A semantic coding method for traffic state monitoring

Abstract

The application discloses a semantic coding method for traffic state monitoring, which is based on a semantic communication framework of deep learning. The semantic coding network for traffic state monitoring is constructed, including a data encoder-decoder network, a discriminator network 1 and a discriminator network 2. The semantic coding network is trained by using a sequential alternating training mechanism of a generative adversarial network mode. The data encoder-decoder network is used for compressing original information, extracting task target semantics and hiding privacy information. The discriminator network 1 is used for controlling the code rate of information transmission. The discriminator network 2 is used for controlling the risk of privacy information leakage. Through the joint optimization strategy among several networks, the control of the transmission code rate, the balance between the quality of task target semantic expression and privacy protection are realized. The method is suitable for the data transmission task for traffic state monitoring in an intelligent transportation system and has good compression performance and privacy leakage prevention ability.

Description

Technical Field

[0001] This invention relates to the field of encoding and decoding technology for traffic condition monitoring and transmission, and in particular to a semantic communication method based on deep learning, for achieving efficient information encoding, reconstruction, and optimized control of target task semantic information and privacy semantic information during the transmission and processing of traffic conditions. Background Technology

[0002] With the rapid development of intelligent transportation systems, road traffic monitoring plays an increasingly important role in ensuring traffic safety, improving traffic efficiency, and supporting urban planning. However, the data collected and transmitted by road traffic monitoring systems often involves sensitive information such as pedestrians and driving behavior, which can easily lead to privacy leaks. In the current environment of limited communication resources, how to efficiently transmit traffic status information while ensuring data privacy has become a significant challenge for intelligent transportation systems.

[0003] Traditional communication systems often employ classic coding schemes from Shannon's information theory, aiming to restore original data without loss, while neglecting semantic information representation. In applications such as intelligent transportation, systems don't always need to recover all the details of the original data; instead, they focus on the key semantic information conveyed, such as "whether a road segment is congested" or "whether a vehicle is engaging in dangerous behavior." Therefore, semantic communication-based transmission mechanisms have emerged as a promising alternative.

[0004] Semantic communication aims to deliver the most valuable information with minimal resource overhead, with the goal of enabling the receiver to accurately understand the sender's "intent" or "task objective." In recent years, thanks to the development of deep learning, neural network-based semantic encoding and decoding frameworks have been widely researched and applied in various fields such as image recognition and natural language processing, achieving significant results. However, in privacy-sensitive scenarios such as traffic condition monitoring, designing a semantic communication system that simultaneously possesses high semantic extraction capabilities and strong privacy protection remains a challenging problem. Summary of the Invention

[0005] This invention proposes a semantic communication method for traffic condition monitoring. By constructing a neural network communication framework that integrates semantic extraction and privacy protection, it achieves an organic combination of efficient semantic-level transmission of traffic data and information security. This method can be widely applied to various intelligent transportation scenarios such as vehicle-road cooperation, traffic scheduling, and anomaly detection, improving system communication efficiency while effectively reducing the risk of sensitive information leakage.

[0006] The technical solution adopted by this invention to solve its technical problem is: A semantic coding method for traffic condition monitoring includes the following steps: Step 1: Collect multi-source data related to traffic condition monitoring, including raw information. Semantic information of the target task and privacy information By using preprocessing techniques, data formats and dimensions are standardized, thereby enhancing data quality.

[0007] Step 2: Construct a semantic coding network for traffic condition monitoring, including: an encoding / decoding network. Discriminator Network 1 Discriminator Network 2 The three networks mentioned above form a multi-objective training framework through joint optimization. By designing a reasonable loss function, the framework can reduce the leakage of privacy information in the reconstruction results while ensuring reconstruction quality and semantic accuracy.

[0008] Step 3: Train the semantic encoding network using a sequential alternating training mechanism in the Generative Adversarial Network (GAN) model.

[0009] Step 4: Deploy the trained semantic coding system to the actual traffic monitoring platform.

[0010] Specifically, in step 1, The multi-source data related to traffic condition monitoring includes traffic camera images, sensor data, and vehicle flow information, constructing a system containing raw information. Semantic information of the target task (such as traffic density, congestion level, vehicle speed prediction, etc.) and privacy information The dataset consists of three parts: license plate information, pedestrian characteristics, and vehicle individual characteristics. Details are as follows: The raw information refers to the multimodal perception data collected from traffic monitoring scenarios, such as raw observation information from vehicles and roadside camera video frames, which has high dimensionality and strong redundancy characteristics. The target task semantic information represents the key semantic targets in traffic condition monitoring, such as the congestion level, traffic density, and anomaly detection indicators of the current road segment. It is the "effective task semantics" that ultimately needs to be maintained in the compressed space. Privacy information within the original data refers to information directly related to an individual or vehicle that may lead to privacy breaches, such as license plate numbers, facial features, vehicle exterior markings, pedestrian patterns, and vehicle trajectory patterns. This type of information requires semantic protection to prevent its leakage.

[0011] The preprocessing methods include: normalization, missing value handling, image / sequence encoding, and label construction.

[0012] In step 2, Constructing semantic coding networks, including data encoding and decoding networks Discriminator Network 1 and discriminator network 2 The semantic coding network is as follows: Codec Network The goal is to minimize reconstruction distortion, maximize semantic expressiveness, minimize the comprehensive mutual information between the original and the reconstructed versions, and minimize the privacy mutual information between the original privacy semantics and the reconstructed versions.

[0013] Discriminator Network 1 Estimate the original information With reconstructing information The comprehensive mutual information is used to guide the coding and decoding network to improve its data compression capabilities.

[0014] Discriminator Network 2 : Estimated privacy information With reconstructing information The privacy information exchanged between them guides the codec to suppress privacy leaks.

[0015] Furthermore, the data encoding / decoding network described above processes traffic data with the support of a semantic coding algorithm. This algorithm, based on lossy source coding theory, transmits the original information while balancing source distortion and compression ratio. The algorithm is as follows: In the formula, This is the original information. For the reconstruction of the original information, For the semantic information of the target task, For the reconstructed task semantic information, For encoding and decoding networks.

[0016] The data encoding / decoding network updates its parameters by reconstructing the distortion of the original information and the accuracy of the semantic information of the target task. When the distortion is not equal to a predetermined value or the accuracy of the semantic information of the target task is not equal to a predetermined value, the parameters of the data encoding / decoding network are updated through the backpropagation mechanism of the neural network. This process is repeated iteratively until the distortion and the accuracy of the semantic information of the target task match the predetermined values, at which point the iteration stops.

[0017] Specifically, the constraint algorithm for matching distortion to a predetermined value is expressed as follows: In the formula, A specific representation of the original information. To reconstruct the specific representation of information, For the number of datasets, The data dimension of the original information. This is a constraint value for distortion, ranging from 0 to 1, and is adjusted according to the required bitrate. When... When the value is 0, the distortion satisfies the constraint condition.

[0018] The constraint algorithm for matching the accuracy of semantic information of the target task to a predetermined value is expressed as follows: In the formula, Classification of semantic information for the target task ( (Total number of categories) Indicates sample The true label or true semantic information, For indicator functions, when When the time condition is met, its value is 1; otherwise, it is 0. To reconstruct the probability distribution of the semantic information of the task, representing the first... Each semantic information is a class The probability distribution; function Cross-entropy is used to measure the difference between the semantic information of the reconstructed task and the semantic information of the real task. This is a constraint value for the semantic accuracy of the target task. To improve the accuracy of semantic transmission, this value is preferably set to 0. When The accuracy meets the constraints.

[0019] Furthermore, the reconstructed information output by the data encoding / decoding network will serve as the input to discriminator network 1. Discriminator network 1 will optimize its network parameters based on the original and reconstructed information through the backpropagation mechanism of the neural network, enabling it to accurately measure the comprehensive mutual information between the reconstructed and original information. The algorithm is expressed as follows: In the formula, The joint distribution of original and reconstructed information. Let be the mathematical expectation of this joint distribution; The marginal distribution of original information and reconstructed information. Let be the mathematical expectation of this marginal distribution; The KL divergence between the joint distribution and the marginal distribution is used to measure... and The difference between the two probability distributions For discriminator network 1, It is a comprehensive mutual information between the original information and the reconstructed information.

[0020] For discriminator network 1 The loss function is as follows: In the formula, The sigmoid function is commonly used in neural networks to perform nonlinear transformations on numerical values. Through backpropagation of the neural network, the parameters of the discriminator network 1 are optimized so that the output of the network can gradually approach the combined mutual information of the original and reconstructed information.

[0021] Furthermore, the reconstructed information output by the data encoding / decoding network will serve as the input to discriminator network 2. Discriminator network 2 will optimize its network parameters based on the privacy information and reconstructed information in the original information, through the backpropagation mechanism of the neural network, enabling it to accurately measure the privacy mutual information between the reconstructed information and the privacy information. The algorithm can be expressed as: In the formula, For private information in the original information, The joint distribution of privacy information and reconstructed information in the original information. Let the expected value of this joint distribution be... The marginal distribution of privacy information and reconstructed information in the original information. Let the expected value of this marginal distribution be... The KL divergence between the joint distribution and the marginal distribution is used to measure... and The difference between the two probability distributions For discriminator network 2, This refers to the privacy information between the original information and the reconstructed information.

[0022] For discriminator network 2 The loss function is as follows: In the formula, This is the sigmoid function. Through backpropagation of the neural network, the parameters of discriminator network 2 are optimized so that its output gradually approximates the privacy-preserving mutual information between the original and reconstructed information.

[0023] Furthermore, the comprehensive mutual information output by the discriminator network 1 And the privacy mutual information output by the discriminator network 2. The parameters of the data encoding and decoding network are adjusted through the backpropagation mechanism of the neural network when the comprehensive mutual information, privacy mutual information, distortion, or accuracy of the target task semantic information are not equal to the predetermined value. This process is repeated until the comprehensive mutual information, privacy mutual information, distortion of the reconstructed information, and accuracy of the target task semantic information are matched to the predetermined value, at which point the iteration stops.

[0024] For data encoding / decoding networks The loss function is as follows: In the formula, The constraint value for integrated mutual information, This is a constraint value for privacy mutual information, and its value ranges from 0 to 1. The closer it is to 0, the smaller the mutual information is, and the closer it is to 1, the larger the mutual information is.

[0025] Step 3, The data encoding / decoding network Discriminator Network 1 and discriminator network 2 They are interconnected and trained sequentially and cyclically using a generative adversarial network (GAN) approach.

[0026] Specifically, in each round of training: Fix discriminator network 1 and discriminator network 2, and update the encoder / decoder network. The parameters are minimized to obtain the total loss function. To simultaneously satisfy the constraints of reconstruction distortion, semantic accuracy, comprehensive mutual information constraints, and privacy mutual information constraints.

[0027] Then, with the encoder-decoder network fixed, discriminator network 1 and discriminator network 2 are trained separately to improve their ability to synthesize mutual information. and privacy information The accuracy of the estimation.

[0028] The training is performed iteratively through the above process until all four metrics (distortion, semantic accuracy, comprehensive mutual information, and privacy mutual information) meet the predetermined constraints, at which point the model converges.

[0029] Step 4, Specifically, the trained The encoding and decoding network is integrated into edge devices or servers and encapsulated as a standard API interface, supporting real-time access to data streams from traffic sensing systems. The compressed representation output by the semantic encoder is used in the uplink data communication of the intelligent transportation system, saving bandwidth while improving semantic expression capabilities, and enabling further analysis and decision-making in the cloud.

[0030] Beneficial effects By employing joint source-channel coding and semantic coding algorithms, based on lossy source coding theory, key traffic semantic information is accurately extracted and transmitted. Compared to traditional bitstream transmission, this significantly reduces unnecessary data volume, alleviates network capacity pressure, lowers end-to-end latency, and meets the real-time requirements of traffic condition monitoring. In high-traffic areas, it can quickly transmit critical information such as congestion and vehicle speed, assisting in traffic command and dispatch.

[0031] Meanwhile, by introducing a discriminator network 2 to measure privacy mutual information and optimize parameters, sensitive information such as driver identity and vehicle trajectory can be effectively protected. Compared with existing mechanisms, this approach more precisely controls the risk of privacy information leakage. Even if data is illegally obtained, attackers will find it difficult to access sensitive content, ensuring the secure transmission of traffic data.

[0032] On the other hand, the data encoding and decoding network updates its parameters based on the reconstruction distortion and the semantic accuracy of the task, iteratively bringing both to predetermined values. This ensures that the receiving end accurately reconstructs the traffic condition, providing a reliable basis for traffic decisions, helping the traffic monitoring center to clearly understand the real-time road conditions, and assisting in reasonable management and control decisions.

[0033] Furthermore, the data encoding / decoding and discriminator networks 1 and 2 are trained sequentially and cyclically through generative adversarial networks, collaborating and optimizing each other to improve model stability and generalization ability. By comprehensively optimizing mutual information, distortion, and semantic accuracy, a balance is achieved in traffic data processing between efficiency, accuracy, and privacy protection, providing strong technical support for the development of intelligent transportation. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the semantic encoding algorithm framework of an embodiment of the present invention; Figure 3 This is a schematic diagram of the system deployment according to an embodiment of the present invention; Figure 4 These are effect diagrams of an embodiment of the present invention; Detailed Implementation The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0035] This invention discloses a semantic-based encoding and decoding method for traffic condition monitoring and transmission, the processing flow of which is as follows: Figure 1 As shown.

[0036] like Figure 2 This is a schematic diagram of the semantic encoding algorithm framework according to an embodiment of the present invention, wherein... This is the original information. For the reconstruction of the original information, For the semantic information of the target task, For the reconstructed task semantic information, This refers to private information in the original data.

[0037] Original information Typically, this consists of image or video frames collected from the vehicle or roadside, which contain task semantic information. and privacy information Task semantic information can be allocated according to the specific task objectives. and privacy semantic information As labels and original information Label and associate.

[0038] Indicates mutual information, A measure of the distortion between the original and reconstructed information. This refers to the accuracy of the original and reconstructed information target task information.

[0039] Figure 2 It contains three networks: an encoder-decoder network, a discriminator network 1, and a discriminator network 2; the encoder-decoder network... and its parameters Discriminator Network 1 and its parameters and discriminator network 2 and its parameters .

[0040] The data encoding / decoding network relies on semantic coding algorithms to achieve efficient data encoding and decoding. Based on lossy source coding theory, this algorithm improves data compression efficiency by adjusting the distortion between the original and reconstructed data. Simultaneously, the network extracts task-oriented semantic information from the original data, ensuring the accuracy and reliability of data transmission while maintaining the accuracy of the target task's semantic information transmission. Furthermore, the network measures the privacy information between the original and reconstructed information to control the risk of privacy information leakage during data transmission.

[0041] In general, in traffic condition monitoring scenarios, the raw information transmitted is video or image information collected at the roadside, which includes semantic information about road conditions, such as traffic congestion, accidents, and violations.

[0042] The data encoding / decoding network processes data with the support of semantic encoding algorithms, performing encoding and decoding on traffic data. Its network structure can be selected from model structures such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), or Transformers, depending on the type of data being processed. It extracts semantic information for the task and reconstructs the original information through encoding and decoding. The prediction algorithm used in this invention is as follows: In the formula, This is the original information. For the reconstruction of the original information, For the semantic information of the target task, For the reconstructed task semantic information, For encoding and decoding networks.

[0043] In the formula, As input to the data encoding / decoding network, This is the output of the encoding / decoding network. The data encoding / decoding network continuously adjusts its weights and thresholds through training on sample data. When the distortion is not equal to a predetermined value or the accuracy of the target task's semantic information is not equal to a predetermined value, the weights and thresholds of the data encoding / decoding network are decreased along the negative gradient direction through the backpropagation mechanism of the neural network. This process is iterated until convergence, so that the distortion and the accuracy of the target task's semantic information match the predetermined values. Distortion matching to a predetermined value can be expressed as: In the formula, A specific representation of the original information. To reconstruct the specific representation of information, For the number of datasets, For the feature dimensions of the original information, This is a constraint value for distortion; the smaller the value, the smaller the distortion. When the value is 0, the distortion satisfies the constraint condition.

[0044] In typical traffic condition monitoring scenarios, It can be represented as the information of each pixel in each frame of the original information. This constraint uses the mean square error (MSE) algorithm, which can accurately measure the degree of distortion between the original information and the reconstructed information at the pixel level.

[0045] The accuracy of matching the semantic information of the target task to a predetermined value can be expressed as: In the formula, For indicator functions, when When the time condition is met, its value is 1; otherwise, it is 0. To reconstruct the probability distribution of the semantic information of the task, the function Cross-entropy is used to measure the difference between the semantic information of the reconstructed task and the semantic information of the true task. This is a constraint value for the semantic accuracy of the task; the smaller the value, the higher the accuracy. The accuracy meets the constraints.

[0046] In typical traffic condition monitoring scenarios, This represents the semantic information of the target task being transmitted. It can represent the actual traffic status category, such as congestion, slow traffic, and smooth traffic. This means that the traffic state predicted by the algorithm based on the reconstructed information belongs to a certain category. The probability of this is determined by the quantitative formula for the accuracy of semantic information in the target task. This allows for a precise assessment of the difference between the reconstructed traffic state information and the actual traffic state, determining whether the algorithm's accuracy in judging traffic conditions meets predetermined requirements. This ensures that the traffic condition monitoring system makes reliable decisions based on accurate road condition semantic information.

[0047] Regarding data compression ratio, the reconstructed information output by the data encoding / decoding network serves as the input to discriminator network 1. Discriminator network 1 employs a multilayer perceptron structure to reflect the comprehensive mutual information between the original and reconstructed information. The mutual information estimation uses a KL (Kullback-Leibler) divergence metric. During optimization, the weights of discriminator network 1 are continuously updated through backpropagation to accurately estimate the comprehensive mutual information between the original and reconstructed information. This feedback informs the encoding / decoding network of its data compression efficiency. The algorithm can be expressed as follows: In the formula, The joint distribution of original and reconstructed information. The marginal distribution of original information and reconstructed information. The KL divergence is the joint distribution of the marginal distribution. For discriminator network 1, It is a comprehensive mutual information between the original information and the reconstructed information.

[0048] For discriminator network 1 The loss function is as follows: In the formula, The sigmoid function is used to optimize the parameters of discriminator network 1 through backpropagation of the neural network, so that the output of the network can gradually approach the combined mutual information of the original information and the reconstructed information.

[0049] Similarly, regarding privacy protection, the reconstructed information output by the data encoding / decoding network will simultaneously serve as the input to discriminator network 2. Discriminator network 2 employs a similar multilayer perceptron structure to measure the mutual information between the reconstructed information and the privacy information in the original information. Discriminator network 2 will also optimize its network parameters based on the backpropagation mechanism of neural networks, enabling it to accurately measure the privacy mutual information between the reconstructed information and the privacy information, thereby feeding back the privacy leakage risk of its encoding / decoding to the encoding / decoding network. The algorithm can be expressed as follows: In the formula, For private information in the original information, The joint distribution of privacy information and reconstructed information in the original information. The marginal distribution of privacy information and reconstructed information in the original information. The KL divergence is the joint distribution of the marginal distribution. For discriminator network 2, This refers to the privacy information between the original information and the reconstructed information.

[0050] In typical traffic condition monitoring scenarios, This refers to privacy information in images or videos collected from roadside locations, such as driver identity and vehicle trajectory.

[0051] For discriminator network 2 The loss function is as follows: In the formula, The sigmoid function is used to optimize the parameters of discriminator network 2 through backpropagation of the neural network, so that the output of the network can gradually approximate the privacy mutual information between the privacy information in the original information and the privacy mutual information of the reconstructed information.

[0052] Furthermore, the comprehensive mutual information output by discriminator network 1 and the privacy mutual information output by discriminator network 2 will be used as inputs to the data encoding and decoding network. When the comprehensive mutual information ≠ the predetermined value, the privacy mutual information ≠ the predetermined value, the distortion ≠ the predetermined value, or the accuracy of the target task semantic information ≠ the predetermined value, the parameters of the data encoding and decoding network will be adjusted through the backpropagation mechanism of the neural network. This process will be repeated iteratively until the comprehensive mutual information, the privacy mutual information, the distortion of the reconstructed information, and the accuracy of the target task semantic information match the predetermined value, at which point the iteration will stop.

[0053] For data encoding / decoding networks The loss function is as follows: In the formula, The constraint value for integrated mutual information, This is a constraint value for privacy mutual information; its value ranges from 0 to 1, with values ​​closer to 0 indicating smaller mutual information and values ​​closer to 1 indicating larger mutual information. (This is from an embodiment.) and All are set to 0.1.

[0054] In the traffic condition monitoring scenario of this invention, in order to ensure the compression rate of transmitted data and the risk of privacy and mutual trust leakage, and The preferred setting is a positive value close to 0. Preferably, this value can be set to 0.1 to ensure that the overall mutual information between the reconstructed information and the original information is small, thereby improving the compression rate of the transmitted data. At the same time, it ensures that the privacy mutual information between the reconstructed information and the privacy information in the original information is small, thereby reducing the risk of privacy information leakage.

[0055] The data encoding / decoding network Discriminator Network 1 and discriminator network 2 The training is interconnected and employs a generative adversarial network (GAN) approach for sequential and iterative training. Specifically, in each training round, the parameters of discriminator network 1 and discriminator network 2 are first fixed, and training data is simultaneously input into all three networks. Only the parameters of the data encoder / decoder network are updated, and its fitting ability is optimized through backpropagation. Subsequently, the already trained data encoder / decoder network is fixed, and data is re-inputted to update the parameters of discriminator network 1 and discriminator network 2, thereby improving their ability to estimate comprehensive mutual information and privacy mutual information. These two steps are repeated alternately, constituting one round of training. The entire process is iterated until the model converges.

[0056] The trained data encoder-decoder network For actual task deployment. For example... Figure 3 As shown, the roadside information collection equipment is responsible for collecting raw information. The collected information is transmitted to the data receiving terminal. During this process, the data encoding / decoding network receives information from the endpoint in real time, compressing the original information while embedding the task objective semantics. and privacy semantic information This method hides the data, thereby improving transmission efficiency and reducing the risk of privacy leaks.

[0057] The effect diagram of the embodiment of the present invention is as follows Figure 4As shown, the original image contains a wealth of detailed information, such as dense traffic on the road and pedestrians on the roadside. After these images are processed by the encoding / decoding network of this invention, effective image compression can be achieved. During this process, by embedding semantic information of the target task, the transmission efficiency of the image is significantly improved while ensuring that key information is not lost.

[0058] It is noteworthy that the privacy information in the original image is properly protected during the image encoding and decoding process. For example, personal information such as the facial features and body posture of pedestrians in the image is effectively hidden, avoiding the risk of privacy information leakage.

[0059] Through this implementation method, the system has achieved remarkable results in traffic condition monitoring. On the one hand, it effectively improves monitoring efficiency, enabling rapid processing and transmission of image data; on the other hand, it greatly enhances monitoring accuracy, relying on accurately embedded semantic information to precisely determine traffic conditions. Simultaneously, the system prevents the leakage of private information during transmission, ensuring personal privacy and security.

[0060] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A semantic coding method for traffic condition monitoring, characterized in that, Includes the following steps: Step 1: Collect multi-source data related to traffic condition monitoring, including raw information. Semantic information of the target task and privacy information By using preprocessing techniques, data formats and dimensions are standardized to enhance data quality; Step 2: Construct a semantic coding network for traffic condition monitoring, including: a data encoding / decoding network. Discriminator Network 1 Discriminator Network 2 The three networks mentioned above form a multi-objective training framework through joint optimization. By designing a reasonable loss function, the reconstruction quality and semantic accuracy are guaranteed while reducing the leakage of privacy information by the reconstruction results. Step 3: Train the semantic encoding network using a sequential alternation training mechanism in the generative adversarial network (GAN) model; Step 4: Deploy the trained semantic coding system to the actual traffic monitoring platform; The multi-source data related to traffic condition monitoring includes traffic camera images, sensor data, and traffic flow information; The data encoding / decoding network processes traffic data with the support of semantic encoding algorithms. The algorithm is as follows: In the formula, This is the original information. For the reconstruction of the original information, For the semantic information of the target task, For the reconstructed task semantic information, For data encoding and decoding networks; The data encoding and decoding network updates its parameters by reconstructing the distortion of the original information and the accuracy of the semantic information of the target task. When the distortion is not equal to the predetermined value or the accuracy of the semantic information of the target task is not equal to the predetermined value, the parameters of the data encoding and decoding network are updated through the backpropagation mechanism of the neural network. This process is repeated until the distortion and the accuracy of the semantic information of the target task match the predetermined value, at which point the iteration stops. The constraint algorithm for matching distortion to a predetermined value is expressed as follows: In the formula, A specific representation of the original information. To reconstruct the specific representation of information, For the number of datasets, The data dimension of the original information. This is a constraint value for distortion, ranging from 0 to 1, and is adjusted according to the bitrate requirements; when When the value is 0, the distortion satisfies the constraint condition. The constraint algorithm for matching the accuracy of semantic information of the target task to a predetermined value is expressed as follows: In the formula, Classification of semantic information for the target task. The total number of categories, Indicates sample The true label or true semantic information, For indicator functions, when When the time condition is met, its value is 1; otherwise, it is 0. To reconstruct the probability distribution of the semantic information of the task, representing the first... Each semantic information is a class The probability distribution; function Cross-entropy is used to measure the difference between the semantic information of the reconstructed task and the semantic information of the real task. The constraint value for the semantic accuracy of the target task, when The accuracy meets the constraints. The reconstructed information output by the data encoding / decoding network will serve as the input to discriminator network 1. Discriminator network 1 will optimize its network parameters based on the original and reconstructed information through the backpropagation mechanism of the neural network, enabling it to accurately measure the comprehensive mutual information between the reconstructed and original information. The algorithm is expressed as follows: In the formula, The joint distribution of original and reconstructed information. Let be the mathematical expectation of this joint distribution; The marginal distribution of original information and reconstructed information. Let be the mathematical expectation of this marginal distribution; The KL divergence between the joint distribution and the marginal distribution is used to measure... and The difference between the two probability distributions For discriminator network 1, This is a comprehensive mutual information between the original information and the reconstructed information; The reconstructed information output by the data encoding / decoding network will serve as the input to discriminator network 2. Discriminator network 2 will optimize its network parameters based on the privacy information and reconstructed information in the original information, using the backpropagation mechanism of the neural network, so that it can accurately measure the privacy mutual information between the reconstructed information and the privacy information. The algorithm is expressed as follows: In the formula, For private information in the original information, The joint distribution of privacy information and reconstructed information in the original information. Let the expected value of this joint distribution be... The marginal distribution of privacy information and reconstructed information in the original information. Let the expected value of this marginal distribution be... The KL divergence between the joint distribution and the marginal distribution is used to measure... and The difference between the two probability distributions For discriminator network 2, This refers to the mutual privacy information between the original information and the reconstructed information. The comprehensive mutual information output by discriminator network 1 And the privacy information output by discriminator network 2 As input to the data encoding and decoding network, when the comprehensive mutual information ≠ the predetermined value, or the privacy mutual information ≠ the predetermined value, or the distortion ≠ the predetermined value, or the accuracy of the target task semantic information ≠ the predetermined value, the parameters of the data encoding and decoding network are adjusted through the backpropagation mechanism of the neural network. This process is repeated iteratively until the comprehensive mutual information, privacy mutual information, the distortion of the reconstructed information, and the accuracy of the target task semantic information match the predetermined value, at which point the iteration stops.

2. The semantic coding method for traffic condition monitoring as described in claim 1, characterized in that, In step 1, a structure containing the original information is constructed. Semantic information of the target task and privacy information The dataset consists of three parts; details are as follows: This refers to raw information, namely multimodal sensing data collected from traffic monitoring scenarios; This refers to the semantic information of the target task, representing the key semantic objectives in traffic condition monitoring; Privacy information in the original information refers to information that is directly related to an individual or vehicle and may lead to privacy leaks. The preprocessing methods include: normalization, missing value handling, image / sequence encoding, and label construction.

3. The semantic coding method for traffic condition monitoring as described in claim 1, characterized in that, In step 2, the semantic coding network includes a data encoding / decoding network. Discriminator Network 1 and discriminator network 2 The details are as follows: Data encoding and decoding networks The objectives are to minimize reconstruction distortion, maximize semantic expressiveness, minimize the comprehensive mutual information between the original and the reconstructed semantics, and minimize the privacy mutual information between the original privacy semantics and the reconstructed semantics. Discriminator Network 1 Estimate the original information With reconstructing information The comprehensive mutual information is used to guide data encoding and decoding networks to improve their data compression capabilities; Discriminator Network 2 : Estimated privacy information With reconstructing information The privacy information exchanged between them guides the codec to suppress privacy leaks.

4. The semantic coding method for traffic condition monitoring as described in claim 3, characterized in that, For discriminator network 1 The loss function is as follows: In the formula, The sigmoid function is used to optimize the parameters of discriminator network 1 through backpropagation of the neural network, so that the output of the network can gradually approach the combined mutual information of the original and reconstructed information.

5. The semantic coding method for traffic condition monitoring as described in claim 3, characterized in that, For discriminator network 2 The loss function is as follows: In the formula, The sigmoid function is used; through backpropagation of the neural network, the parameters of the discriminator network 2 are optimized so that the output of the network can gradually approach the privacy mutual information between the original and reconstructed information.

6. The semantic coding method for traffic condition monitoring as described in claim 3, characterized in that, For data encoding / decoding networks The loss function is as follows: In the formula, The constraint value for integrated mutual information, This refers to the constraint value for privacy information.

7. The semantic coding method for traffic condition monitoring as described in claim 1, characterized in that, In step 3, the data encoding / decoding network Discriminator Network 1 and discriminator network 2 They are interconnected and trained sequentially and cyclically using a Generative Adversarial Network (GAN) approach; Specifically, in each round of training: Fixed discriminator network 1 and discriminator network 2, updated data encoding / decoding network The parameters are minimized to obtain the total loss function. To simultaneously satisfy reconstruction distortion, semantic accuracy, comprehensive mutual information constraints, and privacy mutual information constraints; Then, with the data encoding / decoding network fixed, discriminator network 1 and discriminator network 2 are trained separately to improve their ability to assess comprehensive mutual information. and privacy information The accuracy of the estimation; The training is performed iteratively through the above process until all four metrics—distortion, semantic accuracy, comprehensive mutual information, and privacy mutual information—reach the predetermined constraint standards, at which point the model converges.

8. The semantic coding method for traffic condition monitoring as described in claim 1, characterized in that, Step 4 specifically involves training... The data encoding and decoding network is integrated into edge devices or servers and encapsulated as a standard API interface to support real-time access to data streams from traffic perception systems. The compressed representation output by the semantic encoder is used in the uplink communication of data in intelligent transportation systems, saving bandwidth while improving semantic expression capabilities, and enabling further analysis and decision-making in the cloud.

Images