Unmanned aerial vehicle traffic data acquisition method based on dynamic privacy protection and ground-air cooperation

Abstract

The invention relates to the technical field of intelligent traffic data acquisition and privacy protection, provides an unmanned aerial vehicle traffic data acquisition method and system based on dynamic privacy protection and ground-air cooperation, and effectively solves the problems of high privacy leakage risk, incomplete complex scene coverage, weak cross-regional model generalization ability and the like in traditional traffic data acquisition. The method is suitable for traffic management, logistics optimization, academic research and other multi-element scenes.

Description

Technical Field

[0001] The present invention relates to the field of intelligent traffic data collection and privacy protection technology, and specifically to a method and system for collecting drone traffic data that integrates dynamic privacy masking, data localization federated learning, and ground-air collaborative virtual and real dual domains. The method is suitable for high-security, high-coverage, and high-precision data collection requirements in scenarios such as traffic management, logistics optimization, and academic research. Background Art

[0002] With the widespread use of drones in traffic monitoring, traditional data collection methods face the following technical bottlenecks:

[0003] Privacy leakage risk: Videos collected by drones contain sensitive information such as faces and license plates. Directly uploading or storing raw data can easily lead to personal information leakage.

[0004] Incomplete scene coverage: A single drone’s perspective has blind spots (such as no-fly zones and obstructed areas), and it is difficult to cover complex environments (such as nighttime, rain and fog), resulting in data loss.

[0005] Low data utilization efficiency: The traditional "centralized cloud training" model requires the aggregation of raw data from multiple regions, which has problems such as high transmission costs and high risk of privacy leakage, and the cross-regional model generalization ability is insufficient.

[0006] Although existing technologies (such as YOLO target detection and static blur desensitization) can partially solve the detection problem, they cannot simultaneously meet the multi-objective requirements of "privacy protection-data availability-scene coverage", and a comprehensive solution is urgently needed. Summary of the Invention

[0007] In order to solve the above technical problems, the present invention proposes a method and system for collecting drone traffic data based on dynamic privacy protection and ground-air collaboration. Through the technical chain of "dynamic privacy mask dual-track collection-data localization federated learning-ground-air collaborative virtual and real dual-domain expansion", the problems of privacy leakage, incomplete scene coverage, and weak cross-regional model generalization capabilities in the existing technology are solved. Through technological integration, a multi-objective balance of "privacy protection-data availability-scene coverage" is achieved. In practical applications, it can support real-time perception of traffic conditions, logistics route optimization, accident emergency response and other services, providing a new technical solution with high security, high coverage and high generalization for the refined management of smart transportation. The present invention specifically adopts the following technical solutions:

[0008] A method for collecting drone traffic data based on dynamic privacy protection and ground-air collaboration includes the following steps:

[0009] (1) Dynamic privacy masking and dual-track collection of key traffic data: A lightweight target detection model is used to identify sensitive areas (faces, license plates) and common targets (vehicles, pedestrians) in drone videos in real time. The fuzzy parameters are dynamically adjusted based on environmental parameters and business needs to adaptively blur sensitive areas while extracting key traffic data (trajectory, vehicle type, lane occupancy) of common targets.

[0010] (2) Data localization and privacy-enhanced federated learning: Each regional edge node trains the traffic prediction model based on local desensitized data, only uploading the encrypted model parameters to the cloud, and generating a cross-regional global model through federated aggregation and differential privacy technology;

[0011] (3) Ground-air collaboration and virtual-real dual-domain data collection: Integrate drone videos, roadside sensors, and vehicle-mounted terminal data, complete privacy protection fusion through edge multi-party secure computing (MPC), and combine digital twins with dynamic adversarial generation networks (DAGN) to expand missing scene data such as no-fly zones.

[0012] Preferably, the "dynamic privacy mask and dual-track collection of key traffic data" in step (1) specifically includes:

[0013] (1.1) Use the improved YOLOv8-MobilenetV4 lightweight model to perform target detection on drone video frames and output target categories (faces, license plates, vehicles, pedestrians) and bounding box coordinates;

[0014] (1.2) Obtain environmental parameters (light intensity L, rain and fog level R) and business requirement parameters (such as S=1 around the school), and calculate the fuzzy parameters using the following formula:

[0015] k=k0·α·β·γ

[0016] σ=σ0·α·β·γ

[0017] Among them, k0 and σ0 are the basic kernel size and standard deviation, α is the category weight (face 1.2, license plate 1.0), β is the environmental correction factor β = 1-0.1R + 0.05L, and γ is the business demand factor (1.5 around the school, 1.0 in the general area); (1.3) Gaussian blur or mosaic processing is applied to sensitive areas (face, license plate), and key data such as trajectory, vehicle type, and lane occupancy are extracted from general target areas through optical flow method and high-precision map matching, and stored in a structured format D = {T,C m ,O l}.

[0018] Preferably, the “data localization + privacy-enhanced federated learning” in step (2) specifically includes:

[0019] (2.1) Each regional edge node is based on local desensitized data Di Train the LSTM prediction model M i , the loss function is:

[0020]

[0021] Among them, P i,j is the predicted value of congestion probability, is the real congestion label, R i,j is the accident risk prediction value, is the real accident label, λ is the balance coefficient;

[0022] (2.2) The edge node uses homomorphic encryption (such as Paillier algorithm) to encrypt the model parameters θ i Encryption Generation θ i' =Enc(θ i ), upload to the cloud;

[0023] (2.3) The cloud aggregates parameters through the FedAvg algorithm to generate a global model θ global , the formula is: Among them, N i is the amount of data in region i, N total =∑N i ;

[0024] (2.4) Before uploading the parameters, the gradient g i Add Laplace noise g i , the formula is:

[0025] Where △g is the gradient sensitivity and ∈ is the dynamically adjusted privacy budget.

[0026] Preferably, the "ground-air collaboration + virtual-real dual-domain data collection" in step (3) specifically includes:

[0027] (3.1) Build a ground-air perception network of “UAV + roadside sensor + vehicle terminal” to collect high-altitude video V u , radar point cloud T r 、Sudden braking event B o ;

[0028] (3.2) At the edge node, the encrypted V u 、T r 、B o Perform privacy protection fusion and output accident correlation features F=v min ,d min ,B count (minimum relative speed, minimum vehicle distance, number of emergency brakes);

[0029] (3.3) Build a digital twin system for urban transportation based on OpenStreetMap and generate virtual data V of no-fly zones through the dynamic adversarial generative network (DAGN) sim , the adversarial training loss function of the generator G and the discriminator D is:

[0030]

[0031] Where x is the real data and z is random noise. The generated virtual data is mixed with the real data in proportion ρ = 0.3 to expand the training set D total =D real ∪ρ·D sim .

[0032] A drone traffic data collection system for implementing the method includes: a data collection module: composed of drones, roadside sensors, and vehicle-mounted terminals, for acquiring high-altitude video, radar point clouds, and vehicle-mounted event data; a privacy processing module: integrating a YOLOv8-MobilenetV4 model and a dynamic fuzzy parameter calculation unit, for sensitive area identification, fuzzy processing, and key data extraction; a federated learning module: composed of edge nodes (local model training unit, parameter encryption unit) and a cloud (parameter aggregation unit, differential privacy enhancement unit), for cross-regional model collaborative training;

[0033] Virtual-reality expansion module: includes a digital twin system and a dynamic adversarial generative network (DAGN), which is used to generate and fuse missing scene data such as no-fly zones.

[0034] Preferably, the privacy processing module includes: a target detection submodule: using the YOLOv8-MobilenetV4 lightweight model to output the target category and bounding box coordinates; a fuzzy parameter calculation submodule: based on the environmental parameters L, R and business requirements S, the fuzzy parameters are calculated through the formulas k=k0·α·β·γ and σ=σ0·α·β·γ; a key data extraction submodule: through the optical flow method and high-precision map matching, structured data such as trajectory, vehicle model, lane occupancy, etc. are extracted.

[0035] Preferably, the federated learning module includes: an edge training unit: training the LSTM model based on local desensitized data, supporting homomorphic encryption parameter upload; a cloud aggregation unit: aggregating encryption parameters through the FedAvg algorithm, combining differential privacy technology to generate a global model; a model distribution unit: distributing the global model to each edge node, supporting dynamic model updates.

[0036] Preferably, the virtual-reality expansion module includes: a digital twin submodule: constructing a virtual urban traffic scene based on OpenStreetMap; a DAGN generation submodule: generating virtual data of the no-fly zone through adversarial training of the generator (G) and the discriminator (D); a data fusion submodule: mixing virtual data with real data in a ratio of ρ=0.3 to expand the training set.

[0037] The present invention has the following beneficial effects:

[0038] (1) The dynamic privacy mask and key data dual-track collection technology of the present invention uses a lightweight target detection model to identify sensitive information (such as faces and license plates) and key traffic data (such as vehicle trajectories and lane occupancy) in real time, and dynamically adjusts the fuzzy strategy based on the environment and business needs, thereby strictly protecting privacy while retaining the value of data analysis.

[0039] (2) Data localization federated learning mechanism: each region only uploads encrypted model parameters to the cloud for collaborative training, avoiding the leakage of original data and significantly improving the accuracy of cross-regional traffic prediction.

[0040] (3) Ground-air coordination and virtual-real dual-domain data collection, integrating data from drones, roadside sensors, and vehicle-mounted terminals, and combining digital twin technology to generate virtual data for missing scenarios such as no-fly zones, thus filling in the coverage blind spots of traditional single equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 System architecture diagram.

[0042] Figure 2 Motion blur flowchart.

[0043] Figure 3 Diagram of the federated learning structure. DETAILED DESCRIPTION

[0044] The technical solution of the present invention includes the following core steps (such as Figure 1 shown):

[0045] Step 1: Dynamic privacy mask and dual-track collection of key traffic data

[0046] Purpose: To protect sensitive information (face, license plate) while retaining key data required for traffic analysis (such as vehicle trajectory and lane occupancy).

[0047] 1.1 Multi-target detection and sensitive area positioning:

[0048] The improved YOLOv8-MobilenetV4 lightweight model is used to perform real-time detection on drone video frames and output target categories (pedestrians, vehicles, license plates, faces) and bounding box coordinates. The model structure is:

[0049] YOLOv8-MobilenetV4(x)=Backbone(x)+Neck(x)+Head(x)

[0050] Among them, Backbone uses the depth-wise separable convolution of MobilenetV4 to reduce the amount of computation, Neck is the SPPF (Spatial Pyramid Pooling-Fast) multi-scale feature fusion module, and Head is the decoupled detection head, which outputs the target category confidence C, bounding box coordinates (x1, y1, x2, y2) and key point coordinates (such as facial features).

[0051] 1.2 Dynamic Privacy Mask Strategy:

[0052] According to the target category (face / license plate), environmental parameters (light intensity L, rain and fog level R) and business requirements (such as S=1 around the school), the fuzzy parameters (Gaussian kernel size k, standard deviation σ) are dynamically adjusted, such as Figure 2 As shown. The calculation formula for blur intensity is:

[0053] k=k0·α·β·γ, (1)

[0054] σ=σ0·α·β·γ, (2)

[0055] Where: γ and σ0 are the base kernel size and standard deviation (e.g., k0 = 51, σ0 = 1.5); α is the category weight (α = 1.2 for faces, α = 1.0 for license plates); β is the environment correction factor (β = 1-0.1R + 0.05L, R∈[0,3], L∈[0,1]);

[0056] γ is the business demand factor (γ=1.5 in school surroundings, γ=1.0 in general areas).

[0057] 1.3 Extraction of key traffic data:

[0058] For non-sensitive areas (such as vehicle outlines and pedestrian trajectories), the displacement △(x, y) between consecutive frames is calculated by the optical flow method Lucas-Kanade (I1, I2), and the trajectory is extracted by combining the target detection results. Model C m , Lane Occupancy O l (matching with high-precision maps) and other key data are stored in a structured format D = {T,C m ,O l}.

[0059] Step 2: Data Localization + Privacy-Enhanced Federated Learning

[0060] Purpose: To improve the accuracy of cross-regional traffic prediction by coordinating multi-regional models without transmitting original data.

[0061] 2.1 Local model training:

[0062] Each regional edge node (such as roadside unit) is based on local desensitized data D i (Only contains key data T,C m ,O l ), train the LSTM prediction model M i , output the congestion probability P in the next 15 minutes i,j or accident risk R i,j The model loss function is:

[0063]

[0064] in, is the true label, and λ is the balance coefficient λ = 0.3.

[0065] 2.2 Encrypted parameter upload and aggregation:

[0066] Each edge node uses homomorphic encryption (such as Paillier algorithm) to encrypt the model parameters θ i Encryption, generate θ i' =Enc(θ i ), uploaded to the cloud. The cloud uses the FedAvg algorithm to aggregate parameters, such as Figure 3 As shown:

[0067]

[0068] Among them, N i is the amount of data in region i, N total =∑N i .

[0069] 2.3 Differential Privacy Enhancement:

[0070] Before uploading the parameters, the gradient g i Add Laplace noise ∈ (privacy budget) to ensure:

[0071]

[0072] Here, Δg is the gradient sensitivity Δg=1, ∈ is dynamically adjusted (e.g., ∈=0.5 during busy periods and ∈=1.0 during idle periods).

[0073] Step 3: Ground-air collaboration + virtual-real dual-domain data collection mechanism

[0074] Purpose: To complete data on blind spots such as no-fly zones and improve scene coverage capabilities.

[0075] 3.1 Collaborative collection of multi-source ground and air data:

[0076] Constructing a ground-to-air perception network of "UAV (high-altitude video) + roadside sensor (radar point cloud) + onboard terminal (OBU)": UAV: ​​Collect high-altitude video V u , output desensitized key data D u ; Roadside radar: output vehicle trajectory T r , speed v r ; On-board OBU: output emergency brake event B o , steering angle θ o .

[0077] 3.2 Edge Multi-Party Secure Computing (MPC) Fusion:

[0078] At the edge node (such as the roadside unit), the GMW protocol is used to perform privacy protection fusion of multi-source data: Input: Encrypted D u’ T r’ B o ; Output: F = v after fusion of accident correlation features min ,d min ,B count (Minimum relative speed, minimum vehicle distance, number of emergency brakes).

[0079] 3.3 Virtual and Real Dual Domain Scenario Expansion:

[0080] Build a digital twin system for urban traffic based on OpenStreetMap, and generate virtual data V of no-fly zones through dynamic adversarial generative network (DAGN). sim DAGN's generator G and discriminator D are trained against each other, and the loss function is:

[0081]

[0082] Where x is the real data and z is random noise. The generated virtual data is mixed with the real data in proportion ρ = 0.3 to expand the training set D total =D real ∪ρ·D sim .

[0083] Compared with the prior art, the present invention has the following advantages:

[0084] Balancing privacy protection and data availability: Dynamic privacy masking technology uses adaptive fuzzy parameter adjustment (Formula 1-2) to achieve a face / license plate desensitization rate of ≥99% while retaining ≥95% of key data (such as trajectory continuity and vehicle model recognition accuracy), meeting the requirements of the Personal Information Protection Law.

[0085] Improved cross-region model generalization: Data-localized federated learning (Formula 4-5) avoids uploading raw data to the cloud. This improves cross-region congestion prediction accuracy by 15% compared to the centralized training model (mAP increased from 78% to 93%), with a privacy attack success rate of ≤5%.

[0086] Enhanced scene coverage capability: Ground-air collaboration + virtual-real dual-domain mechanism (Formula 8-9) makes the no-fly zone scene completion rate ≥ 90%, the data missing rate in complex environments (rain, fog, night) is reduced from 25% to 5%, and the accident warning accuracy rate is ≥ 85%.

[0087] The following describes the implementation process of the present invention in conjunction with a specific scenario (traffic monitoring in western Beijing):

[0088] Data collection phase:

[0089] Deploy 10 drones (as in step 3.1) to cover areas such as Haidian and Shijingshan. Use the YOLOv8-MobilenetV4 model to detect pedestrians and vehicles in real time (step 1.1), apply motion blur to faces and license plates (formula 1-2), and extract key data such as trajectory and vehicle model (step 1.3).

[0090] Data fusion and model training:

[0091] The roadside unit fuses drone, radar, and OBU data through the MPC protocol (step 3.2) to generate accident-related features. Each regional edge node trains an LSTM model based on local data (step 2.1), encrypts and uploads parameters to the cloud for aggregation (step 2.2), and combines differential privacy enhancement (step 2.3).

[0092] Scenario expansion and verification:

[0093] A digital twin system was built based on OpenStreetMap. Virtual data of no-fly zones was generated through DAGN (step 3.3). After mixing with real data, the model was trained to verify the no-fly zone scene completion rate and accident warning accuracy.

Claims

1. A method for collecting UAV traffic data based on dynamic privacy protection and ground-air collaboration, characterized in that: The following steps are involved: (1) Dynamic privacy masking and dual-track collection of key traffic data: A lightweight target detection model is used to identify sensitive areas and common targets in drone videos in real time. Blurring parameters are dynamically adjusted based on environmental parameters and business needs to adaptively blur sensitive areas while extracting key traffic data from common targets. (2) Data localization and privacy-enhanced federated learning: Each regional edge node trains the traffic prediction model based on local desensitized data, only uploading the encrypted model parameters to the cloud, and generating a cross-regional global model through federated aggregation and differential privacy technology; (3) Ground-air collaboration and virtual-real dual-domain data collection: Integrate drone videos, roadside sensors, and vehicle-mounted terminal data, complete privacy protection fusion through edge multi-party secure computing (MPC), and combine digital twins with dynamic adversarial generation networks (DAGN) to expand missing scene data such as no-fly zones.

2. The method according to claim 1, characterized in that The dual-track collection of dynamic privacy mask and key traffic data in step (1) specifically includes: (1.1) Use the improved YOLOv8-MobilenetV4 lightweight model to detect targets in drone video frames and output target categories and bounding box coordinates; (1.2) Obtain environmental parameters and business requirement parameters, and calculate fuzzy parameters using the following formula: k=k0·α·β·γ σ=σ0·α·β·γ Among them, k0 and σ0 are the basic kernel size and standard deviation, α is the category weight, β is the environmental correction factor β = 1-0.1R + 0.05L, and γ is the business demand factor; (1.3) Gaussian blur or mosaic processing is applied to sensitive areas, and key data such as trajectory, vehicle type, lane occupancy, etc. are extracted from common target areas through optical flow method and high-precision map matching, and stored in a structured format D = {T, C m ,O l }.

3. The method according to claim 1, characterized in that The data localization and privacy-enhanced federated learning described in step (2) specifically includes: (2.1) Each regional edge node is based on local desensitized data D i Train the LSTM prediction model M i , the loss function is: Among them, P i,j is the predicted value of congestion probability, is the real congestion label, R i,j is the accident risk prediction value, is the real accident label, λ is the balance coefficient; (2.2) The edge node uses homomorphic encryption to encrypt the model parameters θ i Encryption Generation θ i' =Enc(θ i ), upload to the cloud; (2.3) The cloud aggregates parameters through the FedAvg algorithm to generate a global model θ global , the formula is: Among them, N i is the amount of data in region i, N total =∑N i ; (2.4) Before uploading the parameters, the gradient g i Add Laplace noise g i , the formula is: Where △g is the gradient sensitivity and ∈ is the dynamically adjusted privacy budget.

4. The method according to claim 1, wherein The ground-air collaboration and virtual-real dual-domain data collection described in step (3) specifically includes: (3.1) Build a ground-air perception network of "UAV + roadside sensor + vehicle terminal" to collect high-altitude video V u , radar point cloud T r 、Sudden braking event B o ; (3.2) At the edge node, the encrypted V u 、T r 、B o Perform privacy protection fusion and output accident correlation features F=v min ,d min ,B count ; (3.3) Build a digital twin system for urban transportation based on OpenStreetMap and generate virtual data V of no-fly zones through the dynamic adversarial generative network (DAGN) sim , the adversarial training loss function of the generator G and the discriminator D is: Where x is the real data and z is random noise. The generated virtual data is mixed with the real data in proportion ρ = 0.3 to expand the training set D total =D real ∪ρ·D sim .

5. A UAV traffic data collection system for implementing the method according to any one of claims 1 to 4, characterized in that: include: Data acquisition module: composed of drones, roadside sensors, and vehicle-mounted terminals, used to obtain high-altitude video, radar point clouds, and vehicle-mounted event data; Privacy processing module: Integrates the YOLOv8-MobilenetV4 model and dynamic fuzzy parameter calculation unit for sensitive area identification, fuzzy processing, and key data extraction; Federated learning module: Consists of edge nodes and the cloud for cross-regional model collaborative training; Virtual-reality expansion module: includes a digital twin system and a dynamic adversarial generative network (DAGN), which is used to generate and fuse missing scene data such as no-fly zones.

6. The system according to claim 5, characterized in that The privacy processing module includes: a target detection submodule: using the YOLOv8-MobilenetV4 lightweight model to output target categories and bounding box coordinates; a fuzzy parameter calculation submodule: based on environmental parameters L, R and business requirements S, the fuzzy parameters are calculated through the formulas k=k0·α·β·γ and σ=σ0·α·β·γ; a key data extraction submodule: through optical flow method and high-precision map matching, structured data such as trajectory, vehicle model, lane occupancy, etc. are extracted.

7. The system according to claim 5, characterized in that The federated learning module includes: an edge training unit: training the LSTM model based on local desensitized data, supporting the upload of homomorphic encryption parameters; a cloud aggregation unit: aggregating encryption parameters through the FedAvg algorithm, combining differential privacy technology to generate a global model; a model distribution unit: distributing the global model to each edge node, supporting dynamic model updates.

8. The system according to claim 5, wherein: The virtual-reality expansion module includes: a digital twin submodule: building a virtual urban traffic scene based on OpenStreetMap; a DAGN generation submodule: generating virtual data of the no-fly zone through adversarial training of the generator (G) and the discriminator (D); and a data fusion submodule: mixing virtual data with real data in a ratio of ρ = 0.3 to expand the training set.

Images