2, the onboard emergency judgment module runs on the CPU and identifies non-emergency tasks through the emergency judgment engine. The non-emergency tasks are forwarded to the edge GPU computing power service module 2 through the onboard 5G-A / 6G communication module. The GPU computing power service module 2 performs data decompression, format conversion and data labeling preprocessing. The preprocessed data is uploaded to the central GPU computing power aggregation server through Ethernet EtherNet2 for batch and non-real-time processing.

[0040] When the drone enters the service scenario area N, the onboard emergency judgment module runs on the CPU and identifies non-emergency tasks through the emergency judgment engine. The non-emergency tasks are forwarded to the edge GPU computing power service module N through the onboard 5G-A / 6G communication module. The GPU computing power service module N performs data decompression, format conversion and data labeling preprocessing. The preprocessed data is uploaded to the central GPU computing power aggregation server through Ethernet EtherNetN for batch and non-real-time processing.

[0041] After the GPU computing power aggregation server completes the data processing of GPU computing power service module 1, GPU computing power service module 2, and GPU computing power service module N, it generates a global path planning instruction and pushes it to the airborne 5G-A / 6G communication module in real time via the 5G link. After the 5G-A / 6G communication module verifies the data integrity, it forwards it to the airborne decision system algorithm module. The decision system algorithm then integrates real-time sensor data to generate a dynamic optimal flight path and executes it.

[0042] The procedure for the drone to offload non-urgent task data to the GPU computing service module in the distributed GPU computing architecture at the edge computing terminal via the 5G-A / 6G communication module is as follows:

[0043] Program 1: The decision system algorithm module generates data M for broadcasting non-urgent tasks. CPU→GPU And the data encapsulation parameter SP(M) with a bit length of L CPU→GPU ), First auxiliary analytical parameter KP(M CPU→GPU )-Ⅰ and second auxiliary analytical parameters KP(M CPU→GPU )-Ⅱ;

[0044] Non-urgent task data M CPU→GPU Group the data to obtain several groups, and make each group M... CPU→GPU-i The number of bits is L;

[0045] Using data encapsulation parameters SP(M) CPU→GPU For non-urgent task data M CPU→GPU The encapsulation operation was performed, and the specific encapsulation result is as follows:

[0046] The non-urgent task data is XORed sequentially with the data encapsulation parameters, and then the XOR results are concatenated to obtain the non-urgent task encapsulation data, denoted as C. CPU→GPU ;

[0047] C CPU→GPU =[M CPU→GPU-1 ⊕SP(M CPU→GPU )]ǁ[M CPU→GPU-2 ⊕SP(M CPU→GPU )]ǁ...ǁ[M CPU→GPU-i ⊕SP(M CPU→GPU )];

[0048] The result of the XOR operation between the first auxiliary parsing parameter and the data encapsulation parameter is the first auxiliary parsing information, which is F(M). CPU→GPU )-Ⅰ=KP(M CPU→GPU )-Ⅰ⊕SP(M CPU→GPU );

[0049] The result of the XOR operation between the second auxiliary parsing parameter and the data encapsulation parameter is the second auxiliary parsing information, which is F(M). CPU→GPU )-Ⅱ=KP(M CPU→GPU )-Ⅱ⊕SP(M CPU→GPU );

[0050] Non-urgent task encapsulation data is broadcast to the GPU computing power service module, while the data encapsulation and parsing rules, the first auxiliary parsing parameters, the second auxiliary parsing parameters, the first auxiliary parsing information, and the second auxiliary parsing information are pre-transmitted to the GPU computing power service module through a secure channel.

[0051] Program 2: The GPU computing power service module performs the following calculation steps:

[0052] Calculate the data encapsulation parameter SP'(M) CPU→GPU )-a=F(M CPU→GPU )-Ⅰ⊕KP(M CPU→GPU )-Ⅰ;

[0053] Calculate the data encapsulation parameter SP'(M) CPU→GPU )-b=F(M CPU→GPU )-Ⅱ⊕KP(M CPU→GPU )-Ⅱ;

[0054] If the equation SP'(M) CPU→GPU )-a=SP'(M CPU→GPU When -b is true, then SP'(M) can be determined. CPU→GPU )-a or SP'(M CPU→GPU -b is the data encapsulation parameter SP'(M) CPU→GPU );

[0055] That is, there exists SP'(M) CPU→GPU )=SP'(M CPU→GPU )-a=SP'(M CPU→GPU )-b;

[0056] Conversely, incorrect information is returned to the drone;

[0057] Using SP'(M CPU→GPU Encapsulate data C for non-urgent tasks CPU→GPU The parsing operation is performed, and the specific parsing result is as follows:

[0058] Encapsulate non-urgent tasks into data C CPU→GPU Group the data to obtain several grouped data sets, and make each group of data C... CPU→GPU-i The number of bits is L;

[0059] M' CPU→GPU =[C CPU→GPU-1 ⊕SP'(M CPU→GPU )]ǁ[C CPU→GPU-2 ⊕SP'(M CPU→GPU )]ǁ...ǁ[C CPU→GPU-i ⊕SP'(M CPU→GPU )];

[0060] At this point, the non-urgent task data uninstallation process has ended;

[0061] The above data offloading method can efficiently support one-to-many communication, reduce the communication and computing overhead of drones, and ensure data confidentiality.

[0062] Example 2:

[0063] The collaborative system of distributed GPU computing architecture and lightweight NPU architecture for drones has demonstrated unique value in fields such as intelligent urban logistics and distribution, precision agriculture and plant protection operations, emergency rescue and disaster investigation, power and energy facility inspection, urban traffic monitoring and management, mineral resource exploration and mapping, environmental monitoring and ecological protection, and photovoltaic power plant operation and management, thanks to its efficient computing power scheduling, real-time data processing, and dynamic environmental adaptability. Its specific application process is as follows:

[0064] The application process in the field of smart logistics and distribution in cities is as follows: In the last-mile delivery scenario in dense urban areas, the lightweight NPU processes camera and LiDAR data in real time, enabling drones to achieve "centimeter-level" autonomous flight. The YOLOv8-tiny model identifies landing point features and makes obstacle avoidance decisions for dynamic obstacles. Meanwhile, the distributed GPU cluster is responsible for global path planning and, combined with real-time traffic control information, generates dynamically updated optimal routes for drones.

[0065] The application process in the field of precision agriculture and plant protection operations is as follows: In the scenario of tens of thousands of acres of farmland, the system can realize intelligent management of the entire process. The NPU on the drone processes multispectral camera data in real time, identifies crop growth and pest and disease areas through a lightweight UNet model, and controls plant protection nozzles to spray in a variable manner. Meanwhile, the distributed GPU is responsible for integrating the data collected by multiple drones, building a digital twin model of farmland, predicting crop growth trends by combining meteorological data, and generating refined operation plans.

[0066] The application process in the field of emergency rescue and disaster investigation is as follows: At the disaster sites such as earthquakes and floods, the NPU module of the UAV identifies the vital signs of survivors in real time through infrared thermal imaging sensors, and generates a preliminary rescue priority map by penetrating obstacles through millimeter-wave radar. Meanwhile, the distributed GPU cluster integrates the data from multiple UAVs into a panoramic disaster map, and quickly restores the scope of building collapse and the information of flooded areas through three-dimensional reconstruction algorithms, providing decision support for the command center.

[0067] In the field of power and energy facility inspection, the application process is as follows: For large facilities such as high-voltage transmission lines and wind turbine towers, the UAV's NPU analyzes the conductor images captured by high-definition cameras in real time, identifies broken strands and corrosion defects through the ResNet-18 model, while the distributed GPU is responsible for integrating massive inspection data, building equipment health assessment models, and predicting component aging trends.

[0068] In the field of urban traffic monitoring and management, the application process is as follows: In scenarios such as urban main roads and intersections, the UAV's NPU processes camera data in real time, identifies behaviors such as running red lights and illegal parking through lightweight target detection algorithms, and simultaneously calculates parameters such as traffic flow and vehicle speed. Meanwhile, the distributed GPU cluster integrates multi-machine data into a dynamic traffic heat map, combines historical data to predict congestion nodes, and provides optimization suggestions for traffic signal timing. At the scene of a sudden traffic accident, the system supports the UAV to quickly arrive at the scene, the NPU transmits accident images in real time and marks the location of the injured and deceased, and the GPU can simultaneously plan the optimal route for rescue vehicles, shortening the emergency response time.

[0069] The application process in the field of mineral resource exploration and surveying is as follows: In large open-pit mines, the NPU of the UAV processes the lidar point cloud data in real time, identifies the ore-rock boundary and slope cracks through the lightweight PointNet model, and simultaneously calculates whether the slope of the mining face is within the safe threshold. Meanwhile, the distributed GPU is responsible for building a three-dimensional model of the mine, simulating the mining progress by combining blasting vibration data, and optimizing the blasting scheme to reduce resource waste.

[0070] In the field of environmental monitoring and ecological protection, the application process is as follows: In ecologically sensitive areas such as nature reserves and wetlands, the NPU of the UAV processes infrared camera data in real time, accurately identifies the species and activity trajectories of wild animals, and avoids obstacles, while the distributed GPU cluster integrates monitoring data from multiple periods to analyze the trend of species habitat change; In the scenario of marine oil pollution monitoring, the UAV's NPU quickly identifies the extent of oil slick spread, while the GPU combines ocean current models to predict the pollution spread path, providing accurate coordinates for the scheduling of pollution cleanup vessels;

[0071] The specific application process in the field of photovoltaic power plant operation and management is as follows:

[0072] (I) Intelligent Inspection of Photovoltaic Modules: A large number of photovoltaic modules in photovoltaic power plants are exposed to the outdoors for a long time, which is susceptible to problems such as dust cover, cell microcracks, and frame corrosion, which affect the power generation efficiency. The lightweight NPU carried by the drone can process the photovoltaic module images captured by the high-definition camera in real time. Through the lightweight ResNet model, it can accurately identify the dust accumulation area, cell microcracks and frame corrosion on the module surface. At the same time, the NPU can also combine infrared thermal imaging sensor data to quickly detect the temperature anomalies of the module and determine whether there is a hot spot effect. The distributed GPU computing architecture is responsible for integrating the inspection data of multiple drones to build a three-dimensional model of the entire photovoltaic power plant. Through in-depth analysis of massive image data, the GPU can generate a health status report for each photovoltaic module, accurately locate the problematic module, and prioritize according to the severity of the problem, providing clear maintenance guidance for operation and maintenance personnel.

[0073] (II) Planning and Execution of Photovoltaic Panel Cleaning Operations: Dust accumulation on the surface of photovoltaic panels will significantly reduce light absorption efficiency and affect power generation. During the inspection process, the NPU records the dust coverage on the surface of the photovoltaic panels in real time and transmits the data to the distributed GPU. The GPU combines local meteorological data (such as wind speed, rainfall, and solar radiation intensity) and photovoltaic panel power generation data to analyze the impact of dust on power generation efficiency and formulate the optimal cleaning schedule and cleaning path. During the cleaning operation, the NPU on the drone can adjust the flight attitude and working status of the cleaning device in real time according to the path planned by the GPU.

[0074] (III) Monitoring and Safety Early Warning of the Surrounding Environment of the Power Station: Changes in the surrounding environment of the photovoltaic power station may threaten the safe operation of the power station. For example, overgrown weeds may cause fires, and bird nests may cause short circuits in the modules. The NPU of the drone can use LiDAR and camera data to identify the growth of weeds, bird activity tracks and foreign objects that may invade the power station in real time. When the height of weeds exceeds the safety threshold or birds nest near the modules, the NPU will immediately issue an early warning signal and transmit the relevant data to the GPU. The distributed GPU computing architecture will comprehensively analyze this environmental data, combine historical data to predict potential safety risks, predict the possibility of fire based on the growth rate of weeds and wind direction, and plan the weeding operation path of the drone in advance. For foreign objects that may affect the safety of the power station, the GPU can calculate their movement trajectory, determine whether they will cause collision risk to the photovoltaic modules, and dispatch drones to intercept or clear them.

[0075] (iv) Optimization of operation and maintenance of photovoltaic power plants: The distributed GPU computing architecture collects real-time power generation data, equipment operating parameters and inspection and cleaning data of photovoltaic power plants. Through big data analysis and machine learning algorithms, a power generation efficiency prediction model of photovoltaic power plants is established. This model can accurately predict the power generation under different weather conditions and different component states, helping power plant operators to formulate reasonable power generation plans. At the same time, the GPU can also analyze the failure modes and life trends of equipment, predict possible failures in advance, such as inverter failures and cable aging, and promptly remind operation and maintenance personnel to carry out preventive maintenance to reduce downtime.

[0076] The above application scenarios collectively embody the core value of the system: through the collaborative mode of "onboard NPU architecture implementation decision + edge-side global optimization", it breaks through the technical bottlenecks of drones in terms of computing power, endurance, and environmental adaptability.

Claims

1. A drone autopilot processing system based on a service scenario-driven distributed GPU computing architecture and a lightweight drone NPU architecture, characterized in that, The drone autopilot processing system consists of airborne components, edge computing components, and a drone-end GPU computing power aggregation server. When a drone enters the service area, the emergency judgment engine identifies urgent tasks with high real-time requirements and directly assigns them to the onboard components for processing via the NPU computing architecture. Non-urgent tasks are forwarded to the edge GPU computing power service module for preprocessing. The preprocessed data is then uploaded to the central GPU computing power aggregation server via Ethernet for batch processing. After the GPU computing power aggregation server completes data processing, it generates global path planning instructions and pushes them to the onboard components. The decision system algorithm then integrates real-time sensor data to generate a dynamically optimal flight path and executes it.

2. The unmanned aerial vehicle autopilot processing system according to claim 1, characterized in that, The airborne components include an emergency judgment module, an image perception algorithm module, a path planning optimization algorithm module, and a decision system algorithm module.

3. The unmanned aerial vehicle autopilot processing system according to claim 1, characterized in that, The image perception algorithm module is supported by the neural network processor (NPU).

4. The unmanned aerial vehicle autopilot processing system according to claim 1, characterized in that, The service scenario deploys computing power nodes according to different service scenarios. Its components are distributed GPU computing power architecture, including several GPU computing power service modules. In the drone flight hotspot area, multiple GPU computing power nodes are deployed to form a dense computing power network. In remote areas or specific operation areas, computing power nodes are reasonably set according to the operation range and task requirements to ensure coverage of the drone operation area.

5. The unmanned aerial vehicle autopilot processing system according to claim 1, characterized in that, The emergency judgment module identifies urgent tasks with high real-time requirements through the emergency judgment engine and directly assigns them to the airborne image perception algorithm module for processing.

6. The unmanned aerial vehicle autopilot processing system according to claim 1, characterized in that, The image perception algorithm module extracts semantic segmentation results and visual features from image data through a deep feature aggregation network, constructs a 3D semantic map, and outputs the UAV pose by fusing multi-sensor data.

7. The unmanned aerial vehicle autopilot processing system according to claim 1, characterized in that, The path planning optimization algorithm module is based on visual features and semantic information, integrates dynamic obstacle prediction to update the dynamic cost map, uses an improved A* algorithm to search for global paths, and generates time-parameterized trajectories through B-spline optimization.

8. The unmanned aerial vehicle autopilot processing system according to claim 1, characterized in that, The decision-making system's algorithm module integrates UAV pose, time-parameterized trajectory, and real-time scene information, and makes decisions collaboratively with a rule engine and a deep learning model.

9. The unmanned aerial vehicle autopilot processing system according to any one of claims 1-8, characterized in that, The procedure for the drone to offload non-urgent task data to the GPU computing service module in the distributed GPU computing architecture at the edge computing terminal via the 5G-A / 6G communication module is as follows: Program 1: The decision-making system algorithm module generates data M for broadcasting non-urgent tasks. CPU→GPU And the data encapsulation parameter SP(M) with a bit length of L CPU→GPU ), First auxiliary analytical parameter KP(M CPU→GPU )-Ⅰ and second auxiliary analytical parameters KP(M CPU→GPU )-Ⅱ; Non-urgent task data M CPU→GPU Group the data to obtain several groups, and make each group M... CPU→GPU-i The number of bits is L; Using data encapsulation parameters SP(M) CPU→GPU For non-urgent task data M CPU→GPU The encapsulation operation was performed, and the specific encapsulation result is as follows: The non-urgent task data is XORed sequentially with the data encapsulation parameters, and then the XOR results are concatenated to obtain the non-urgent task encapsulation data, denoted as C. CPU→GPU ; C CPU→GPU =[M CPU→GPU-1 ⊕SP(M CPU→GPU )]ǁ[M CPU→GPU-2 ⊕SP(M CPU→GPU )]ǁ...ǁ[M CPU→GPU-i ⊕SP(M CPU→GPU )]; The result of the XOR operation between the first auxiliary parsing parameter and the data encapsulation parameter is the first auxiliary parsing information, which is F(M). CPU→GPU )-Ⅰ=KP(M CPU→GPU )-Ⅰ⊕SP(M CPU→GPU ); The result of the XOR operation between the second auxiliary parsing parameter and the data encapsulation parameter is the second auxiliary parsing information, which is F(M). CPU→GPU )-Ⅱ=KP(M CPU→GPU )-Ⅱ⊕SP(M CPU→GPU ); Non-urgent task encapsulation data is broadcast to the GPU computing power service module, while the data encapsulation and parsing rules, the first auxiliary parsing parameters, the second auxiliary parsing parameters, the first auxiliary parsing information, and the second auxiliary parsing information are pre-transmitted to the GPU computing power service module through a secure channel. Program 2: The GPU computing power service module performs the following calculation steps: Calculate the data encapsulation parameter SP'(M) CPU→GPU )-a=F(M CPU→GPU )-Ⅰ⊕KP(M CPU→GPU )-Ⅰ; Calculate the data encapsulation parameter SP'(M) CPU→GPU )-b=F(M CPU→GPU )-Ⅱ⊕KP(M CPU→GPU )-Ⅱ; If the equation SP'(M) CPU→GPU )-a=SP'(M CPU→GPU When -b is true, then SP'(M) can be determined. CPU→GPU )-a or SP'(M CPU→GPU -b is the data encapsulation parameter SP'(M) CPU→GPU ); That is, there exists SP'(M) CPU→GPU )=SP'(M CPU→GPU )-a=SP'(M CPU→GPU )-b; Conversely, incorrect information is returned to the drone; Using SP'(M CPU→GPU Encapsulate data C for non-urgent tasks CPU→GPU The parsing operation is performed, and the specific parsing result is as follows: Encapsulate non-urgent tasks into data C CPU→GPU Group the data to obtain several grouped data sets, and make each group of data C... CPU→GPU-i The number of bits is L; M' CPU→GPU =[C CPU→GPU-1 ⊕SP'(M CPU→GPU )]ǁ[C CPU→GPU-2 ⊕SP'(M CPU→GPU )]ǁ...ǁ[C CPU→GPU-i ⊕SP'(M CPU→GPU )]; At this point, the non-urgent task data uninstallation process has ended.

10. The unmanned aerial vehicle autopilot processing system according to claim 9, characterized in that, The drone autopilot processing system based on a service scenario-based distributed GPU computing architecture and a lightweight NPU architecture has demonstrated its application value in fields such as urban intelligent logistics and distribution, precision agriculture and plant protection operations, emergency rescue and disaster investigation, power and energy facility inspection, urban traffic monitoring and management, mineral resource exploration and mapping, environmental monitoring and ecological protection, and photovoltaic power station operation and management.

Images