Vehicle identification method and system based on dynamic behavior analysis and federated learning

Abstract

The invention discloses a vehicle identification method and system based on dynamic behavior analysis and federated learning, the method is applied to local ends in one-to-one correspondence with pre-detection areas, and the method comprises the steps: detecting a vehicle entering the pre-detection area, and obtaining the first multi-mode perception data of the vehicle; inputting the first multi-modal perception data of the vehicle into a trained exclusive identification model to extract exclusive identification features of the vehicle for identification to obtain an identification result of the vehicle, the identification result being used for indicating a model type, a behavior type and a virtual identity label of the vehicle; wherein the exclusive recognition model and the universal recognition model are jointly trained under the constraint of a total loss function, basic parameters of the universal recognition model are standard universal parameters issued by the cloud end at the last time, and the standard universal parameters are obtained by the cloud end through aggregating fuzzy network parameters of all the universal recognition models based on federal learning. The method is better in recognition effect and has no risk of privacy leakage.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) technology, specifically relating to a vehicle recognition method and system based on dynamic behavior analysis and federated learning. Background Technology

[0002] With the expansion of universities and the deepening of smart campus construction, the composition of vehicles on campus is becoming increasingly complex, including private cars of faculty and staff, logistics support vehicles, vehicles of visitors from outside the campus, and an increasing number of delivery and courier electric vehicles. This places higher demands on campus traffic safety, area access control, and order management. Vehicle recognition is the foundation of vehicle management, and traditional recognition methods mainly include two types: one based on radio frequency chips and the other based on license plates.

[0003] Radio frequency (RFID) chip-based identification methods typically involve equipping campus vehicles with RFID tags and fixing card readers at access control points for identification. This method requires tagging each vehicle for both temporary and long-term access, and RFID chips have low security, making them easy to copy and impersonate. Therefore, it is costly, has poor security, and cannot achieve sophisticated management.

[0004] License plate-based visual recognition methods typically use cameras to capture license plate numbers for identification, and are currently the mainstream vehicle recognition method. However, the installation locations of cameras on campus are limited, with many trees and teaching buildings, resulting in numerous blind spots. This makes it difficult to identify small, easily concealed targets such as unlicensed electric bikes and bicycles. Furthermore, when using cameras to capture license plates, facial information of users is inevitably captured simultaneously, recording users' travel trajectories and posing a risk of privacy data leakage and misuse.

[0005] Vehicle recognition methods effective in other scenarios, such as the centralized machine learning solutions used by traffic management departments, are not suitable for on-campus environments. This is because data such as vehicle trajectories and driving habits are far more sensitive than static license plate information. The responsibilities of the university and the traffic management system differ, and centralized collection of such information inevitably leads to serious infringement issues. Meanwhile, different management systems on campus, such as those for the cafeteria, main gate access control, parking lots, and dormitories, often belong to different units, resulting in disconnected data. Departments such as the security department, logistics group, and apartment management center, due to their division of responsibilities and data privacy concerns, cannot aggregate their data onto a central platform, creating "data silos" that prevent comprehensive vehicle behavior analysis and collaborative management.

[0006] Therefore, current vehicle identification methods used for vehicle management on campus suffer from poor identification performance or the risk of privacy leaks. Summary of the Invention

[0007] This invention provides a vehicle recognition method and system based on dynamic behavior analysis and federated learning, which can solve the above-mentioned technical problems.

[0008] In a first aspect, embodiments of the present invention provide a vehicle recognition method based on dynamic behavior analysis and federated learning. The method is applied to a local terminal corresponding to a pre-detection area, the local terminal being used to identify vehicles within its corresponding pre-detection area. The method includes: A vehicle entering the pre-inspection area is detected, and the first multimodal perception data of the vehicle is acquired; The first multimodal perception data of the vehicle is input into a trained dedicated recognition model to extract the vehicle's dedicated recognition features for recognition, thereby obtaining the vehicle's recognition result. The recognition result is used to indicate the vehicle's model category, behavior category, and virtual identity identifier. The specific recognition model and the general recognition model are trained together under the constraint of the total loss function. The basic parameters of the general recognition model are the standard general parameters issued by the cloud in the last time. The standard general parameters are obtained by the cloud based on federated learning to aggregate the fuzzy network parameters of all general recognition models.

[0009] Secondly, embodiments of the present invention provide a vehicle recognition system based on dynamic behavior analysis and federated learning, including a cloud and a local terminal corresponding one-to-one with the pre-detection area; The cloud is used to aggregate the blurred network parameters of the general recognition model based on federated learning to obtain standard general parameters, and then send the standard general parameters to the local terminal. The local terminal is used to train a specific recognition model and a general recognition model under the constraint of the total loss function to obtain a trained specific recognition model, wherein the basic parameters of the general recognition model are the standard general parameters; After detecting a vehicle entering the pre-detection area corresponding to the local terminal, the local terminal is also used to acquire the first multimodal perception data of the vehicle; input the first multimodal perception data into the trained dedicated recognition model to extract the vehicle's dedicated recognition features for vehicle recognition, and obtain the vehicle's recognition result, which is used to indicate the vehicle's model category, behavior category and virtual identity identifier.

[0010] The beneficial effects of the embodiments of the present invention compared with the prior art are as follows: The present invention can improve the vehicle recognition effect by collecting multimodal perception data that reflects the vehicle behavior characteristics instead of a single license plate or radio frequency chip; the model used for recognition is trained based on federated learning, rather than directly exchanging vehicle perception data, which can avoid privacy data leakage, realize data exchange under confidentiality, and break the "data silo" effect. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the structure of a dedicated recognition model provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a general recognition model provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a vehicle recognition system based on dynamic behavior analysis and federated learning provided in an embodiment of the present invention; Figure 4 A flowchart illustrating the implementation of a training method for a federated learning-based proprietary recognition model and a general recognition model, provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of a vehicle recognition scenario provided by an embodiment of the present invention; Figure 6 This is a flowchart illustrating the implementation of a vehicle recognition method based on dynamic behavior analysis and federated learning, as provided in an embodiment of the present invention. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0013] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0014] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0015] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0016] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0017] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0018] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0019] Example 1 Figure 1 The diagram shown illustrates the structure of a proprietary recognition model provided in an embodiment of the present invention. As an example and not a limitation, model 1 may include a microscale convolution module 11, a macroscale convolution module 12, a feature fusion module 13, a temporal modeling and attention module 14, and a first classification module 15.

[0020] In one possible implementation, the micro-scale convolution module 11 and the macro-scale convolution module 12 can be used to extract the temporal features of the multimodal sensing data at micro-scale and macro-scale, respectively, to obtain local temporal features at micro-scale and macro-scale. The feature fusion module 13 can align and concatenate local temporal features at different scales to obtain fused features. The temporal modeling and attention module 14 can capture driving details and the macro-environment at the moment of the driving event based on the fused features to obtain specific recognition features. The first classification module 15 can map the specific recognition features to the first predicted recognition result.

[0021] In one example, the microscale convolution module 11 and the macroscale convolution module 12 can be convolution modules with the same structure, but their network parameters differ due to the different scales of the data they process.

[0022] For example, the convolutional module can consist of 2 to 3 one-dimensional convolutional layers, using small convolutional kernels (such as 3 or 5), in conjunction with ReLU activation functions and batch normalization layers to extract local temporal features (such as minor steering adjustments and brief accelerations).

[0023] In one example, the feature fusion module 13 can specifically align local temporal features of different scales in the time dimension, and then concatenate or weightedly add them to obtain fused features.

[0024] In one example, the temporal modeling and attention module 14 can be composed of cascaded gated recurrent submodules 141 and sub-attention layers 142. The gated recurrent submodule 141 can capture driving details and macro environment at the moment the driving event occurs; the sub-attention layer 142 can enable the model to focus on the most critical moment in the entire driving event, generating more discriminative and specific recognition features.

[0025] In one example, the first classification module 15 can be a fully connected classifier.

[0026] Figure 2 The diagram shown illustrates the structure of a general recognition model provided in an embodiment of the present invention. As an example and not a limitation, model 2 may include a cascaded mesoscale convolutional module 21, a gated recurrent module 22, an attention layer 23, and a second classification module 24.

[0027] For example, the mesoscale convolution module 21 can extract mesoscale local temporal features based on the second multimodal perception data, and the gated recurrent module 22 can capture the long-range dependencies and contextual information before and after the occurrence of the driving event (such as the entire process from "starting to decelerate" to "coming to a complete stop") based on the mesoscale local temporal features to obtain hidden features of driving behavior. The attention layer 23 can perform attention operations on the hidden features of driving behavior to make the general recognition model focus on the key moment of decision-making in the driving event (such as the moment of lane change decision), generating a fixed-length, semantically rich general recognition feature.

[0028] Optionally, the structures of the mesoscale convolution module 21, the gated recurrent module 22, the attention layer 23, and the second classification module 24 can be the same as those of the convolution module, the gated recurrent submodule 141, the sub-attention layer 142, and the first classification module 15 in the dedicated recognition model 1, only the specific network parameters are different.

[0029] Example 2 Figure 3 The diagram shown illustrates the structure of a vehicle recognition system based on dynamic behavior analysis and federated learning, as provided in an embodiment of the present invention. As an example and not a limitation, system 3 may include a cloud terminal 31 and local terminals 32 corresponding one-to-one with pre-detection areas.

[0030] In one example, the cloud-based system 31 can obtain standard general parameters by aggregating the blurred network parameters of the general recognition model based on federated learning and then send them to the local end. The local end 32 trains its own general recognition model and its own specific recognition model under the constraint of the total loss function.

[0031] For example, the basic parameters of the general recognition model during this training are the standard general parameters previously distributed from the cloud.

[0032] In one example, after the model is trained, the local device can obtain the vehicle's first multimodal perception data when it detects a vehicle entering its corresponding pre-detection area. This data is then input into the trained dedicated recognition model to extract the vehicle's dedicated recognition features and obtain the vehicle recognition result.

[0033] For example, the vehicle identification results can indicate the vehicle's model category, behavior category, and virtual identity.

[0034] Example 3 Figure 4 The diagram illustrates a training method for a federated learning-based specialized recognition model and a general recognition model, as provided in an embodiment of the present invention. This method is illustrative and not limiting, and can be applied to the aforementioned system 3. The method may include steps S401-S410, which are described below.

[0035] S401, the local end inputs the first and second multimodal perception data of the sample vehicles in the t-th round into the dedicated recognition model and the general recognition model trained in the (t-1)-th round, respectively, to extract the dedicated recognition features and general recognition features of the t-th round, and obtain the first and second predicted recognition results of the t-th round.

[0036] For example, the first multimodal sensing data includes micro-scale and macro-scale multimodal sensing data, and the second multimodal sensing data includes meso-scale multimodal sensing data. The data window sizes for micro-scale, meso-scale, and macro-scale data increase sequentially; for example, the sizes of the three data windows could be 2s, 5s, and 30s, respectively.

[0037] Specifically, the multimodal sensing data sequence can be expanded forward and backward from the moment the driving event occurs to obtain multimodal sensing data of the corresponding scale.

[0038] For example, t is a positive integer less than or equal to Tmax, and Tmax is the preset maximum number of iterations.

[0039] Specifically, the dedicated and general recognition models after the 0th round of training can be the dedicated and general recognition models trained in the previous round, respectively. The model parameters (i.e., basic model parameters) used in the general recognition model trained in the previous round can be the standard general parameters issued in the cloud last time.

[0040] S402, the local end determines the total loss of round t based on the exclusive and general identification features of round t, and the first and second prediction identification results of round t.

[0041] In one example, the total loss can be calculated using the total loss function, which satisfies the following formula:

[0042] in, , , , These are the total loss function, the cross-entropy loss function for the first and second predicted recognition results, and the distillation loss function, respectively. , There are two weight parameters. For example, the distillation loss function is used to constrain the general recognition model to learn the decision style and feature representation of the specific recognition model, and can be the sum of the soft target distillation loss function and the feature alignment loss function.

[0043] Specifically, the soft target distillation loss function and the feature alignment loss function can satisfy the following formulas respectively:

[0044] in, , These are the soft target distillation loss function and the feature alignment loss function, respectively. This represents the mean square error. , The specific identification features of the sample vehicles extracted by the dedicated identification model and the general identification model, respectively. General identification features , Denotes KL divergence, , These are the first and second predicted identification results for the sample vehicles, respectively.

[0045] S403, the local end updates the network parameters of the dedicated recognition model and the general recognition model after training in round t-1 according to the total loss in round t, and obtains the dedicated recognition model and the general recognition model after training in round t.

[0046] In one example, model parameters can be updated based on the backpropagation algorithm.

[0047] S405, the local end determines whether the convergence condition is met.

[0048] In one example, if the convergence condition is met, step S406 can be performed.

[0049] For example, the convergence condition can be that t is greater than Tmax, or the total loss in round t is less than a preset convergence threshold.

[0050] In another example, if the convergence condition is not met, we can let t=t+1 and start the next round of iterative training from step S401.

[0051] S406, the local end uses the dedicated recognition model and the general recognition model trained in the tth round as the trained dedicated recognition model and the pre-trained general recognition model, respectively.

[0052] S407: The local device sends the blurred network parameter update values ​​of the pre-trained general recognition model to the cloud.

[0053] For example, the updated value refers to the difference between the network parameters of the pre-trained general recognition model and its basic parameters.

[0054] In one example, Gaussian noise can be used to blur the updated network parameter values.

[0055] Accordingly, the cloud receives the blurred network parameter update values.

[0056] The parameters of the general recognition model of this invention participate in federated learning, enabling it to absorb and transmit global knowledge. It encodes common and consensus-based vehicle behavior patterns from different local endpoints (e.g., most vehicles slow down before speed bumps) into the model's own parameters and shares them with other local endpoints through federated aggregation. Meanwhile, the parameters of the dedicated recognition model remain local, enabling it to discover and protect local characteristics and focus on fine-grained behavior patterns specific to the pre-detection area corresponding to the local endpoint (e.g., the specific delivery route and loitering behavior of delivery vehicles).

[0057] S408, the cloud aggregates all the blurred network parameters to obtain the general parameters of this standard.

[0058] In one possible implementation, the cloud can perform weighted aggregation of the blurred network parameters based on contribution.

[0059] For example, the contribution of a model can be reflected by a performance improvement parameter, which can be the improvement in the accuracy of the model's predictions.

[0060] In one example, the standard general parameters can be calculated using the following formula:

[0061] These are standard, universal parameters. For the first The aggregate weight of each local endpoint For the first Updated network parameters of a pre-trained general recognition model. Indicates to Add Gaussian noise for blurring; in:

[0062] For a weight parameter, For the first The number of sample vehicles used by each local terminal during training. This represents the total number of sample vehicles used during all local training sessions. For the first Local performance improvement parameters K This represents the total number of local endpoints.

[0063] S409, the cloud sends the general parameters of this standard to the local terminal.

[0064] Accordingly, the local end receives the general parameters of this standard.

[0065] S410, the local end updates the pre-trained general recognition model according to the general parameters of this standard to obtain the trained general recognition model.

[0066] This invention trains two models based on a federated learning framework, uploading only the updated model parameters to the cloud. The original vehicle trajectory and behavior data (such as multimodal perception data and specific and general recognition features) remain locally, effectively avoiding the risk of privacy data leakage. Simultaneously, federated learning allows different local endpoints to collaborate (train models) without sharing the original data, aggregating "data value," such as performance improvement parameters, rather than the data itself. This fully utilizes data information scattered across different local endpoints, breaking down the "data silo" effect.

[0067] Furthermore, models on the local machine can automatically learn new behavioral patterns over time, continuously improving the model and enhancing its adaptability. Distributing the computational tasks of feature extraction and local model training to the local machine can reduce the computational pressure on the cloud, and reasonable allocation of computational tasks can improve computational efficiency.

[0068] Furthermore, in traditional federated learning, the model responsible for identification and sharing is usually the same model. However, this invention enables the general identification model to share parameters, aggregate and absorb global knowledge, and then pass the global knowledge to the dedicated identification model through the constraints of the loss function. At the same time, the dedicated identification model is retained locally to adapt to local characteristics, and the local-specific knowledge is promptly passed to the general identification model through the knowledge distillation process. This allows the dedicated identification model of this invention to ensure both good generalization ability in the global scope and excellent performance in each local scenario, significantly improving the identification effect.

[0069] Example 4 Figure 5The diagram shown is a scenario illustration of vehicle recognition provided by an embodiment of the present invention.

[0070] As an example, see Figure 5 Local terminals can be electronic devices deployed in different areas of the campus, such as computers and servers in guard rooms, security offices, and monitoring rooms. Each local terminal has its own pre-detection area for monitoring and identification (see...). Figure 5 In addition to local processing equipment, the campus also has detection devices such as cameras, radar, and geomagnetic sensors. The detection devices located in the pre-detection area corresponding to the local terminal will upload the data they detect to the local terminal. The local terminal can process the perception data sequence based on data windows of different sizes to obtain the first multimodal perception data, and perform vehicle recognition based on this.

[0071] For example, multimodal perception data can be behavioral data that characterizes the driving characteristics of vehicles on campus, such as vibration patterns when a vehicle passes over speed bumps / roadblocks, deceleration-stop-start speed sequences of vehicles at intersections and zebra crossings, low-speed cruising and loitering behavior in specific areas (such as under dormitory buildings), and steering and micro-manipulation sequences when entering / entering parking spaces.

[0072] For example, there are speed bumps in the corresponding pre-inspection area (see...). Figure 5 The purple-filled pre-detection area indicates the moment a vehicle crosses a speed bump, which can be defined as the moment a driving event occurs. Within the pre-detection area, detection devices such as accelerometers, cameras, and Bluetooth devices collect multimodal data, including the vehicle's acceleration, video trajectory, and Bluetooth beacon positioning data, to obtain a multimodal perception data sequence. Then, using the moment of the driving event as the midline, data is extracted from this data sequence using different data windows to obtain the first and second multimodal perception data.

[0073] For example, the vehicle identification results can be used to indicate the vehicle's model category, behavior category, and virtual identity.

[0074] For example, when an unknown vehicle enters the campus, the dedicated identification model captures its unique identification features and matches them with existing features in the database. If the match is successful, its identity is identified, for example, it is determined to be a registered delivery tricycle that is driving normally, and its virtual identity is A1256. If the match is unsuccessful, it is marked as abnormal, and other local terminals are notified to pay attention to the vehicle. After the administrator confirms that it is harmless, a new virtual identity is granted to it and its unique identification features are saved to the database.

[0075] Figure 6The diagram shown illustrates a vehicle recognition method based on dynamic behavior analysis and federated learning, as provided in an embodiment of the present invention. As an example and not a limitation, the recognition method may include steps S601-S603, which are described below.

[0076] S601, a vehicle entering the pre-inspection area corresponding to the local terminal is detected, and the first multimodal perception data of the vehicle is obtained.

[0077] For example, during the use of a trained model, typically only the first multimodal perception data of the vehicle is acquired for identification. Alternatively, the first and second multimodal perception data of some vehicles can be sampled periodically. After confirmation by management personnel, the identification results of these sampled vehicles can be used as sample vehicles for the next training of the model.

[0078] Optionally, the administrator can periodically conduct spot checks on the vehicle recognition results, and start a new training of the model whenever the model's recognition accuracy is less than the preset retraining threshold; or directly set the model to be retrained every preset period.

[0079] S602, the first multimodal perception data is input into the trained dedicated recognition model to extract the vehicle's dedicated recognition features for recognition, and the recognition result of the vehicle is obtained.

[0080] This invention improves vehicle recognition by collecting multimodal perception data that reflects vehicle behavior characteristics, rather than relying on a single license plate or radio frequency chip. It also avoids privacy data leakage by training the recognition model based on federated learning, rather than directly exchanging vehicle perception data, thus enabling data exchange under confidentiality and breaking the "data silo" effect.

[0081] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

Claims

1. A vehicle recognition method based on dynamic behavior analysis and federated learning, characterized in that, The method is applied to a local terminal that corresponds one-to-one with a pre-inspection area. The local terminal is used to identify vehicles within its corresponding pre-inspection area. The method includes: A vehicle entering the pre-inspection area is detected, and the first multimodal perception data of the vehicle is acquired; The first multimodal perception data of the vehicle is input into a trained dedicated recognition model to extract the vehicle's dedicated recognition features for recognition, thereby obtaining the vehicle's recognition result. The recognition result is used to indicate the vehicle's model category, behavior category, and virtual identity identifier. The specific recognition model and the general recognition model are trained together under the constraint of the total loss function. The basic parameters of the general recognition model are the standard general parameters issued by the cloud in the last time. The standard general parameters are obtained by the cloud based on federated learning to aggregate the fuzzy network parameters of all general recognition models.

2. The method according to claim 1, characterized in that, The training methods for the specific recognition model and the general recognition model include: Under the constraint of the total loss function, the previously trained dedicated recognition model and general recognition model are trained based on the first and second multimodal perception data of the sample vehicles to obtain the currently trained dedicated recognition model and the currently pre-trained general recognition model. The basic parameters of the previously trained general recognition model are the previous standard general parameters. Send the blurred network parameter update values ​​of the general recognition model after this pre-training to the cloud so that the cloud can aggregate all the blurred network parameters to obtain the standard general parameters for this time; The pre-trained general recognition model is updated based on the standard general parameters to obtain the trained general recognition model.

3. The method according to claim 2, characterized in that, The total loss function includes: in, , , , These are the total loss function, the cross-entropy loss function of the general recognition model, the cross-entropy loss function of the specific recognition model, and the distillation loss function, respectively. , There are two weight parameters; the distillation loss function is used to constrain the general recognition model to learn the decision style and feature representation of the specific recognition model.

4. The method according to claim 3, characterized in that, The distillation loss function is the sum of the soft target distillation loss function and the feature alignment loss function, and the soft target distillation loss function and the feature alignment loss function respectively satisfy the following formulas: in, , These are the soft target distillation loss function and the feature alignment loss function, respectively. This represents the mean square error. , These are the dedicated recognition features and general recognition features of the sample vehicles extracted by the dedicated recognition model and the general recognition model, respectively. Denotes KL divergence, , These are the first and second predicted recognition results for the sample vehicle, respectively, provided by the dedicated recognition model and the general recognition model.

5. The method according to claim 4, characterized in that, The standard general parameters satisfy the following formula: For the aforementioned standard general parameters, For the first The aggregate weight of each local endpoint For the first Updated network parameters of a pre-trained general recognition model. Indicates to Add Gaussian noise for blurring; in: For a weight parameter, For the first The number of sample vehicles used by each local terminal during training. This represents the total number of sample vehicles used during all local training sessions. For the first Local performance improvement parameters K This represents the total number of local endpoints.

6. The method according to claim 5, characterized in that, The performance improvement parameter refers to the improvement in accuracy.

7. The method according to claim 4, characterized in that, The first multimodal sensing data includes micro-scale and macro-scale multimodal sensing data, and the second multimodal sensing data includes meso-scale multimodal sensing data, with the data window size increasing sequentially for micro-scale, meso-scale, and macro-scale. The distribution of the specific identification features is related to the pre-detection area corresponding to the local terminal, and the general identification features are used to characterize the general driving behavior features of the vehicle.

8. The method according to claim 7, characterized in that, The dedicated recognition model includes a micro-scale convolution module, a macro-scale convolution module, a feature fusion module, a temporal modeling and attention module, and a first classification module; The micro-scale and macro-scale convolution modules are used to extract the temporal features of the micro-scale and macro-scale multimodal sensing data, respectively, to obtain the local temporal features at the micro-scale and macro-scale. The feature fusion module is used to align and stitch together the local temporal features at the micro-scale and macro-scale to obtain fused features; The temporal modeling and attention module is used to capture driving details and macroscopic environment at the moment of a driving event based on the fused features, and obtain the exclusive recognition features; The first classification module is used to map the specific identification features to the first predicted identification result.

9. The method according to claim 7, characterized in that, The general recognition model includes a cascaded mesoscale convolution module, a gated recurrent module, an attention layer, and a second classification module; The mesoscale convolution module is used to extract mesoscale local temporal features based on the second multimodal sensing data; The gated loop module is used to capture the long-range dependencies and contextual information before and after the occurrence of a driving event based on the mesoscale local temporal features, and to obtain hidden features of driving behavior. The attention module is used to perform attention operations on the hidden features of the driving behavior so that the general recognition model focuses on key decision moments to obtain the general recognition features. The second classification module is used to map the general recognition features to the second predicted recognition result.

10. A vehicle recognition system based on dynamic behavior analysis and federated learning, characterized in that, This includes both cloud-based and local terminals that correspond one-to-one with the pre-inspection areas; The cloud is used to aggregate the blurred network parameters of the general recognition model based on federated learning to obtain standard general parameters, and then send the standard general parameters to the local terminal. The local terminal is used to train a specific recognition model and a general recognition model under the constraint of the total loss function to obtain a trained specific recognition model, wherein the basic parameters of the general recognition model are the standard general parameters; After detecting a vehicle entering the pre-detection area corresponding to the local terminal, the local terminal is also used to acquire the first multimodal perception data of the vehicle; input the first multimodal perception data into the trained dedicated recognition model to extract the vehicle's dedicated recognition features for vehicle recognition, and obtain the vehicle's recognition result, which is used to indicate the vehicle's model category, behavior category and virtual identity identifier.

Images