Internet of vehicles target identification method based on personalized federal learning
Through personalized federated learning methods, category confidence and clustering technology are used to optimize the collaborative relationship of mobile vehicles in the Internet of Vehicles, the data heterogeneity problem is solved, the training performance and efficiency of the object detection model are improved, and user data privacy is protected.
Patent Information
- Application Number
- CN202510254252.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-05
- Publication Date
- 2025-07-11
AI Technical Summary
The existing machine learning methods have data heterogeneity problems in the Internet of Vehicles, resulting in poor training performance and slow convergence speed of target detection models, and the risk of privacy leakage and high transmission costs for centralized training.
Using a personalized federated learning method, through the collaboration between edge servers and mobile vehicles, the data similarity between users is reasoned using category confidence, clustering and hierarchical model parameters are updated, and the collaboration relationship is optimized, and personalized models are customized for mobile vehicles with different data distributions.
It improves the training performance of the object detection model, shortens the model convergence time, improves the processing efficiency of object detection tasks in the Internet of Vehicles scenarios, and protects user data privacy.
Smart Images
Figure CN120298656A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a vehicle networking target recognition method based on personalized federated learning, belonging to the technical field of target recognition. Background Art
[0002] With the rapid development of vehicle networking technology, the perception and communication capabilities of modern vehicle networking devices are constantly enhanced. In-vehicle intelligent devices can collect a large amount of high-quality sensing data, which is crucial for application scenarios such as traffic management and intelligent driving. Based on machine learning technology, this data can provide intelligent services for vehicle networking. However, most existing machine learning methods adopt a centralized training method, which requires uploading the data to a central server for unified training. Due to data privacy issues and the distributed characteristics of a large number of moving vehicles in the vehicle networking environment, directly uploading the data to the central server for training will face a large risk of privacy leakage and high transmission costs.
[0003] To solve this dilemma, federated learning emerges as a distributed machine learning method. Different from the traditional centralized model training process, federated learning allows each mobile device in the vehicle networking to save the local data on its own device and only share the model update parameters, which effectively protects the user's data privacy. However, due to the heterogeneity of the data of each mobile vehicle in the vehicle networking, the data during the training process exhibits non-independent and identically distributed (non-IID) characteristics, resulting in problems such as poor training performance and slow convergence speed when the target detection model performs federated learning. To address this challenge, the personalized federated learning method for vehicle networking scenarios has become a research hotspot. The personalized federated learning customizes personalized models for mobile vehicles with different data distributions to better adapt to the specific needs and data characteristics of each mobile vehicle.
[0004] Therefore, how to design a vehicle networking target recognition algorithm based on personalized federated learning is an urgent problem to be solved currently. Summary of the Invention
[0005] The technical problem to be solved by the present invention is: to provide a vehicle networking target recognition method based on personalized federated learning, by optimizing the cooperation relationship between mobile vehicles, customizing personalized models for mobile vehicles with different data distributions, so as to solve the data heterogeneity problem in vehicle networking scenarios, improve the training performance of the target detection model, shorten the model convergence time, and improve the accuracy of the model.
[0006] The present invention adopts the following technical solutions to solve the above technical problems:
[0007] A vehicle networking target recognition method based on personalized federated learning. For a vehicle networking scenario including N moving vehicles, the following steps 1 - 4 are executed to complete vehicle target recognition in the vehicle networking scenario;
[0008] Step 1: Build a personalized federated learning architecture, including an edge server and local target detection models set in each moving vehicle. The local dataset corresponding to the moving vehicle i includes the vehicle networking scenario images collected by the moving vehicle i and the corresponding class labels on the images. The public dataset corresponding to the edge server includes images with all class labels; Set the maximum global iteration number T of the edge server and the local iteration training number R of each local target detection model;
[0009] Step 2: For the first round of global iteration, the edge server distributes the pre - trained initial global parameters to each local target detection model for federated learning. After each local target detection model performs R - round gradient descent iterative training using the corresponding local dataset, it uploads the trained model parameters to the edge server;
[0010] Step 3: The edge server receives the model parameters uploaded by each moving vehicle, performs forward inference on the public dataset, calculates the class confidence of each moving vehicle, infers the similarity between each moving vehicle according to the class confidence, and performs logical clustering on all moving vehicles;
[0011] Step 4: Define the batch normalization layer of each local target detection model as the personalized layer, and the remaining layers as the non - personalized layers, that is, the model parameters uploaded in Step 2 include personalized layer parameters and non - personalized layer parameters; The edge server updates the personalized layer parameters based on the clustering results in Step 3, and updates the non - personalized layer parameters by linearly combining the non - personalized layer parameters of all moving vehicles;
[0012] Step 5: The edge server distributes the updated personalized layer parameters and non - personalized layer parameters to the local target detection models of each moving vehicle, and continues the second round of global iteration. After the moving vehicle performs R - round gradient descent iterative training, it uploads the trained model parameters to the edge server. The edge server executes Step 4 to update the model parameters. The subsequent global iterations repeat the process of the second round of global iteration. When the preset maximum global iteration number T is reached, the global iteration terminates; The model update parameters obtained in the last round of global iteration are used as the final model parameters for vehicle target recognition.
[0013] Compared with the prior art, the present invention adopts the above technical solutions and has the following technical effects:
[0014] 1. The present invention infers the data similarity between users based on class confidence, can accurately optimize the cooperation relationship between users, and further customize personalized models for mobile vehicles with different data distributions.
[0015] 2. By optimizing the cooperation and training strategies between mobile vehicles, the present invention shortens the convergence time of the model, enables the personalized model to complete training more quickly and efficiently in the vehicle networking environment, and thus improves the processing efficiency of target detection tasks in the vehicle networking scenario. Brief Description of the Drawings
[0016] Figure 1 is the federated system model diagram in the vehicle networking scenario proposed by the present invention;
[0017] Figure 2 is the flowchart of the vehicle networking target recognition method based on personalized federated learning proposed by the present invention. Detailed Embodiments
[0018] The following details the embodiments of the present invention, and the examples of the embodiments are shown in the drawings. The embodiments described below with reference to the drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0019] The present invention proposes a vehicle networking target recognition method based on personalized federated learning, applies FL to the vehicle networking environment, and uses YOLOv7 as the target detection model, and proposes a novel personalized mechanism. First, the data distribution relationship between different users is measured by class confidence, a similarity matrix between mobile vehicles is established, and a hierarchical clustering algorithm is used to cluster mobile vehicles to provide guidance for model parameter update. The model parameter update adopts hierarchical update (personalized layer, non-personalized layer), so as to solve the data heterogeneity problem in the vehicle networking scenario, improve the training performance of the target detection model, shorten the model convergence time, and improve the accuracy of the model.
[0020] The participating users of the distributed federated learning architecture in the vehicle networking of the present invention are distributed as Figure 1 shown. It is assumed that there is an edge server and N mobile vehicles in the vehicle networking scenario. The local data set of each mobile vehicle is D i ={x i ,y i}, where x i ,y i respectively represent the training sample data and the corresponding label of vehicle i. The data sets owned by different users have the characteristics of data heterogeneity, specifically manifested as long-tail distribution and uneven data distribution. A public data set D pub, the public dataset has evenly distributed class information. In the first round, the mobile users initialize the model parameters and perform R rounds of local training, then upload the trained parameters to the edge server. The edge server establishes the collaboration relationship among the mobile users, calculates and updates the personalized training model, and redistributes it to the mobile users. The mobile users continue to perform local training and uploading in the second round and subsequent rounds. The edge server updates and distributes the personalized model based on the collaboration relationship established in the first round. Repeat the global iteration process until the total number of iteration rounds is reached. The hierarchical update strategy is adopted to calculate and update the personalized training model parameters, that is, the personalized layer is calculated in the cluster, and the non-personalized layer is calculated and updated among all vehicles. The overall goal of federated training is to minimize the total training loss of all mobile vehicles and maximize the working efficiency of federated training.
[0021] The process of the vehicle networking target recognition method based on personalized federated learning proposed by the present invention is as Figure 2 shown, and includes the following steps:
[0022] Step 1: The mobile vehicle uses the local dataset to train for R rounds and uploads the trained model parameters to the edge server
[0023] In the first round of the global iteration stage, the edge server distributes the initialized global model to the selected mobile vehicles for training. The mobile vehicles adopt the YOLOV7 target detection network, and the dataset used for training is D i ={x i ,y i}}. The training loss function L i (w i ) includes three parts: bounding box loss, target detection loss, and classification loss. The overall goal during the federated learning process is:
[0024] L i (w i ) = L box (w i ) + L obj (w i ) + L cls (w i )
[0025]
[0026] Among them, L box (w i ) is the bounding box loss, L obj (w i ) is the target detection loss, L cls (w i ) is the classification loss, M i is the local dataset of vehicle i, Represents the sum of all the training data for training the mobile vehicles.
[0027] Step 2: The edge server receives the model parameters transmitted by the user, performs forward inference on the public dataset, calculates the class confidence of each user's model, and infers the similarity between users based on the class confidence
[0028] The edge server stores the model parameters uploaded in the pre-training rounds and performs forward inference on the public dataset D pub to evaluate the characteristics of the mobile vehicle data distribution, and then obtains the prediction probability for each class. These prediction probabilities reflect the model's recognition ability for different class data. Define the class confidence, which represents the average value of the prediction probabilities of the model for each class on the public dataset. The calculation is as follows:
[0029]
[0030] where, P i represents the class confidence of mobile vehicle i, M pub is the size of the public dataset, f i,n is the local model f of mobile vehicle i i The result of inferring the nth sample of the public dataset, that is, the class probability.
[0031] Using the class confidence as a measure of the client data distribution is intuitive and reasonable. The class confidence represents the level of certainty that the model has in its designated class predictions, usually expressed as a value between 0 and 1, and the higher the value, the greater the confidence. Compared with directly using the model parameters, the class confidence has the advantage that it effectively avoids the influence of permutation symmetry and the curse of dimensionality in neural networks.
[0032] After calculating the class confidence of each mobile vehicle, next, the cosine similarity between vehicles is calculated to measure the similarity of their data distributions. The calculation is as follows:
[0033]
[0034] where, P i is the class confidence obtained by vehicle i through forward inference on the public dataset M pub P j is the class confidence obtained by vehicle j through forward inference on the public dataset M pub is the class confidence obtained by vehicle j through forward inference on the public dataset M.
[0035] The similarity calculation can not only reveal which vehicles have similar data distributions, but also provide a basis for subsequent clustering, personalized learning, and grouping of mobile vehicles in federated learning.
[0036] Step 3: Construct a similarity matrix, perform logical clustering using the hierarchical classification method, and establish a collaboration relationship between vehicles
[0037] After calculating the similarity between vehicles according to Step 2, construct it into an N-dimensional symmetric similarity matrix A. Each element A of the matrix i,j represents the similarity between vehicle i and vehicle j. The closer the value of A i,j is to 1, the higher the similarity between vehicle i and vehicle j, and the more similar the data distributions are. To more intuitively establish the collaboration relationship between vehicles and perform effective model sharing and updating between vehicles, the concept of clustering is introduced. By clustering similar vehicles into the same group, it is possible to better identify moving vehicles with similar data distributions, thereby reducing the impact of data heterogeneity between different moving vehicles and improving the effect of federated learning. In this process, the hierarchical clustering algorithm is used for logical clustering. The advantage of hierarchical clustering is that it does not require a preset number of clusters, but controls the clustering accuracy by setting a distance threshold or selecting a suitable merging criterion. Hierarchical clustering can effectively capture the hierarchical structure between data, and thus generate appropriate clusters according to the similarity between vehicles. The steps of the hierarchical clustering algorithm for clustering are as follows:
[0038] (a) Initialization: Consider each moving vehicle as a separate cluster, for a total of N clusters.
[0039] (b) Calculate the distance matrix: For each pair of vehicles i and j, the distance matrix D and the similarity matrix A satisfy:
[0040] D i,j = I - A i,j
[0041] where I is generally taken as 1, such that the D i,j distance satisfies that the greater the similarity, the smaller the distance;
[0042] (c) Calculate the distance between clusters and select the merging criterion: Calculate the distance between clusters using average linkage, that is, the average of the distances between all pairs of points in the two clusters. In each aggregation step, select the two clusters with the smallest distance for merging.
[0043]
[0044] (d) Merge the two closest clusters: According to the selected merging criterion, find the two clusters with the smallest distance in the current clusters for merging, update the distance matrix, delete the two merged clusters, and add the distance between the new clusters.
[0045] (e) Repeat step (d) until the termination condition (the distance of clustering exceeds the preset threshold) is met.
[0046] In this way, vehicles with high similarity are assigned to the same group, and finally the cluster labels of each moving vehicle are obtained.
[0047] Step 4: The edge server performs a hierarchical update strategy, updates the personalized layer and the non-personalized layer with different strategies respectively, and generates a personalized model for each vehicle for the next round of training
[0048] After the end of one round of training, the parameters are uploaded to the edge server, including the personalized layer parameters and the non-personalized layer parameters. The Batch Normalization (BN) layer is defined as the personalized layer. The reason is that the BN layer can be dynamically adjusted according to the data distribution of each moving vehicle, so it can effectively cope with the data heterogeneity problem, thereby improving the adaptability and performance of the model. The remaining layers are regarded as non-personalized layers. Compared with the personalized layer, the parameters of the non-personalized layer do not have strong data dependence. The functions of these layers are mainly to perform general feature extraction or inference, so their parameters are shared among moving vehicles. The parameters of the non-personalized layer will be aggregated among different moving vehicles in a weighted summation manner, so as to ensure a certain global consistency while ensuring personalized performance.
[0049] (a) Personalized layer update: The update of the personalized layer is carried out within the cluster. Within the same cluster, the data distributions among moving vehicles are similar, so their personalized layers can share more information. To ensure the effective update of the personalized layer, the server will introduce regularization constraints and guide the update process based on the similarity matrix among moving vehicles.
[0050]
[0051] where A i,j is the similarity between moving vehicles i and j. The higher the similarity, the more consistent the update of the personalized layer. The global step size η is the parameter during the server regularization update, which controls the adjustment amplitude of the model parameters during the regularization process to achieve a smooth and stable optimization process. is the personalized layer parameter of moving vehicle j with the same cluster label as moving vehicle i. is the personalized layer parameter uploaded by moving vehicle i in the t-th round, is the personalized layer parameter of moving vehicle i for the (t + 1)-th round of training.
[0052] Through regularization adjustment, it has the following advantages:
[0053] (1) Improve generalization ability: By reducing the differences in the personalized layers among moving vehicles, regularization helps the trained model to generalize better and avoid overfitting to the local data of a certain moving vehicle.
[0054] (2)Enhance the collaborative training effect: Regularization ensures that the personalized layer updates of in-cluster moving vehicles are more consistent in data distribution, thus facilitating the improvement of the overall training efficiency and effect.
[0055] (3)Smooth convergence: By regularizing the update pace of the personalized layer, large oscillations and drastic parameter changes are avoided, making the training process smoother and more stable.
[0056] (b) Non-personalized layer update: Neural network layers other than the BN layer are defined as non-personalized layers. When the server updates the parameters of the non-personalized layer, the updated model parameters are generated by linearly combining the parameters of the non-personalized layers of all moving vehicles. To ensure that the contributions of different moving vehicles to the update of the non-personalized layer are unequal, the server constructs a weighted coefficient matrix μ to weight the updates of the non-personalized layers of each moving vehicle, represented as an N-dimensional matrix. Each weighted coefficient μ i,j depends on the similarity A between moving vehicle i and moving vehicle j i,j , to ensure that the contribution ratio of each moving vehicle is relative, the similarity matrix constructed in step 3 is normalized and expressed as:
[0057]
[0058] where μ i,j is the aggregation ratio of moving vehicle j in the non-personalized model layer of moving vehicle i, satisfying For 1 ≤ i ≤ N, represents the result after aggregation of the non-personalized layer of moving vehicle i, is the result after updating the non-personalized layers of all moving vehicles.
[0059]
[0060] where, is all non-personalized layers of moving vehicle 1 before all updates, is the first non-personalized layer of moving vehicle 1.
[0061] The non-personalized layer update method based on similarity weighting has the following advantages:
[0062] (1) It can more accurately integrate the parameters of the non-personalized layers of each moving vehicle. When the data distributions of two vehicles are similar, their updates of the non-personalized layers should contribute more prominently to each other; while clients with larger data differences contribute less.
[0063] (2) Improve the stability of the local model, reduce the interference of clients with large data differences on the update of the local non-personalized layer, and thus make the update smoother and more stable.
[0064] (3) Improve the adaptability of the model to data heterogeneity. The weighting mechanism enables vehicles with different data distributions to adjust their update contributions to the non-personalized layer according to their similarity to other vehicles, thereby making the updated model more adaptable to different types of data distributions.
[0065] The update of the personalized layer can enhance the personalization of each client model, reduce the differences between clients in the same cluster, and effectively address the data heterogeneity problem through regularization and similarity weighting. The update of the non-personalized layer ensures that the shared part of the model can more accurately reflect the characteristics of each client through similarity weighting, improving the adaptability and stability of the model. Through our update strategy, better performance improvement and optimization can be achieved based on personalized requirements by leveraging the shared information of other similar clients.
[0066] Step 5: Receive the personalized model sent by the edge server and perform iterations and updates after the second round until the preset global number of rounds is reached; otherwise, return to Step 4.
[0067] When the vehicle user conducts the next round of training, the edge server sends the update parameters of the current round to each client (i.e., the vehicle user), including the model weights of the personalized layer and the non-personalized layer. The vehicle user applies the received parameters to its local model and continues to train on its local dataset. During the training process, the YOLOV7 model will use the default fine-tuning hyperparameter settings, with a learning rate of 0.01 during the warm-up period and 0.001 during the decay period. Perform R rounds of gradient descent iterations: After the training is completed, the vehicle user sends the locally updated model parameters back to the edge server and repeats Step 4.
[0068] Based on the same inventive concept, an embodiment of this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the foregoing vehicle network target recognition method based on personalized federated learning.
[0069] Based on the same inventive concept, an embodiment of this application provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the steps of the foregoing vehicle network target recognition method based on personalized federated learning.
[0070] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.
[0071] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices produce means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.
[0072] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.
[0073] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.
[0074] The above embodiments are only for illustrating the technical idea of the present invention, and the protection scope of the present invention cannot be limited thereby. Any modification made on the basis of the technical solution according to the technical idea proposed by the present invention falls within the protection scope of the present invention.
Claims
1. A vehicle networking target recognition method based on personalized federated learning, characterized in that, For a vehicle-to-everything (V2X) scenario including N moving vehicles, perform the following steps 1 - 4 to complete vehicle target recognition in the V2X scenario; Step 1: Build a personalized federated learning architecture, including an edge server and local object detection models set in each moving vehicle. The local dataset corresponding to moving vehicle i includes V2X scenario images collected by moving vehicle i and the corresponding class labels on the images. The public dataset corresponding to the edge server includes images with all class labels; Set the maximum number of global iterations T of the edge server and the number of local iterative training rounds R of each local object detection model; Step 2: For the first round of global iteration, the edge server distributes the pre-trained initial global parameters to each local object detection model for federated learning. After each local object detection model performs R rounds of gradient descent iterative training using the corresponding local dataset, it uploads the trained model parameters to the edge server; Step 3: The edge server receives the model parameters uploaded by each moving vehicle, performs forward inference on the public dataset, calculates the class confidence of each moving vehicle, infers the similarity between each moving vehicle based on the class confidence, and performs logical clustering on all moving vehicles; Step 4: Define the batch normalization layer of each local object detection model as the personalized layer, and the remaining layers as the non-personalized layer. That is, the model parameters uploaded in Step 2 include personalized layer parameters and non-personalized layer parameters; The edge server updates the personalized layer parameters based on the clustering results in Step 3, and updates the non-personalized layer parameters by linearly combining the non-personalized layer parameters of all moving vehicles; Step 5: The edge server distributes the updated personalized layer parameters and non-personalized layer parameters to the local object detection models of each moving vehicle, and continues the second round of global iteration. After the moving vehicle performs R rounds of gradient descent iterative training, it uploads the trained model parameters to the edge server. The edge server executes Step 4 to update the model parameters. The subsequent global iterations repeat the process of the second round of global iteration. When the preset maximum number of global iterations T is reached, the global iteration terminates; Take the model update parameters obtained in the last round of global iteration as the final model parameters for vehicle target recognition.
2. The method for vehicle network target recognition based on personalized federated learning according to claim 1, wherein In Step 2, the local object detection model is the YOLOV7 object detection model, and the overall goal of federated learning is: L i (w i ) = L box (w i ) + L obj (w i ) + L cls (w i ) Among them, L i (w i ) is the application parameter w of the local target detection model corresponding to the moving vehicle i i The loss function during training, L box (w i ) is the bounding box loss, L obj (w i ) is the target detection loss, L cls (w i ) is the classification loss; M i is the local dataset of the moving vehicle i, and M is the sum of the local datasets of all moving vehicles 3. The vehicle network target recognition method based on personalized federated learning according to claim 1, wherein, The specific process of Step 3 is as follows: The edge server receives the model parameters uploaded by each moving vehicle, performs forward inference on the public dataset, and calculates the class confidence of each moving vehicle, that is: Among them, P i represents the class confidence obtained by the moving vehicle i through forward inference on the public dataset D pub and M pub represents the size of the public dataset D pub and f i,n represents the result of the moving vehicle i's local object detection model f i inferring the nth sample of the public dataset, that is, the class probability; Infer the similarity between each moving vehicle based on the class confidence: Among them, A i,j represents the similarity between moving vehicles i and j, and P j represents the class confidence obtained by forward inference of moving vehicle j on the public dataset D pub . α(·) represents the non - negative cosine similarity function, <P i ·P j > represents the inner product of P i and P j . ||P i || and ||P j || respectively represent the L2 norms of P i and P j . A similarity matrix A is constructed based on A i,j . Based on the similarity matrix A, use the hierarchical clustering algorithm to perform logical clustering on all moving vehicles. The specific process is as follows: a. Consider each moving vehicle as a separate cluster, for a total of N clusters; b. Calculate the distance D between moving vehicles i and j i,j : D i,j = I - A i,j where I takes 1; c. Use the average linkage method to calculate the distance between any two clusters: Among them, D(C p , C q ) represents the distance between cluster C p and C q , |C p | and |C q | respectively represent the number of vehicles included in cluster C p and C q ; d. Find the two clusters with the smallest distance and merge them; e. For all the merged clusters, update the distance between any two clusters in all the merged clusters based on the average linkage method, and determine whether the distance between any two clusters exceeds a preset threshold. If so, regard all the merged clusters as the final clusters; otherwise, return to step d to continue the merging until the distance between any two clusters in all the merged clusters exceeds the preset threshold.
4. The method for identifying vehicle networking targets based on personalized federated learning according to claim 3, wherein, In step 4, the update of the personalized layer parameters is as follows: Based on the clustering results of step 3, introduce regularization constraints, and update the personalized layer parameters of the local object detection models of moving vehicles with the same cluster label based on the similarity matrix between moving vehicles: Among them, represents the personalized layer parameters used for training the local target detection model of mobile vehicle i in the (t + 1)-th round of global iteration, represents the personalized layer parameters uploaded by the local target detection model of mobile vehicle i in the t-th round of global iteration, is the global step size, μ is the batch size of the local target detection model, R is the number of local iteration training rounds of the local target detection model, and η is the parameter during the regularization update of the edge server, represents the personalized layer parameters of the local target detection model of mobile vehicle j with the same cluster label C as mobile vehicle i; The update of the non-personalized layer parameters is as follows: Similarity A between moving vehicles i and j i,j Calculate the weighting coefficient μ i,j : Among them, A i,k represents the similarity between moving vehicles i and k, and satisfies According to μ i,j Construct a weighted coefficient matrix μ, and update the non-personalized layer parameters of all moving vehicles according to μ: Among them, represents the non-personalized layer parameter matrix for training all mobile vehicle local target detection models in the (t + 1)-th round of global iteration, represents the first, second, …, m-th non-personalized layer parameters uploaded by the local target detection model of mobile vehicle 1 in the t-th round of global iteration, represents the first, second, …, m-th non-personalized layer parameters uploaded by the local target detection model of mobile vehicle 2 in the t-th round of global iteration, represents the first, second, …, m-th non-personalized layer parameters uploaded by the local target detection model of mobile vehicle N in the t-th round of global iteration, where N is the number of all mobile vehicles.
5. The method for vehicle networking target recognition based on personalized federated learning according to claim 4, characterized in that, In each round of global iteration, after the local object detection model undergoes R rounds of gradient descent iteration, upload the iteratively obtained model parameters to the edge server. The gradient descent iteration formula is as follows: Among them, respectively represent the model parameters obtained from the (r + 1)-th and r-th local iterations of the local object detection model in the (t + 1)-th global iteration, represents the gradient of the loss function of vehicle i with respect to the model parameters.
6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the vehicle network object recognition method based on personalized federated learning according to any one of claims 1 to 5.
7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the vehicle network object recognition method based on personalized federated learning according to any one of claims 1 to 5.