An ai-based low-altitude economy unmanned aerial vehicle data processing system and method thereof

Abstract

This invention discloses an AI-based low-altitude economic unmanned aerial vehicle (UAV) data processing system and method, belonging to the field of UAV intelligent algorithm technology. The system includes: a flight path planning engine built based on a multimodal AI large model; an UAV equipped with an image sensor and a wireless transmission module; a ground processing center with a built-in AI intelligent processing model, a multimodal pre-annotation algorithm module, and an annotation algorithm feedback module; a model adaptive training engine with a built-in dynamic training strategy library, configured to automatically access labeled data that has passed algorithm verification; and an application feedback module configured to deploy iteratively optimized models, automatically analyze preprocessed real-time image data, and output structured application results. This invention constructs a fully intelligent closed-loop architecture encompassing "flight path planning - data processing - model training - scenario application," achieving real-time data linkage and collaboration across all stages. Core processes do not require deep manual intervention, significantly reducing labor costs and improving operational efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent algorithm technology for unmanned aerial vehicles (UAVs), and more specifically, to an AI-based data processing system and method for low-altitude economical UAVs. Background Technology

[0002] The low-altitude economy is an economic model that utilizes low-altitude aircraft such as drones as its core equipment to conduct production and services in low-altitude airspace. It has been widely applied in surveying and mapping, urban governance, digital rural development, and environmental monitoring. Drone data processing is a core support for the low-altitude economy. Specifically, it refers to the technological system that uses image sensors such as high-definition cameras to collect scene image data, and through processes such as preprocessing, feature analysis, and intelligent modeling, transforms the raw images into structured information that meets application requirements. AI technology provides crucial support for its automated and intelligent upgrades and is an important driving force for the high-quality development of the low-altitude economy.

[0003] In the context of large-scale and complex low-altitude economic drone scenarios, the following key issues exist: the flight path planning stage lacks dynamic adaptation and end-to-end collaboration capabilities, making it impossible to dynamically adjust flight strategies based on subsequent data processing results and application requirements; the automation level of data preprocessing and annotation is limited, and manual intervention is still required to filter and process the data even after automatic annotation. In view of this, we propose an AI-based low-altitude economic drone data processing system and method. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based low-altitude economic unmanned aerial vehicle (UAV) data processing system and method to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An AI-based low-altitude economic drone data processing system includes:

[0007] The route planning engine is built on a multimodal AI large model. It is configured to take into account the task requirements and real-time environmental data of public service scenarios. Through the built-in dynamic decision-making algorithm, it can autonomously generate the optimal flight path, waypoint coordinates and flight parameters, control the drone to take off, cruise and land automatically, and receive feedback information from scenario applications for subsequent route optimization.

[0008] The drone is equipped with an image sensor and a wireless transmission module. It is configured to collect scene image data according to the route generated by the route planning engine, and transmit the collected data back to the ground processing center in real time through the wireless transmission module. The acquisition parameters can be adjusted according to the data processing feedback.

[0009] The ground processing center is equipped with an AI intelligent processing model, a multimodal pre-annotation algorithm module, and an annotation algorithm feedback module. It is configured to perform pre-processing operations such as noise reduction, stitching, and format standardization on the returned data through the AI ​​intelligent processing model, achieve automatic annotation of image targets through the multimodal pre-annotation algorithm module, receive manual inspection and review results and optimize annotation logic through the annotation algorithm feedback module, and link with the acquisition module of the UAV through the collaborative control interface.

[0010] The model adaptive training engine has a built-in dynamic training strategy library. It is configured to automatically access labeled data that has passed algorithm verification, select the optimal training algorithm and adjust training parameters based on data type, data volume changes and application feedback results, trigger incremental training or full update training, and optimize the model's feature extraction network and decision logic. The labeled data that has passed algorithm verification is the labeled data output by the ground processing center and verified by the labeling algorithm feedback module.

[0011] The application feedback module is configured to deploy the iteratively optimized model, automatically analyze the preprocessed real-time image data and output structured application results, and extract error data and user adjustment records during the application process to form model optimization feedback information that is sent back to the model adaptive training engine. The feedback information is also synchronized to the route planning engine.

[0012] Preferably, the flight path planning engine does not require manual pre-setting of key parameters, can adapt to complex environments, and does not require manual intervention to adjust flight parameters; the UAV's image sensor is a high-definition camera, and the wireless transmission module supports real-time transmission of collected data; the annotation algorithm feedback module feeds back the results of manual spot checks to the multimodal pre-annotation algorithm module in real time, dynamically optimizing the annotation logic.

[0013] Preferably, the multimodal pre-annotation algorithm module, in conjunction with the AI ​​automatic scoring model, accurately filters low-quality labeled data, and manual adjustments and reviews are only made for low-scoring data, without the need to screen and correct all labeled data one by one.

[0014] Preferably, the dynamic training strategy library pre-configures data volume thresholds and task type matching rules. In response to the addition of flight tasks or the accumulation of collected data to a preset threshold, it automatically determines the training mode—incremental training is triggered when the data volume accumulates to a preset incremental threshold, and full update training is triggered when the data volume accumulates to a preset full update threshold or when a new data type is added. The optimization results of the model adaptive training engine are synchronized to the UAV edge and the ground processing center to achieve unified update and deployment with consistent model versions and synchronous effects.

[0015] Preferably, the structured application results output by the application feedback module do not require full manual verification; only minor adjustments and reviews are needed for the results of key scenarios.

[0016] A data processing method for low-altitude economic drones based on AI includes the following steps:

[0017] S1, Intelligent Route Planning and Dynamic Adaptation: Based on a multimodal AI large model, the route planning engine takes into account the task requirements of public service scenarios and real-time environmental data, and autonomously generates the optimal flight path, waypoint coordinates and flight parameters through the built-in dynamic decision-making algorithm to control the drone to take off, cruise and land automatically.

[0018] S2. Automated data acquisition, processing and annotation: The UAV collects scene image data through image sensors according to the route planned in S1 and transmits it back to the ground processing center in real time. The ground processing center completes data preprocessing through AI intelligent processing model, realizes automatic annotation of image targets through multimodal pre-annotation algorithm, and manually checks and reviews a very small number of abnormal data. The review results are fed back to the annotation algorithm module to optimize the annotation logic. At the same time, the UAV acquisition parameters are adjusted according to the preprocessed image quality data.

[0019] S3. Automated model iterative training based on job data: The model adaptive training engine automatically accesses the verified labeled data output by S2. Based on the data type, data volume changes and application feedback results, it selects the optimal training algorithm from the dynamic training strategy library, adjusts the training parameters, triggers incremental training or full update training, and optimizes the model using the newly added data.

[0020] S4, Model Application and Closed-Loop Feedback: The model optimized by S3 is deployed at the edge of the UAV or the ground processing center. The real-time image data preprocessed by S2 is automatically analyzed and structured application results are output. Error data and user adjustment records in the application process are extracted to form feedback information, which is sent back to the model adaptive training engine of S3 and the flight path planning engine of S1 to complete the closed loop of the whole process.

[0021] Preferably, in step S1, the route planning engine does not require manual preset of key parameters and can adapt to complex environments; in step S2, data preprocessing includes noise reduction, splicing, and format standardization operations, and the data acquisition and processing processes are linked through a collaborative control interface.

[0022] Preferably, step S2 adopts a collaborative mode of "automatic annotation + AI automatic scoring + manual targeted adjustment". Low-quality annotated data is filtered out by AI automatic scoring, and only low-scoring data is processed manually, reducing the workload of manual intervention.

[0023] Preferably, in step S3, the model adaptive training engine responds to the accumulation of data and automatically selects incremental training or full update training mode, and the optimized feature extraction network and decision logic are synchronously applied to subsequent data processing and scenario application stages.

[0024] Preferably, the feedback information in step S4 includes model optimization feedback information and route optimization data support information, which are used to optimize the model training direction and route planning for subsequent similar tasks, respectively; the structured application results only require manual review of key scenario results with minor adjustments, without the need for comprehensive verification.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] (1) This invention constructs a fully intelligent closed-loop architecture encompassing "route planning - data processing - model training - scenario application," enabling real-time data linkage and collaboration across all stages. This solves the problems of fragmented processes and the need for manual intervention in existing technologies. Core processes do not require deep manual intervention, only a small amount of manual sampling and review, significantly reducing labor costs and improving operational efficiency. Based on the accumulation of operational data, the model achieves automated iterative training. As flight missions increase, the amount of collected data continues to grow. The model automatically optimizes parameters and algorithm logic through incremental training, solving the shortcomings of weak generalization ability and inability to improve accuracy in existing technologies. This achieves continuous improvement in model accuracy and precision, adapting to diverse public service scenarios.

[0027] (2) The route planning stage of this invention adopts a multimodal AI large model to autonomously decide the optimal route, which does not require manual parameter preset and can adapt to the environment. This solves the problem that the route planning of existing technologies relies on manual intervention and requires manual adjustment to adapt to complex environments, thus improving the intelligence and adaptability of route planning. The collaborative mode of "automatic annotation + AI automatic scoring + manual targeted adjustment" is adopted. The annotation algorithm and scoring model are optimized by feedback, and the data collection and processing processes are coordinated and linked. This solves the problems of insufficient accuracy of automatic annotation, the need for manual screening and correction, and process disconnect in existing technologies. The AI ​​automatic scoring can accurately filter low-quality annotation data. The human operator only needs to adjust the low-scoring data. Under the premise of ensuring the quality of the annotation data, the workload of manual intervention is minimized and the annotation efficiency is improved. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0030] Example:

[0031] Please see Figure 1 A data processing system for low-altitude economic drones based on AI, comprising:

[0032] The flight path planning engine, built on a multimodal AI model, is configured to take into account the task requirements and real-time environmental data of public service scenarios (such as surveying, urban governance, digital villages, and environmental monitoring). Through a built-in dynamic decision-making algorithm, it autonomously generates the optimal flight path, waypoint coordinates, and flight parameters, controlling the drone's automatic takeoff, cruise, and landing. It can also receive feedback from scenario applications for subsequent flight path optimization. The flight path planning engine requires no manual pre-setting of key parameters, can adapt to complex environments, and requires no manual intervention to adjust flight parameters. The drone's image sensor is a high-definition camera, and the wireless transmission module supports real-time data transmission. The annotation algorithm feedback module feeds back the results of manual spot checks to the multimodal pre-annotation algorithm module in real time, dynamically optimizing the annotation logic.

[0033] The drone, equipped with an image sensor and a wireless transmission module, is configured to collect scene image data according to the flight path generated by the flight path planning engine, and transmit the collected data back to the ground processing center in real time through the wireless transmission module. It can also adjust the collection parameters (such as camera shooting resolution, frame rate, etc.) based on the data processing feedback.

[0034] The ground processing center is equipped with an AI intelligent processing model, a multimodal pre-annotation algorithm module, and an annotation algorithm feedback module. It is configured to perform pre-processing operations such as noise reduction, stitching, and format standardization on the returned data through the AI ​​intelligent processing model, achieve automatic annotation of image targets through the multimodal pre-annotation algorithm module, receive manual spot check results and optimize annotation logic through the annotation algorithm feedback module, and link with the acquisition module of the UAV through a collaborative control interface.

[0035] Among them, the multimodal pre-annotation algorithm module, together with the AI ​​automatic scoring model, accurately filters low-quality labeled data. Manual adjustments and reviews are only made for low-scoring data, without the need to screen and correct all labeled data one by one.

[0036] The model adaptive training engine has a built-in dynamic training strategy library. It is configured to automatically access labeled data that has passed algorithm verification (i.e., labeled data output by the ground processing center and verified). Based on the data type, data volume changes, and application feedback results, it selects the optimal training algorithm, adjusts training parameters, triggers incremental training or full update training, and optimizes the model's feature extraction network and decision logic. The labeled data that has passed algorithm verification is the labeled data output by the ground processing center and verified by the labeling algorithm feedback module.

[0037] Specifically, the dynamic training strategy library pre-configures data volume thresholds and task type matching rules. In response to the addition of flight tasks or the accumulation of collected data to a preset threshold, it automatically determines the training mode—incremental training is triggered when the data volume accumulates to a preset incremental threshold, and full update training is triggered when the data volume accumulates to a preset full update threshold or when a new data type is added. The optimization results of the model adaptive training engine are synchronized to the UAV edge and the ground processing center, achieving unified update and deployment with consistent model versions and synchronous effects on both.

[0038] The application feedback module is configured to deploy the iteratively optimized model, automatically analyze the preprocessed real-time image data, and output structured application results that meet public service needs (such as surveying and mapping data reports, urban governance anomaly early warning information, and environmental monitoring analysis conclusions). Simultaneously, it extracts error data and user adjustment records during the application process, forming model optimization feedback information that is fed back to the model adaptive training engine and synchronized to the flight route planning engine. The structured application results output by the application feedback module do not require full manual verification; only minor adjustments and reviews are needed for results in key scenarios.

[0039] The multimodal AI model for the route planning engine adopts a finely tuned version based on Qwen2-VL-72B-Instruct. The input data includes text-based task requirements (such as "monitoring street vendors in urban areas"), image-based environmental data (satellite maps, real-time weather images), and numerical environmental parameters (wind speed, temperature, coordinates of no-fly zones). The dynamic decision-making algorithm adopts a fusion scheme of "reinforcement learning + rule constraints". The reward function is set as "shortest path length + highest data collection coverage + optimal environmental adaptability". The constraints include "avoiding no-fly zones, flight altitude not less than 50 meters, and flight speed not exceeding 10 m / s".

[0040] AI intelligent processing model (data preprocessing): The denoising algorithm adopts "adaptive median filtering + BM3D joint denoising", which can effectively remove Gaussian noise and salt-and-pepper noise from the image; the stitching algorithm adopts "SIFT feature matching + RANSAC robust alignment" to ensure that the panoramic image after stitching multiple frames of images has no obvious distortion; the format is standardized and unified as "JPEG image (resolution 1920×1080) + JSON annotation file (compliant with COCO format)" for easy use in subsequent model training.

[0041] Multimodal pre-annotation algorithm + AI automatic scoring model: The pre-annotation algorithm adopts the "YOLOv8x + CLIP multimodal fusion" architecture, which can simultaneously achieve target box selection and category classification (such as identifying vehicles occupying the road, illegal buildings, pollutants, etc.); the AI ​​automatic scoring model adopts the "logistic regression + feature weighting" scheme, and the scoring dimensions cover "target box accuracy, classification accuracy, occlusion handling effect, and boundary integrity". The scoring range is 0-100 points, with a preset threshold of 80 points. Annotated data with a score below this threshold is judged as low-quality data and triggers a manual review process.

[0042] The model adaptive training engine features a dynamic training strategy library: it includes three training algorithms—ResNet50 transfer learning, VisionTransformer incremental training, and Faster R-CNN full update—which can be adaptively selected based on data characteristics. The training parameter adjustment logic is as follows: when the data volume is less than 10,000 images, the learning rate is set to 0.001, and the number of iterations is 100 rounds; when the data volume is greater than or equal to 10,000 images, the learning rate decays to 0.0001, and the number of iterations is 200 rounds. The training mode trigger conditions are: the incremental training threshold is "adding ≥5,000 new labeled data images", and the full update threshold is "adding ≥20,000 new labeled data images or adding ≥3 new target categories" (e.g., adding "tents" or "mobile vendors" to the original "vehicles" monitoring).

[0043] Flight parameters: The optimal flight path is "circular cruise 5 meters inward from the boundary of the target area". The waypoint coordinates adopt the WGS84 coordinate system (example: starting point longitude 116.3974°, latitude 39.9088°, ending point longitude 116.3956°, latitude 39.9075°), flight speed 8m / s, flight altitude 50 meters, shooting frame rate 15fps, camera depression angle 30°;

[0044] Data volume thresholds: The incremental training threshold is set to 5,000 labeled images, and the full update threshold is set to 20,000 labeled images. The size of a single image is limited to ≤5MB to avoid storage and transmission pressure.

[0045] Manual review ratio: The review ratio for abnormal data sampling is 3%-5% of the total labeled data; the manual adjustment ratio for low-quality labeled data (score < 80 points) is 100%; and the review ratio for results in key scenarios (such as intersections, school gates, and pollution-sensitive areas) is 10%.

[0046] A data processing method for low-altitude economic drones based on AI includes the following steps:

[0047] S1, Intelligent Route Planning and Dynamic Adaptation: Based on a multimodal AI large model, the route planning engine takes into account the task requirements and real-time environmental data of public service scenarios (such as surveying, urban governance, digital villages, and environmental monitoring), and autonomously generates the optimal flight path, waypoint coordinates, and flight parameters through the built-in dynamic decision-making algorithm to control the drone to take off, cruise, and land automatically. The route planning engine does not require manual preset of key parameters and can adapt to complex environments.

[0048] S2. Automated data acquisition, processing, and annotation: Following the flight path planned in S1, the drone collects scene image data via image sensors and transmits it back to the ground processing center in real time. The ground processing center performs data preprocessing using an AI intelligent processing model and automatically annotates image targets using a multimodal pre-annotation algorithm. A small number of abnormal data points are manually sampled and reviewed. The review results are fed back to the annotation algorithm module to optimize the annotation logic. Simultaneously, the drone's acquisition parameters (such as shooting resolution and frame rate) are adjusted based on the preprocessed image quality data. Data preprocessing includes noise reduction, stitching, and format standardization. The data acquisition and processing processes are linked through a collaborative control interface. Specifically, a collaborative model of "automatic annotation + AI automatic scoring + manual targeted adjustment" is adopted. AI automatic scoring filters out low-quality annotated data, and manual processing is only performed on low-scoring data, reducing the workload of manual intervention.

[0049] S3. Automated iterative training of the model based on job data: The model adaptive training engine automatically accesses the verified labeled data output by S2. Based on the data type, data volume changes, and application feedback results, it selects the optimal training algorithm from the dynamic training strategy library, adjusts the training parameters, and triggers incremental training or full update training to optimize the model using the new data. The model adaptive training engine responds to the data volume accumulation and automatically selects incremental training (when the data volume accumulates to the preset incremental threshold) or full update training (when the data volume accumulates to the preset full update threshold or when new data types are added). The optimized feature extraction network and decision logic are synchronously applied to subsequent data processing and scenario application stages.

[0050] S4, Model Application and Closed-Loop Feedback: The model optimized in S3 is deployed at the edge of the UAV or a ground processing center. It automatically analyzes the real-time image data preprocessed in S2 and outputs structured application results (such as surveying data reports, urban governance anomaly warnings, environmental monitoring analysis results, etc.). Error data and user adjustment records during the application process are extracted to form feedback information, which is then sent back to the model adaptive training engine in S3 and the flight path planning engine in S1, completing the entire closed-loop process. Specifically, the feedback information includes model optimization feedback information and flight path optimization data support information, used to optimize the model training direction and subsequent flight path planning for similar tasks, respectively. The structured application results only require minor manual adjustments and reviews of key scenario results, without the need for comprehensive verification.

[0051] Public service scenario cases

[0052] S1 (Intelligent Route Planning and Dynamic Adaptation): Input task requirements: "Monitor street vendor activities on XX section (2km long) in Hanbin District, Ankang City", real-time environmental data: "Wind speed level 3, temperature 25℃, no no-fly zone, visibility 10km"; The route planning engine does not require manual parameter presets and generates the optimal path through a dynamic decision-making algorithm: "Start point A (109.0215°, 32.6826°) → bidirectional cruise along the road section → End point B (109.0087°, 32.6792°)", and simultaneously outputs flight parameters: "flying altitude 50m, speed 8m / s, frame rate 15fps, camera depression angle 30°". Since the environmental data meets the flight conditions, no manual adjustment is required.

[0053] S2 (Automated Data Acquisition, Processing, and Labeling): The drone collects image data (1200 4K images) along a planned flight path and transmits it back to the ground processing center in real time via a 5G module. The ground processing center first performs preprocessing using "adaptive median filtering + SIFT stitching" to generate one panoramic image of the road segment and 1200 local detail images. The image sharpness after preprocessing is ≥0.8. Then, using the "YOLOv8x + CLIP" multimodal pre-labeling algorithm, 32 "suspected vehicles obstructing the road" and 15 "street vendors" are automatically selected. AI... After automatic scoring, 8 images received scores below 80 due to severe target occlusion. Manual adjustments were made to these 8 images (correcting 3 target bounding boxes and supplementing 2 classifications). The review results were fed back in real time through the annotation algorithm feedback module to optimize the "occluded target recognition logic" of the pre-annotation algorithm. At the same time, the ground processing center detected 10 blurry images (clarity <0.8) and sent parameter adjustment instructions to the UAV through the collaborative control interface to increase the shooting frame rate to 20fps and adjust the exposure time to 1 / 500s.

[0054] S3 (Automated Iterative Model Training Based on Operational Data): The model adaptive training engine automatically accesses 1198 verified and qualified labeled data (including 8 manually adjusted data), combined with 18000 historical cumulative data. Since the full update threshold (20000 data) has not been reached, incremental training is triggered, using the Vision Transformer algorithm with a learning rate of 0.0001 and 150 iterations. After training, the model's accuracy in identifying "street vending" targets has increased from 86% to 93%. The optimized model version is uniformly "V2.1" and is simultaneously deployed to the drone edge and ground processing center to ensure consistent data processing standards at both ends.

[0055] S4 (Model Application and Closed-Loop Feedback): After deploying the V2.1 model, the real-time collected image data is automatically analyzed, and the structured results are output: "28 vehicles occupying the road (including license plate information: Shaanxi GXXXX, etc.), 13 street vendors (including location information: 100m from XX section, etc.)". The urban management department staff only review the results for 3 key scenarios (intersections and school gates) and correct 1 misjudged "non-occupying vehicle". The application feedback module extracts the error data "misjudgment rate 0.8%" and user adjustment records to form feedback information: "The model's recognition accuracy for vehicles obstructed at intersections needs to be improved. It is recommended to optimize the local feature extraction logic. The flight path can be adjusted to an intersection flight height of 45m and a shooting frame rate of 25fps". This information is sent back to the model adaptive training engine as the optimization direction for the next training, and synchronized to the flight path planning engine to provide flight path optimization support for subsequent monitoring of similar road sections, completing the entire closed-loop process.

[0056] Technical effect verification data

[0057] Human intervention ratio: The solution of this invention requires only 3%-5% human intervention (sampling only low-quality labeled data and key scenario results), while existing technologies require 40%-60% human intervention (full screening of labeled data + comprehensive verification of application results), reducing the human intervention ratio by 87.5%-92.5%;

[0058] Model recognition accuracy: When the cumulative labeled data reaches 20,000 images, the model recognition accuracy of the present invention reaches 93%, while the accuracy of the existing fixed model (without iterative training) is only 82%, representing an improvement of 13.4%.

[0059] Route planning adaptability: In complex environments (such as winds of level 3-6, building obstruction, and visibility of 5-8km), the route planning adaptability of this invention reaches 95%, while existing technologies require manual adjustment for adaptation, with an adaptability rate of only 70%, representing an improvement of 35.7%;

[0060] Operational efficiency: For the task of monitoring street vendors occupying a 2km section of road, the solution of this invention takes 30 minutes (including the entire process of data collection, processing, and analysis), while the existing technology takes 60 minutes for the same task (including manual intervention), thus improving operational efficiency by 50%.

[0061] Annotation efficiency: For annotating 10,000 images, the solution of this invention takes 2 hours (automatic annotation + a small amount of manual adjustment), while the existing technology mainly uses manual annotation and takes 8 hours, thus improving the annotation efficiency by 75%.

[0062] Abnormal situation handling

[0063] Data transmission interruption: If the wireless transmission is interrupted for more than 10 seconds, the drone will automatically switch to the WiFi 6 backup link, and at the same time pause data collection and hover and wait; if the interruption is interrupted for more than 30 seconds, the drone will return to the starting point according to the preset return path, and the collected data will be temporarily stored in the local SD card and automatically retransmitted after the link is restored.

[0064] Excessive low-quality data: If the proportion of low-quality labeled data in a single batch of collected data exceeds 20%, the ground processing center will automatically trigger a second data collection by the UAV (only for the flight segment corresponding to the low-quality data), and at the same time dynamically adjust the "occlusion target recognition weight" and "feature matching threshold" of the pre-labeling algorithm to reduce the proportion of low-quality data in the future.

[0065] Model training failure: If the loss function does not decrease after 50 iterations during training (loss value > 0.5), the model adaptive training engine will automatically switch the training algorithm (e.g., from Vision Transformer to ResNet50), re-trigger training, and record the reason for failure (e.g., uneven data distribution, unreasonable parameter settings). The system will also send a prompt message to the user through the ground processing center, supporting manual adjustment of training parameters.

[0066] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data processing system for low-altitude economic unmanned aerial vehicles based on AI, characterized in that, include: The route planning engine is built on a multimodal AI large model. It is configured to take into account the task requirements and real-time environmental data of public service scenarios. Through the built-in dynamic decision-making algorithm, it can autonomously generate the optimal flight path, waypoint coordinates and flight parameters, control the drone to take off, cruise and land automatically, and receive feedback information from scenario applications for subsequent route optimization. The drone, equipped with an image sensor and a wireless transmission module, is configured to collect scene image data according to the route generated by the route planning engine, and transmit the collected data back to the ground processing center in real time through the wireless transmission module, and can adjust the collection parameters according to the data processing feedback. The ground processing center is equipped with an AI intelligent processing model, a multimodal pre-annotation algorithm module, and an annotation algorithm feedback module. It is configured to perform preprocessing operations such as denoising, stitching, and format standardization on the returned data through the AI ​​intelligent processing model; automatically annotate image targets through the multimodal pre-annotation algorithm module; receive manual inspection and review results and optimize the annotation logic through the annotation algorithm feedback module; and link with the drone's acquisition module through a collaborative control interface. The multimodal pre-annotation algorithm module, in conjunction with an AI automatic scoring model, accurately filters low-quality annotated data, requiring manual adjustments and reviews only for low-scoring data, eliminating the need to individually screen and correct all annotated data. The model adaptive training engine has a built-in dynamic training strategy library. It is configured to automatically access labeled data that has passed algorithm verification, select the optimal training algorithm and adjust training parameters based on data type, data volume changes, and application feedback results, triggering incremental training or full update training to optimize the model's feature extraction network and decision logic. The labeled data that has passed algorithm verification is the data output from the ground processing center and verified by the labeling algorithm feedback module. The dynamic training strategy library is pre-configured with data volume thresholds and task type matching rules. In response to new flight tasks or data accumulation reaching preset thresholds, it automatically determines the training mode—incremental training is triggered when data accumulates to a preset incremental threshold, and full update training is triggered when data accumulates to a preset full update threshold or when a new data type is added. The optimization results of the model adaptive training engine are synchronized to the UAV edge and the ground processing center, achieving unified update deployment with consistent model versions and synchronous effects on both. The application feedback module is configured to deploy the iteratively optimized model, automatically analyze the preprocessed real-time image data and output structured application results, and extract error data and user adjustment records during the application process to form model optimization feedback information that is sent back to the model adaptive training engine. The feedback information is also synchronized to the route planning engine.

2. The AI-based low-altitude economic unmanned aerial vehicle (UAV) data processing system according to claim 1, characterized in that: The flight path planning engine does not require manual preset of key parameters, can adapt to complex environments, and does not require manual intervention to adjust flight parameters; the image sensor of the UAV is a high-definition camera, and the wireless transmission module supports real-time transmission of collected data; the annotation algorithm feedback module feeds back the results of manual spot checks to the multimodal pre-annotation algorithm module in real time, dynamically optimizing the annotation logic.

3. The AI-based low-altitude economic drone data processing system according to claim 1, characterized in that: The structured application results output by the application feedback module do not require full manual verification; only minor adjustments and reviews are needed for the results in key scenarios.

4. An AI-based data processing method for low-altitude economic unmanned aerial vehicles, applied to the system described in any one of claims 1-3, characterized in that, Includes the following steps: S1, Intelligent Route Planning and Dynamic Adaptation: Based on a multimodal AI large model, the route planning engine takes into account the task requirements of public service scenarios and real-time environmental data, and autonomously generates the optimal flight path, waypoint coordinates and flight parameters through the built-in dynamic decision-making algorithm to control the drone to take off, cruise and land automatically. S2. Automated data acquisition, processing and annotation: The UAV collects scene image data through image sensors according to the route planned in S1 and transmits it back to the ground processing center in real time. The ground processing center completes data preprocessing through AI intelligent processing model, realizes automatic annotation of image targets through multimodal pre-annotation algorithm, and manually checks and reviews a very small number of abnormal data. The review results are fed back to the annotation algorithm module to optimize the annotation logic. At the same time, the UAV acquisition parameters are adjusted according to the preprocessed image quality data. S3. Automated model iterative training based on job data: The model adaptive training engine automatically accesses the verified labeled data output by S2. Based on the data type, data volume changes and application feedback results, it selects the optimal training algorithm from the dynamic training strategy library, adjusts the training parameters, triggers incremental training or full update training, and optimizes the model using the newly added data. S4, Model Application and Closed-Loop Feedback: The model optimized by S3 is deployed at the edge of the UAV or the ground processing center. The real-time image data preprocessed by S2 is automatically analyzed and structured application results are output. Error data and user adjustment records in the application process are extracted to form feedback information, which is sent back to the model adaptive training engine of S3 and the flight path planning engine of S1 to complete the closed loop of the whole process.

5. The AI-based low-altitude economic drone data processing method according to claim 4, characterized in that: The route planning engine described in step S1 does not require manual preset of key parameters and can adapt to complex environments; the data preprocessing described in step S2 includes noise reduction, splicing, and format standardization operations, and the data acquisition and processing processes are linked through a collaborative control interface.

6. The AI-based low-altitude economic drone data processing method according to claim 4, characterized in that: Step S2 adopts a collaborative mode of "automatic annotation + AI automatic scoring + manual targeted adjustment". Low-quality annotated data is filtered out by AI automatic scoring, and manual processing is only used for low-scoring data, reducing the workload of manual intervention.

7. The AI-based low-altitude economic drone data processing method according to claim 4, characterized in that: In step S3, the model adaptive training engine responds to the accumulation of data and automatically selects incremental training or full update training mode. The optimized feature extraction network and decision logic are synchronously applied to subsequent data processing and scenario application stages.

8. The AI-based low-altitude economic drone data processing method according to claim 4, characterized in that: The feedback information mentioned in step S4 includes model optimization feedback information and route optimization data support information, which are used to optimize the model training direction and route planning for subsequent similar tasks, respectively; the structured application results only require manual review and adjustment of key scenario results, without full verification.

Images