Collaborative vehicle perception-oriented adaptive semantic feature compression and transmission method

Abstract

The invention discloses a cooperative vehicle perception-oriented adaptive semantic feature compression and transmission method, and the method comprises the steps: representing the distribution density of task related information in a space through a confidence map generated by a cooperative vehicle detection network, and screening key semantic features at different spatial positions according to the distribution density; and priority compression and reliable transmission of a high-value region are realized, so that the redundant burden of an irrelevant region is reduced. Meanwhile, semantic feature compression and digital communication coding are jointly optimized in an end-to-end mode, so that the compression strength and redundancy are adaptively adjusted under different channel conditions, and the sensing performance of the cooperative vehicle in a limited bandwidth and noise channel environment is improved. The communication efficiency and the target detection precision of the cooperative vehicle system are effectively improved, and the method can be widely applied to multi-vehicle cooperative sensing tasks in intelligent traffic and automatic driving scenes.

Description

Technical Field

[0001] This invention relates to the fields of collaborative perception and semantic communication, and in particular to an end-to-end adaptive semantic feature compression and transmission method for collaborative vehicle perception. Background Technology

[0002] In the context of intelligent machine communication, communication needs are evolving from point-to-point information exchange between single devices to task-oriented multi-machine collaboration. Machine collaboration refers to multiple intelligent agents with sensing, communication, and control capabilities working together to complete processes such as perception fusion, task allocation, and decision execution based on shared information. For example, in the vehicle-to-everything (V2X) scenario, collaborative perception and decision-making among multiple vehicles can overcome the physical limitations of a single vehicle's perspective, enabling the completion of information in obscured areas and real-time coordination of multi-vehicle paths. Machine collaboration has become a crucial foundational capability supporting the operation of future large-scale intelligent systems and a key indicator of the transition from "individual intelligence" to "swarm intelligence" in intelligent systems. To achieve efficient collaborative perception, several feature fusion methods have recently been proposed, such as V2VNet, V2X-ViT, When2Com, and Where2Com. Among them, the literature [R.Xu,H.Xiang,Z.Tu,X.Xia,M.-H.Yang,and J.Ma,“V2x-vit:Vehicle-to-everythingcooperative perception with vision transformer,”in European conference on computer vision,pp.107–124,Springer,2022.] adopts a heterogeneous multi-agent self-attention mechanism to address the problem of device heterogeneity; the literature [Y.Hu,S.Fang,Z.Lei,Y.Zhong,and S.Chen,“Where2comm:Communication-efficient collaborative perception via spatial confidencemaps,”Advances in Neural Information Processing Systems,vol.35,pp.4874–4886,2022.] minimizes bandwidth requirements through a spatial confidence perception system.

[0003] Traditional communication systems prioritize accurate data reproduction, focusing primarily on lossless reconstruction of the information source at the receiving end. This design struggles to meet the demands of task-oriented machine collaboration. In complex and dynamic scenarios, the true purpose of communication is not precise data restoration, but rather supporting information exchange and task execution between intelligent agents. Therefore, communication systems urgently need to shift from "data restoration" to "task service," introducing information representation models and performance metrics that characterize task-related semantic value. Through methods such as compressed sensing, feature extraction, and end-to-end optimization, they can achieve efficient encoding and compact transmission of key semantics, ensuring the response speed and accuracy of collaborative tasks under limited bandwidth conditions.

[0004] Current communication systems generally adopt a separate design of "source coding and channel coding". While this architecture has clear module boundaries and a good implementation foundation in traditional theory and engineering practice, it neglects the global optimality of both under the given task objective. This makes it prone to a "cliff effect" under low signal-to-noise ratio and low bandwidth conditions, where performance drops sharply as channel conditions deteriorate. In recent years, works such as [E. Bourtsoulatze, DB. Kurka, and D. Gündüz, "Deep joint source-channel coding for wireless image transmission," IEEE Transactions on Cognitive Communications and Networking, vol.5, no.3, pp.567–579, 2019.] have attempted to achieve end-to-end joint optimization of source and channel coding using deep neural networks, demonstrating performance advantages under simulation conditions. However, these schemes typically rely on entirely new coding architectures, have poor compatibility with existing communication protocols and hardware systems, and are difficult to deploy in practice.

[0005] The above problems are particularly prominent in machine-cooperative semantic communication systems for real-time tasks. Traditional semantic communication models often use average performance metrics as optimization targets, making them suitable for scenarios with low real-time requirements. However, in safety-critical applications such as connected vehicles, it is more necessary to ensure the lower limit of system performance and avoid security risks caused by performance crashes at individual moments. Therefore, it is necessary to design a communication mechanism that balances task-driven approaches, bitrate constraints, and reliability for real-time machine-cooperative scenarios, breaking through the limitations of existing methods in terms of architecture and coding strategies. Summary of the Invention

[0006] This invention addresses machine collaboration scenarios characterized by limited bandwidth, high task complexity, and stringent timeliness requirements. It proposes an adaptive semantic feature compression and transmission method for collaborative vehicle perception to solve the problem that existing technologies struggle to simultaneously balance system performance, real-time performance, and deployability.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] An adaptive semantic feature compression and transmission method for cooperative vehicle perception includes the following steps:

[0009] S1. Extract semantic features from raw sensor data, and perform region selection and transmission filtering of semantic features based on confidence maps;

[0010] S2. Source encoding and decoding based on ROI and entropy encoding and decoding: adaptive compression, entropy encoding and decoding of selected semantic features;

[0011] S3, Digital Channel Coding and Channel Modeling;

[0012] S4. The semantic features of the receiving side are fused and reconstructed, and target detection and perceptual reasoning are performed based on the fused features;

[0013] S5. Jointly optimize the deep neural network parameters, the proportion of Region of Interest (ROI), and the channel code rate of each collaborator to minimize end-to-end task distortion while satisfying the constraint on average channel occupancy to construct an optimization problem.

[0014] S6. Estimate the maximum reconstruction distortion of the semantic features output by different collaboration models during end-to-end transmission and reconstruction, and calculate the upper bound of end-to-end task distortion.

[0015] S7. An adaptive source-channel joint optimization algorithm is used to solve the optimization problem. The algorithm mainly includes two stages: 1) model lookup table construction; 2) joint control of model selection and channel coding rate. By executing the algorithm, the optimal model for each channel is found. The smallest one is taken as the optimal model, and the corresponding optimal channel coding rate is calculated.

[0016] Furthermore, the specific process of step S1 includes:

[0017] Let X ego and X j Let f represent the lidar point cloud data collected by the autonomous vehicle and the j-th collaborator, respectively, where j = 1, 2, ..., J. All vehicles use a semantic feature encoder f with shared parameters. enc (·), whose trainable parameter is μ, to generate a bird's-eye view semantic feature map:

[0018] S ego =f enc (X ego ;μ),S j =f enc (X j ;μ)

[0019] The autonomous vehicle generates a spatial confidence map through a detection network, selects the top r% of high-confidence elements and sets them to 1, and sets the remaining elements to 0, thereby obtaining a binarized ROI mask M. j Where r>0 represents the ROI ratio; autonomous vehicles directly set Y. ego =S ego .

[0020] Furthermore, the specific process of step S2 includes:

[0021] Select feature map Y j =S j ⊙M j Y j Through analytical transformation g with learnable parameter φ a (·) is mapped to latent variables:

[0022] F j =g a (Y j ;φ)

[0023] Quantification was then performed. And using the probability mass function Entropy encoding is performed to obtain a bitstream. Where L j Indicates the bit length; the expected code rate of the source is calculated as follows: The elements of the ROI mask are only 0 or 1, and the 1s in the mask are concentrated in the spatial location of the detected object. Run-length encoding is used for M. j The compressed run length is used as a new information source, and then entropy coding is performed to obtain... Its source code rate

[0024] Autonomous vehicles at the receiving end Decode to recover M j and utilize right Decode to obtain The autonomous vehicle uses a synthesis transformation g with parameter ψ. s (·) Reconstruct the feature map of the collaborators to obtain:

[0025] Furthermore, the specific process of step S3 includes:

[0026] After source coding is completed, bit stream B j The data is divided into fixed-length blocks, each containing k information bits; for each block, a (k,n) block code is used, with a channel coding rate of [missing information]. Used to generate n channel symbols; the formula for calculating the total number of packets is: in This indicates a round-up operation; the bandwidth ratio is defined as... K is used to measure the number of channel uses required to transmit one semantic feature map element. S S represents j The number of elements in the middle.

[0027] Furthermore, the specific process of step S4 includes:

[0028] semantic features from various collaborators After decoding and reconstruction, the autonomous vehicle uses a non-parametric feature fusion module to integrate the reconstructed semantic features from collaborators. With its own semantic features Y ego The fusion is performed to generate the fused semantic representation G:

[0029]

[0030] Finally, the autonomous vehicle uses a semantic decoder with learnable parameters ω to convert the fused semantic representation G into a task-related output.

[0031]

[0032] The semantic decoder is used to implement the 3D target detection task.

[0033] Furthermore, the specific process of constructing the optimization problem in step S5 includes:

[0034] For the j-th collaborator, its bandwidth constraint is expressed as:

[0035]

[0036] Where, d j k represents the channel usage budget allocated to this collaborator. CRC k represents the length of the CRC in the case of polar codes. CRC =24, k in the case of random encoding CRC =0, This represents the channel code rate of the ROI mask when the block error rate is guaranteed to be below the threshold ρ0;

[0037] A unified channel allocation strategy is adopted, meaning that all collaborators have the same channel occupancy budget: d j =d; based on approximation relationship The optimization problem can be expressed as:

[0038]

[0039] in Φ = {μ, φ, θ, ψ, ω} represents the parameters of the deep neural network, and r is the proportion of the ROI. The channel code rate for each collaborator.

[0040] Furthermore, the training objective of the model in step S6 is:

[0041]

[0042] Where L task(s) β is a hyperparameter representing the distortion of cooperative sensing tasks under channelless conditions;

[0043] A ROI reconstruction regularization term is introduced during training, and the overall loss function is defined as follows:

[0044]

[0045] Where λ is a weighting coefficient balancing task accuracy and feature fidelity, and the expectation term... Estimation was performed using Monte Carlo sampling;

[0046] During training, the ROI proportion r is uniformly sampled, i.e., r∈[r] min ,r max ].

[0047] Furthermore, step S6, calculating the end-to-end reconstruction distortion, includes:

[0048] End-to-end reconstruction distortion is defined as:

[0049]

[0050] Distortion is decomposed into the source-induced part and the channel-induced part, and defined as D respectively. s and D c Among them, the channel-induced distortion term The upper bound is represented as

[0051]

[0052] Furthermore, step S6, calculating the upper bound of end-to-end task distortion, includes:

[0053] Under ideal error-free transmission conditions, the baseline task loss is defined based on the target detection results output by the cooperative sensing system; source distortion is defined as... in This represents the detection result under ideal error-free transmission conditions. Reconstruct features for the corresponding collaborators;

[0054] Under the condition of channel bit error, based on the fusion network f fuse(·), decoder f dec (·) and detection loss function The Lipschitz continuity of the given property, and let its Lipschitz constant be C. ω ,have to:

[0055]

[0056] The upper bound of the difference between the actual task loss and the baseline task loss is represented as an additive term related to the collaborative feature reconstruction error, thus obtaining the upper bound of the end-to-end task distortion:

[0057]

[0058] Furthermore, the adaptive source-channel joint optimization algorithm in step S7 specifically includes:

[0059] 1) Model lookup table construction

[0060] First, predefine a set of optional ROI ratios. For each (β, r), where Estimate the feature source coding rate R on the validation set respectively. s ROI mask coding rate R s,mask Task loss caused by information source and constant terms With C ω ;in, With C ω The upper bounds of the feature reconstruction error and the task loss caused by the channel are obtained by fitting the linear regression. After this estimation process, the corresponding models of each (β,r) will be stored in a lookup table for subsequent model selection and channel code rate control.

[0061] 2) Joint control of model selection and channel coding rate

[0062] A joint algorithm for model selection and channel coding rate control is used to find the optimal model (β) in a lookup table. * The task with the minimum loss is r0, and the specific process is as follows:

[0063] Algorithm 1: Model Selection and Encoding Rate Control

[0064] (1) Establish a model lookup table and obtain all models that can achieve reliable transmission under the current channel;

[0065] (2) Substitute the parameters according to the channel coding method used to estimate the upper bound of end-to-end semantic distortion.

[0066] (3) Find the models in each model The smallest one is taken as the optimal model, and the corresponding... As the optimal channel coding rate;

[0067] Or, when part of the link is saturated to P max When the truncated channel inversion is triggered, the system freezes the target model parameter β selected in step (1). * The ROI mask ratio is dynamically adjusted for each collaborator only to adapt to channel fluctuations; this process is executed in each frame, and the specific flow is as described in Algorithm 2:

[0068] Algorithm 2: Adaptive Adjustment of ROI Ratio

[0069] (1) Based on the current instantaneous signal-to-noise ratio γ j Identify all cases where the target signal-to-noise ratio γ cannot be achieved. t vehicle collection And calculate the channel capacity C for each collaborator. j =log2(1+γ) j );

[0070] (2) For each power-constrained collaborator j, keep the model parameter β = β * Remain unchanged, generate a candidate ratio set

[0071] (3) For each candidate proportion In the corresponding channel capacity C j The feasibility check step for reusing Algorithm 1 is performed; for the candidate proportions that pass the check, their optimal coding rate is calculated. And estimate the corresponding task losses.

[0072] (4) Among all feasible ratios, select the ROI ratio that minimizes the task loss.

[0073]

[0074] (5) Set the model of collaborator j as follows: And adopt the corresponding optimal coding rate If no proportion meets the conditions, the data transmitted by that collaborator is discarded.

[0075] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0076] This invention proposes an adaptive semantic feature compression and transmission method for cooperative vehicle perception. It utilizes a confidence map generated by a collaborating vehicle detection network to characterize the distribution density of task-related information in space, and uses this map to select key semantic features at different spatial locations. This prioritizes compression and reliable transmission of high-value areas, thereby reducing redundancy in irrelevant areas. Simultaneously, this invention employs end-to-end joint optimization of semantic feature compression and digital communication coding, adaptively adjusting compression strength and redundancy under different channel conditions to improve the perception performance of cooperative vehicles in environments with limited bandwidth and noisy channels. This invention effectively improves the communication efficiency and target detection accuracy of cooperative vehicle systems and can be widely applied to multi-vehicle cooperative perception tasks in intelligent transportation and autonomous driving scenarios.

[0077] Furthermore, this invention was simulated on the cooperative sensing dataset OPV2V and compared with models such as source-channel joint coding and separate coding. The results show that this invention significantly improves task detection accuracy and communication efficiency under the same bandwidth; it surpasses the PointPillar method for transmitting raw data under high SNR conditions. It maintains stable sensing performance in fast time-varying channel scenarios, with the adaptive ROI strategy effectively improving the sensing reliability of frames affected by Channel Truncation Inversion (TCI). Attached Figure Description

[0078] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0079] Figure 1 The diagram shows a collaborative perception task scenario and a system architecture diagram provided by this invention.

[0080] Figure 2 The flowchart illustrates the spatial feature selection, source encoding, and source decoding processes performed by collaborators in this invention.

[0081] Figure 3 The model selection results under different SNRs provided by this invention.

[0082] Figure 4 The present invention provides end-to-end task distortion and block error rate under different SNRs.

[0083] Figure 5 The detection accuracy of the present invention and the comparative scheme is compared under different signal-to-noise ratios with AP@0.5 average bandwidth.

[0084] Figure 6The detection accuracy of the present invention and the comparative scheme is compared under different signal-to-noise ratios with AP@0.5 average bandwidth.

[0085] Figure 7 The average bandwidth ratio of AP@0.5 at a signal-to-noise ratio of 4dB provided by this invention is compared with the detection performance of the present invention and the comparative scheme.

[0086] Figure 8 The average bandwidth ratio of AP@0.7 at a signal-to-noise ratio of 4dB provided by this invention is compared with the detection performance of the present invention and the comparative scheme.

[0087] Figure 9 This is a bird's-eye view of the vehicle detection results of the method and comparative scheme of the present invention. Detailed Implementation

[0088] The main contents of this invention include:

[0089] 1. Semantic compression and selective transmission mechanism for collaborative computing in vehicle-to-everything (V2X) networks

[0090] Starting from typical vehicle-to-everything (V2X) collaborative perception and decision-making scenarios, this paper analyzes the redundant structure of multi-vehicle perception results in the spatiotemporal dimensions and constructs a representation model for task-related semantic information. Based on this, a semantic compression and selective transmission mechanism adapted to bandwidth limitations is designed, prioritizing the encoding and transmission of only key targets, key areas, and key moments to achieve efficient fusion of multi-vehicle perception results and reduce the communication resources occupied by redundant information.

[0091] 2. Task-driven communication optimization goal and resource mapping relationship modeling

[0092] Based on semantic representation and collaborative perception models, we further abstract and summarize task-oriented communication optimization objectives, construct the mapping relationship between task completion (such as detection accuracy, decision correctness, and vehicle safety indicators) and communication resources (such as bit rate, bandwidth, and latency budget), and reveal the trade-off mechanism between the importance of different semantic information and communication costs and task benefits in the process of multi-node information interaction, providing theoretical guidance and optimization criteria for system design.

[0093] 3. Deployable structurally weakly coupled joint optimization scheme

[0094] To address the poor compatibility of existing end-to-end deep joint coding schemes with engineering systems, this paper proposes a structurally weakly coupled coding and transmission model guided by multi-user machine collaborative optimization theory. While maintaining the traditional "source-channel" module boundary, an artificial intelligence model is introduced to jointly optimize feature extraction, semantic compression, and task-related rate allocation. An unequal error protection mechanism is designed at the channel coding level, and at the receiver, decoding correctness and confidence information are combined for collaborative fusion decision-making. Through these methods, task completion and system reliability under low signal-to-noise ratio conditions are improved without fundamentally altering existing communication protocols and hardware architectures.

[0095] Through the above design, this invention can effectively improve the bandwidth utilization efficiency and task performance lower limit of machine collaborative semantic communication systems for real-time tasks while ensuring the integrability of existing communication systems, and provide an engineering-applicable communication and collaboration solution for large-scale swarm intelligence systems in typical scenarios such as vehicle networking.

[0096] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0097] This invention first constructs a collaborative perception framework based on feature extraction, such as... Figure 1 As shown, the collaborative perception framework includes a semantic feature extraction and selection module, a ROI-based source codec, an entropy encoder, a digital channel codec, a semantic information fusion module, and a target detection decoder.

[0098] Based on this framework, the adaptive semantic feature compression and transmission method for cooperative vehicle perception proposed in this invention includes the following steps:

[0099] S1. Extract semantic features from raw sensor data, and perform region selection and transmission filtering of semantic features based on confidence maps.

[0100] The specific process of step S1 includes:

[0101] Let X ego and X j Let f represent the lidar point cloud data collected by the autonomous vehicle and the j-th collaborator, respectively, where j = 1, 2, ..., J. All vehicles use a semantic feature encoder f with shared parameters. enc (·), whose trainable parameter is μ, to generate a bird's-eye view semantic feature map:

[0102] S ego =f enc (X ego ;μ),S j =f enc (X j ;μ)

[0103] The autonomous vehicle generates a spatial confidence map through a detection network, selects the top r% of high-confidence elements and sets them to 1, and sets the remaining elements to 0, thereby obtaining a binarized ROI mask M. j Where r>0 represents the ROI ratio; autonomous vehicles do not perform this step and directly set Y. ego =S ego .

[0104] S2. Source encoding and decoding based on ROI and entropy encoding and decoding: adaptive compression, entropy encoding and decoding of selected semantic features.

[0105] The specific process of step S2 includes:

[0106] Select feature map Y j =S j ⊙M j Y j Through analytical transformation g with learnable parameter φ a (·) is mapped to latent variables:

[0107] F j =g a (Y j ;φ)

[0108] Quantification was then performed. And using the probability mass function Entropy encoding is performed to obtain a bitstream. Where L j Indicates the bit length; the expected code rate of the source is calculated as follows: Since the elements of the ROI mask are only 0 or 1, and the 1s in the mask are concentrated in the spatial location of the detected object, run-length encoding is used for M. j The compressed run length is used as a new information source, and then entropy coding is performed to obtain... Its source code rate

[0109] Autonomous vehicles at the receiving end Decode to recover M j and utilize right Decode to obtain To adapt to various downstream tasks and avoid overfitting to specific tasks, the autonomous vehicle employs a synthetic transformation g with parameter ψ. s (·) Reconstruct the feature map of the collaborators to obtain:

[0110] S3, Digital Channel Coding and Channel Modeling.

[0111] The specific process of step S3 includes:

[0112] After source coding is completed, bit stream B j The data is divided into fixed-length blocks, each containing k information bits; for each block, a (k,n) block code is used, with a channel coding rate of [missing information]. Used to generate n channel symbols; since each symbol may carry multiple bits (e.g., 2 bits in QPSK modulation), for higher-order modulation schemes, R c The value may be greater than 1.

[0113] The formula for calculating the total number of groups is: in This indicates a round-up operation; the bandwidth ratio is defined as... K is used to measure the number of channel uses required to transmit one semantic feature map element. S S represents j The number of elements in the middle.

[0114] After channel coding is completed, the collaborator transmits the constellation points to the vehicle via the digital channel. The relationship between the constellation points transmitted by the collaborator and the constellation points received by the vehicle in the digital channel is as follows:

[0115]

[0116] in and Let m be the m-th constellation point in the i-th code block sent by collaborator j, and let m be the corresponding receiving constellation point. This represents the complex channel coefficients of the link between collaborator j and the vehicular vehicle. This represents the average symbol energy sent by collaborator j. The average power is It is a circularly symmetric complex Gaussian noise that is independently and identically distributed between symbols.

[0117] In this invention, it is assumed that the channel between all potential collaborators and the voluntary vehicle has quasi-static characteristics, that is, the channel coefficients remain unchanged in all code blocks within any frame, but can change between frames. Simultaneously, this study assumes that the collaborators can obtain completely accurate Channel State Information (CSI). Based on this, the collaborators will use CSI to perform Truncated Channel Inversion (TCI) power control. The specific process is as follows: (1) The voluntary vehicle pre-determines that the received signal-to-noise ratio (SNR) of all collaborators' transmitted signals must reach a predetermined value γ. (2) Based on the predetermined SNR and the channel coefficients of the current frame, each potential collaborator calculates the required average symbol energy of the constellation points. (3) If the required energy Not exceeding the collaborator's energy limit Then potential collaborators are activated, and according to... If the required energy exceeds the limit, potential collaborators will not transmit features in the current frame; conversely, if the required energy exceeds the limit, potential collaborators will not transmit features in the current frame. Therefore, it is assumed that the total number of collaborators eventually activated is J≤J0.

[0118] Finally, the digital channel decoder located on the vehicle restores the received constellation points from each active cooperator to the corresponding bit sequence through digital demodulation and channel decoding, obtaining... and

[0119] Errors in decoding the ROI mask can severely impact the source decoding of features. To ensure correct reconstruction of the ROI mask, this invention employs Unequal Error Protection (UEP) during channel coding of both the ROI mask and features. Specifically, the channel coding rate for the ROI mask is guaranteed to have a block error rate below a pre-set threshold ρ0 = 1e-5, thereby minimizing the possibility of errors; however, no strict requirements are placed on the block error rate of the information code rate used for the features.

[0120] This invention considers two channel coding schemes: random coding, as a theoretical benchmark; and polar coding, as a practical implementation scheme in the vehicle-to-everything (V2X) scenario.

[0121] The process of collaborators performing spatial feature selection, source coding, and source decoding is as follows: Figure 2 As shown.

[0122] S4. The semantic features of the receiving side are fused and reconstructed, and target detection and perceptual reasoning are performed based on the fused features.

[0123] The specific process of step S4 includes:

[0124] semantic features from various collaborators After decoding and reconstruction, the autonomous vehicle uses a non-parametric feature fusion module to integrate the reconstructed semantic features from collaborators. With its own semantic features Y ego The fusion is performed to generate the fused semantic representation G:

[0125]

[0126] Finally, the autonomous vehicle uses a semantic decoder with learnable parameters ω to convert the fused semantic representation G into a task-related output.

[0127]

[0128] The semantic decoder is used to implement the 3D target detection task.

[0129] S5. Jointly optimize the deep neural network parameters, the proportion of Region of Interest (ROI), and the channel code rate of each collaborator to minimize end-to-end task distortion while satisfying the constraint on average channel occupancy to construct an optimization problem.

[0130] Specifically, the collaborative perception framework accepts 3D point cloud data as input, converts it into BEV (Browser-Eye View) through discretization, and then uses an anchor-based PointPillar method to obtain feature maps of the BEV perspective from the 3D point cloud data through a feature encoder.

[0131] Definition and approximation of the optimization problem:

[0132] In collaborative perception, the goal is to maximize the perceived truth value O and the features from collaborators. And the characteristics of autonomous vehicles Y ego Mutual information between Mutual information is defined as follows:

[0133]

[0134] H(O) is a constant and can be ignored during the optimization process.

[0135] To make the optimization process feasible, the conditional entropy term is relaxed, and the following variational upper bound is minimized.

[0136]

[0137] Where, O = (O cls O loc O dir ) represents the real detection target used for classification, localization, and orientation; q represents the variational approximation distribution generated by the semantic fusion module and the semantic decoder.

[0138] For the classification part O cls With direction part O dir Its output distribution and It is modeled as a multinomial distribution over discrete categories, which is equivalent to minimizing the cross-entropy loss with the true label. For the localization part f loc We assume that the prediction error follows an independent Laplace distribution in each dimension. Specifically, the conditional distribution is modeled as follows:

[0139]

[0140] in, This indicates the predicted mean location, and b is the scale parameter.

[0141] therefore, Loss can be generated from 3D target detection Approximately:

[0142]

[0143] in, and Focal Loss is used. Smooth L1 Loss is used. and These represent the corresponding prediction results.

[0144] In the above equation, end-to-end task distortion depends on the joint distribution. This distribution is implicitly determined by factors such as deep neural network parameters, channel transmission rate, and signal-to-noise ratio, thus its closed-form analytical expression cannot be obtained. To facilitate subsequent optimization of the proposed framework, this invention first derives an upper bound on end-to-end reconstruction distortion, and then further derives an upper bound on end-to-end task distortion based on this.

[0145] Step S5 involves constructing the optimization problem, specifically including:

[0146] For the j-th collaborator, its bandwidth constraint is expressed as:

[0147]

[0148] Where, d j k represents the channel usage budget allocated to this collaborator. CRC k represents the length of the CRC in the case of polar codes. CRC =24, in the case of random encoding CRC =0, This represents the channel code rate of the ROI mask when the block error rate is guaranteed to be below the threshold ρ0.

[0149] This invention employs a unified channel allocation strategy, meaning that all collaborators have the same channel occupancy budget: d j =d; based on approximation relationship The optimization problem can be expressed as:

[0150]

[0151] in Φ = {μ, φ, θ, ψ, ω} represents the parameters of the deep neural network, and r is the proportion of the ROI. The channel code rate for each collaborator.

[0152] S6. Estimate the maximum reconstruction distortion of the semantic features output by different collaboration models during end-to-end transmission and reconstruction, and calculate the upper bound of end-to-end task distortion.

[0153] The transmitting end uses a pre-built semantic distortion estimation module to estimate the maximum reconstruction distortion of semantic features output by different cooperative models during end-to-end transmission and reconstruction. Based on this, it provides the corresponding upper bound of end-to-end task distortion, providing a constraint basis for subsequent adaptive semantic feature compression and channel resource allocation.

[0154] The process of calculating end-to-end reconstruction distortion includes:

[0155] End-to-end reconstruction distortion is defined as:

[0156]

[0157] Distortion is decomposed into the source-induced part and the channel-induced part, and defined as D respectively. s and D c D s D c The calculation method is as follows:

[0158]

[0159] Among them, the distortion caused by the channel The upper bound is represented as

[0160]

[0161] The process of calculating the upper bound of end-to-end task distortion includes:

[0162] Under ideal error-free transmission conditions, the baseline task loss is defined based on the target detection results output by the cooperative sensing system; source distortion is defined as... in This represents the detection result under ideal error-free transmission conditions. Reconstruct features for the corresponding collaborators;

[0163] Under the condition of channel bit error, based on the fusion network f fuse (·), decoder f dec (·) and detection loss function The Lipschitz continuity of the given property, and let its Lipschitz constant be C. ω ,have to:

[0164]

[0165] The upper bound of the difference between the actual task loss and the baseline task loss is represented as an additive term related to the collaborative feature reconstruction error, thus obtaining the upper bound of the end-to-end task distortion:

[0166]

[0167] To efficiently solve optimization problems, this invention first extends the DIB theory to the training of end-to-end neural network models without considering channel errors. Subsequently, an adaptive source-channel joint optimization algorithm is designed to select the optimal model and channel code rate based on the channel signal-to-noise ratio.

[0168] This invention employs a three-stage training strategy to train a set of deep neural network models to balance the task distortion caused by source compression. With source coding rate R s The trade-off between them.

[0169] Phase 1: Skip the feature selection and source coding modules (i.e., set...) The parameters {μ,ω} of the training detection backbone network are used to optimize the following:

[0170] Phase Two: Assuming no channel error Under the given conditions, with {φ,θ,ψ} fixed, the feature selection and source coding related module {μ,ω} is trained.

[0171] The third stage: Based on the pre-training of the first two stages, end-to-end joint optimization is performed on the entire network parameter set Φ.

[0172] In the second and third phases of training, an optimization objective based on DIB theory is adopted to achieve a joint balance between the source code rate and the task loss caused by the source:

[0173]

[0174] Where j=0 corresponds to an autonomous vehicle, satisfying and

[0175] For ease of solution, mutual information terms Approximating the source code rate, i.e. At the same time, the conditional entropy term Replace with the expected detection loss, i.e. Furthermore, to ensure end-to-end differentiability, the discrete quantization process is relaxed using additive uniform noise, i.e. in

[0176] The training objective of the model in this invention is:

[0177]

[0178] Where L task(s) β is a hyperparameter representing the distortion of cooperative sensing tasks under channelless conditions.

[0179] To enhance the model's generalization ability and prevent overfitting, an ROI reconstruction regularization term is introduced during training. The overall loss function is defined as follows:

[0180]

[0181] Where λ is a weighting coefficient balancing task accuracy and feature fidelity, and the expectation term... Estimation was performed using Monte Carlo sampling.

[0182] To improve the robustness of the model under different ROI ratios, the ROI ratio r is uniformly sampled during training, i.e., r∈[r min ,r max ].

[0183] S7. An adaptive source-channel joint optimization algorithm is used to solve the optimization problem. The algorithm mainly includes two stages: 1) model lookup table construction; 2) joint control of model selection and channel coding rate. By executing the algorithm, the optimal model for each channel is found. The smallest one is taken as the optimal model, and the corresponding optimal channel coding rate is calculated.

[0184] The adaptive source-channel joint optimization algorithm proposed in this invention is used to solve the defined optimization problem, specifically including:

[0185] 1) Model lookup table construction

[0186] First, predefine a set of optional ROI ratios. For each (β, r), where Estimate the feature source coding rate R on the validation set respectively. s ROI mask coding rate R s,mask Task loss caused by information source and constant terms With C ω ;in, With C ω The upper bounds of the feature reconstruction error and the task loss caused by the channel are obtained by fitting the linear regression. After this estimation process, the corresponding models of each (β,r) will be stored in a lookup table for subsequent model selection and channel code rate control.

[0187] 2) Joint control of model selection and channel coding rate

[0188] After determining the deep neural network parameters Φ and the ROI ratio r, the original optimization problem can be simplified to solving the following optimization problem independently for each cooperative vehicle j:

[0189]

[0190] At a smaller block error rate ρ j Under certain conditions, approximate relationships can be used. This simplifies the original optimization problem to:

[0191] in, The available characteristic transmission channel budget is defined as follows: It is calculated based on the theoretical formula for block error rate under the condition of fixed block error rate ρ = ρ0.

[0192] Model Feasibility Check: To determine the feasibility of a candidate model under given channel conditions, this study further proposes feasibility judgment criteria. To ensure reliable transmission of both the ROI mask and features, the following conditions must be met:

[0193]

[0194] The proof is as follows: (i) According to Shannon's channel coding theorem, reliable transmission requires that the coding rate does not exceed the channel capacity, i.e. Therefore, we can conclude that... (ii) Considering the block length constraint, each feature must occupy at least one channel resource of length n, which can be obtained by rounding. In summary, the model is feasible when the conditions in the formula are met.

[0195] Optimal channel coding rate: Furthermore, based on different channel coding methods, the optimal coding rate for the problem can be obtained. when When the random encoding is large enough, we have:

[0196]

[0197] For polar coding

[0198]

[0199] Since all cooperating vehicles use the same allocation strategy, the subscript j is omitted in the formula.

[0200] The proposed model selection and channel coding rate joint control algorithm is used to find the optimal model (β) in a lookup table. * The task with the minimum loss is r0, and the specific process is as follows:

[0201] Algorithm 1: Model Selection and Encoding Rate Control

[0202] (1) Establish a model lookup table and obtain all models that can achieve reliable transmission under the current channel;

[0203] (2) Substitute the parameters according to the channel coding method used to estimate the upper bound of end-to-end semantic distortion.

[0204] (3) Find the models in each model The smallest one is taken as the optimal model, and the corresponding... As the optimal channel coding rate;

[0205] Optionally, when a portion of the link becomes power saturated to P max When the truncated channel inversion is triggered, the system freezes the target model parameter β selected in step (1). * The ROI mask ratio is dynamically adjusted for each collaborator only to adapt to channel fluctuations; this process is executed in each frame, and the specific flow is as described in Algorithm 2:

[0206] Algorithm 2: Adaptive Adjustment of ROI Ratio

[0207] (1) Based on the current instantaneous signal-to-noise ratio γ j Identify all cases where the target signal-to-noise ratio γ cannot be achieved. t vehicle collection And calculate the channel capacity C for each collaborator. j =log2(1+γ) j );

[0208] (2) For each power-constrained collaborator j, keep the model parameter β = β * Remain unchanged, generate a candidate ratio set

[0209] (3) For each candidate proportion In the corresponding channel capacity C j The feasibility check step for reusing Algorithm 1 is performed; for the candidate proportions that pass the check, their optimal coding rate is calculated. And estimate the corresponding task losses.

[0210] (4) Among all feasible ratios, select the ROI ratio that minimizes the task loss.

[0211]

[0212] (5) Set the model of collaborator j as follows: And adopt the corresponding optimal coding rate If no proportion meets the conditions, the data transmitted by that collaborator is discarded.

[0213] In summary, this invention proposes an adaptive semantic feature compression and transmission method (ASC-CP) for cooperative vehicle perception, which has the following significant advantages compared to existing perception-oriented communication systems:

[0214] 1. Stronger task relevance and higher compression efficiency

[0215] This invention constructs an ROI (Region of Interest) region selection mechanism, determines the importance of semantic features based on the target detection confidence, and performs adaptive compression and encoding only on the key regions of the task, thereby significantly reducing the transmission overhead of redundant features and achieving efficient collaborative perception under bandwidth-constrained conditions.

[0216] 2. End-to-end optimization is possible, and sensing performance is stable and reliable.

[0217] This invention derives an upper bound for end-to-end task distortion based on the Lipschitz continuity derivation of deep neural networks, and decomposes task-related distortion into two parts: the source end and the channel end. It can jointly optimize model selection, ROI compression ratio and channel coding parameters, effectively suppress the sensitivity of sensing results to channel noise, and improve the overall robustness and sensing accuracy of the system.

[0218] 3. Adaptive resource allocation, still advantageous under complex channel conditions.

[0219] This invention constructs a rate and model adaptive selection algorithm based on end-to-end task distortion. It can dynamically select the model structure, optimize the source-channel rate, and adjust the ROI coding ratio in real time under a given channel signal-to-noise ratio, enabling the system to adaptively adjust the transmission quality according to the instantaneous channel state and improve the availability performance in weak channel and fast fading scenarios.

[0220] 4. Highly deployable and compatible with existing communication systems.

[0221] This invention employs digital channel coding and a configurable source coding structure, supports unequal error protection strategies, is highly compatible with existing vehicle network communication protocols and hardware systems, can be deployed directly without reconstructing the communication stack, has low engineering implementation difficulty, and has the potential for large-scale deployment.

[0222] To verify the effectiveness of the end-to-end adaptive semantic feature compression and transmission method for cooperative vehicle perception of the present invention, the following verification comparative experiment was conducted.

[0223] 1. Simulation Experiment Setup

[0224] This invention is verified through simulation in a vehicle-to-everything (V2X) cooperative perception scenario. The experiment uses the OPV2V public dataset as the cooperative perception data source. This dataset contains point cloud and image information from multiple vehicle perspectives and provides 3D bounding box annotations for training and testing the cooperative perception model.

[0225] The wireless transmission section adopts an OFDM scheme based on the 3GPP standard. The carrier frequency is 5.9 GHz, the total bandwidth is 20 MHz, the subcarrier spacing is 30 kHz, and each physical resource block contains 12 subcarriers and 10 resource blocks to form a subchannel. Assuming that each cooperating vehicle occupies one subchannel in one transmission, the available bandwidth for a single user is approximately 3.6 MHz, and the symbol rate is approximately 3.6 MBaud. The maximum transmit power for a single user is calculated based on the total transmit power constraint. Noise figure and other RF parameters are consistent with existing V2V communication systems.

[0226] The deep source encoding part employs a factorized prior-based compressed network as the semantic feature encoder and decoder, and uses a PointPillars structure as the backbone network for 3D object detection. It maps the original LiDAR point cloud into a bird's-eye view semantic feature map with a size of 64×100×352 and a spatial resolution of 0.4m / pixel. All neural network models are implemented in the PyTorch framework, using the Adam optimizer to train multiple models with different trade-off parameters in stages, which are then used to construct a lookup table for subsequent adaptive model selection.

[0227] In both the proposed and separate schemes, the channel coding section considers two implementation forms: ideal random coding and polar codes, uniformly adopting a code length n = 1024. In the polar code implementation, the modulation scheme is adaptively selected based on the current signal-to-noise ratio (QPSK for high SNR, BPSK for low SNR), and the optimal channel code rate is calculated by a rate control algorithm, thereby constructing the corresponding polar code encoder and decoder.

[0228] To comprehensively evaluate the performance of this invention, ASC-CP is compared with three benchmark schemes: the first is an end-to-end deep JSCC scheme based on deep neural networks, including standard DJSCC without using ROI and DJSCC-ROI which only simulates transmission in the ROI region; the second is the traditional "source-channel separation" scheme, which, under the constraints of fixed channel code rate and bit error rate, first determines the ROI ratio by the feature selection module, then completes bit-level compression using a factorized prior model, and finally transmits through polar codes and QPSK modulation; the third is the PointPillars baseline scheme without feature selection and compression, which relies solely on the original semantic features for collaborative detection. Perceptual performance is evaluated using the average precision (AP) metric at different IoU thresholds.

[0229] Through the above experimental configuration, the present invention can be fairly compared with a variety of typical solutions in a unified vehicle network communication environment, verifying the comprehensive advantages of the present invention in terms of detection accuracy and communication efficiency.

[0230] 2. Model selection and distortion analysis under different signal-to-noise ratio conditions;

[0231] The invention was simulated under a fixed average bandwidth ratio. Figure 3 The variation trends of the system model selection parameter β and ROI ratio r under different signal-to-noise ratio conditions are presented. It can be seen that as the signal-to-noise ratio increases, the selected β value decreases step by step, while the ROI ratio gradually increases. This indicates that the system can adaptively improve feature fidelity and retain more key semantics of the task when the channel quality is improved. Figure 4 The paper presents the changes in end-to-end task distortion and block bit error rate with signal-to-noise ratio (SNR). It can be seen that task distortion exhibits a monotonically decreasing trend with SNR, and both source distortion and channel distortion are effectively suppressed. The block bit error rate fluctuates periodically with SNR, which reflects the mechanism of dynamic balancing of performance and reliability under the joint scheduling of model and channel bit rate in this invention.

[0232] 3. Performance Comparison and Resource Utilization Advantages

[0233] Under the constraints of fixed bandwidth and fixed signal-to-noise ratio, this invention is compared with the source-channel separation scheme, the deep JSCC scheme (including the ROI version), and the PointPillars baseline without communication constraints. Figure 5 and Figure 6 The detection accuracy of the present invention and the comparative scheme (AP@0.7 and AP@0.5) under different signal-to-noise ratios are shown. Figure 7 and Figure 8 The detection performance at different average bandwidth ratios under a 4dB signal-to-noise ratio (SNR) is demonstrated. Results show that this invention significantly alleviates the "cliff effect" of traditional split-mode schemes in low SNR regions and avoids premature performance saturation at high SNRs, exhibiting a smoother overall performance improvement curve. With similar bandwidth overhead, the performance achieved by this invention is close to or better than DJSCC-ROI, and significantly better than the baseline model of basic DJSCC and uncompressed collaboration. This difference indicates that explicitly introducing feature selection and communication budget during the training phase enables the encoder to learn semantic information more closely aligned with the target detection task. At the same detection accuracy, this invention can reduce channel occupancy by approximately 15%–25% (see...). Figure 7 – Figure 8 This reflects bandwidth utilization efficiency and resource saving capabilities.

[0234] Furthermore, Figure 9 The paper presents a comparison of the 3D detection results generated by various schemes under different signal-to-noise ratios (SNRs). It can be seen that the present invention outperforms the separate and deep JSCC coding methods at the same bandwidth in terms of both the number of detection boxes and the localization integrity at a 4dB SNR, while maintaining a stable advantage at 10dB. This demonstrates that the present invention possesses good robustness and scalability in low-bandwidth and time-varying channel scenarios.

[0235] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive semantic feature compression and transmission method for cooperative vehicle perception, characterized in that, Includes the following steps: S1. Extract semantic features from raw sensor data, and perform region selection and transmission filtering of semantic features based on confidence maps; S2. Source encoding and decoding based on ROI and entropy encoding and decoding: adaptive compression, entropy encoding and decoding of selected semantic features; S3, Digital Channel Coding and Channel Modeling; S4. The semantic features of the receiving side are fused and reconstructed, and target detection and perceptual reasoning are performed based on the fused features; S5. Jointly optimize the deep neural network parameters, the proportion of Region of Interest (ROI), and the channel code rate of each collaborator to minimize end-to-end task distortion while satisfying the constraint on average channel occupancy to construct an optimization problem. S6. Estimate the maximum reconstruction distortion of the semantic features output by different collaboration models during end-to-end transmission and reconstruction, and calculate the upper bound of end-to-end task distortion. S7. An adaptive source-channel joint optimization algorithm is used to solve the optimization problem. The algorithm mainly includes two stages: 1) model lookup table construction; 2) joint control of model selection and channel coding rate. By executing the algorithm, the optimal model for each channel is found. The smallest one is taken as the optimal model, and the corresponding optimal channel coding rate is calculated.

2. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, The specific process of step S1 includes: Let X ego and X j Let f represent the lidar point cloud data collected by the autonomous vehicle and the j-th collaborator, respectively, where j = 1, 2, ..., J. All vehicles use a semantic feature encoder f with shared parameters. enc (·), whose trainable parameter is μ, to generate a bird's-eye view semantic feature map: S ego =f enc (X ego ;μ),S j =f enc (X j (m) The autonomous vehicle generates a spatial confidence map through a detection network, selects the top r% of high-confidence elements and sets them to 1, and sets the remaining elements to 0, thereby obtaining a binarized ROI mask M. j Where r>0 represents the ROI ratio; autonomous vehicles directly set Y. ego =S ego .

3. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, The specific process of step S2 includes: Select feature map Y j =S j ⊙M j Y j Through analytical transformation g with learnable parameter φ a (·) is mapped to latent variables: F j =g a (Y j ;f) Quantification was then performed. And using the probability mass function Entropy encoding is performed to obtain a bitstream. Where L j Indicates the bit length; the expected code rate of the source is calculated as follows: The elements of the ROI mask are only 0 or 1, and the 1s in the mask are concentrated in the spatial location of the detected object. Run-length encoding is used for M. j The compressed run length is used as a new information source, and then entropy coding is performed to obtain... Its source code rate Autonomous vehicles at the receiving end Decode to recover M j and utilize right Decode to obtain The autonomous vehicle uses a synthesis transformation g with parameter ψ. s (·) Reconstruct the feature map of the collaborators to obtain:

4. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, The specific process of step S3 includes: After source coding is completed, bit stream B j The data is divided into fixed-length blocks, each containing k information bits; for each block, a (k,n) block code is used, with a channel coding rate of [missing information]. Used to generate n channel symbols; the formula for calculating the total number of packets is: in This indicates a round-up operation; the bandwidth ratio is defined as... K is used to measure the number of channel uses required to transmit one semantic feature map element. S S represents j The number of elements in the middle.

5. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, The specific process of step S4 includes: semantic features from various collaborators After decoding and reconstruction, the autonomous vehicle uses a non-parametric feature fusion module to integrate the reconstructed semantic features from collaborators. With its own semantic features Y ego The fusion is performed to generate the fused semantic representation G: Finally, the autonomous vehicle uses a semantic decoder with learnable parameters ω to convert the fused semantic representation G into a task-related output. The semantic decoder is used to implement the 3D target detection task.

6. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, Step S5 involves constructing the optimization problem, specifically including: For the j-th collaborator, its bandwidth constraint is expressed as: Where, d j k represents the channel usage budget allocated to this collaborator. CRC k represents the length of the CRC in the case of polar codes. CRC =24, k in the case of random encoding CRC =0, This represents the channel code rate of the ROI mask when the block error rate is guaranteed to be below the threshold ρ0; A unified channel allocation strategy is adopted, meaning that all collaborators have the same channel occupancy budget: d j =d; based on approximation relationship The optimization problem can be expressed as: in Φ = {μ, φ, θ, ψ, ω} represents the parameters of the deep neural network, and r is the proportion of the ROI. The channel code rate for each collaborator.

7. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, The training objective of model S6 is: Where L task(s) β is a hyperparameter representing the distortion of cooperative sensing tasks under channelless conditions; A ROI reconstruction regularization term is introduced during training, and the overall loss function is defined as follows: Where λ is a weighting coefficient balancing task accuracy and feature fidelity, and the expectation term... Estimation was performed using Monte Carlo sampling; During training, the ROI proportion r is uniformly sampled, i.e., r∈[r] min ,r max ].

8. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, Step S6, calculating end-to-end reconstruction distortion, includes: End-to-end reconstruction distortion is defined as: Distortion is decomposed into the source-induced part and the channel-induced part, and defined as D respectively. s and D c Among them, the channel-induced distortion term The upper bound is represented as 9. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, Step S6, calculating the upper bound of end-to-end task distortion, includes: Under ideal error-free transmission conditions, the baseline task loss is defined based on the target detection results output by the cooperative sensing system; source distortion is defined as... in This represents the detection result under ideal error-free transmission conditions. Reconstruct features for the corresponding collaborators; Under the condition of channel bit error, based on the fusion network f fuse (·), decoder f dec (·) and detection loss function The Lipschitz continuity of the given property, and let its Lipschitz constant be C. ω ,have to: The upper bound of the difference between the actual task loss and the baseline task loss is represented as an additive term related to the collaborative feature reconstruction error, thus obtaining the upper bound of the end-to-end task distortion:

10. The adaptive semantic feature compression and transmission method for cooperative vehicle perception according to claim 1, characterized in that, Step S7, the adaptive source-channel joint optimization algorithm, specifically includes: 1) Model lookup table construction First, predefine a set of optional ROI ratios. For each (β, r), where Estimate the feature source coding rate R on the validation set respectively. s ROI mask coding rate R s,mask Task loss caused by information source and constant terms With C ω ;in, With C ω The upper bounds of the feature reconstruction error and the task loss caused by the channel are obtained by fitting the linear regression. After this estimation process, the corresponding models of each (β,r) will be stored in a lookup table for subsequent model selection and channel code rate control. 2) Joint control of model selection and channel coding rate A joint algorithm for model selection and channel coding rate control is used to find the optimal model (β) in a lookup table. * The task with the minimum loss is r0, and the specific process is as follows: Algorithm 1: Model Selection and Encoding Rate Control (1) Establish a model lookup table and obtain all models that can achieve reliable transmission under the current channel; (2) Substitute the parameters according to the channel coding method used to estimate the upper bound of end-to-end semantic distortion. (3) Find the models in each model The smallest one is taken as the optimal model, and the corresponding... As the optimal channel coding rate; Or, when part of the link is saturated to P max When the truncated channel inversion is triggered, the system freezes the target model parameter β selected in step (1). * The ROI mask ratio is dynamically adjusted for each collaborator only to adapt to channel fluctuations; this process is executed in each frame, and the specific flow is as described in Algorithm 2: Algorithm 2: Adaptive Adjustment of ROI Ratio (1) Based on the current instantaneous signal-to-noise ratio γ j Identify all cases where the target signal-to-noise ratio γ cannot be achieved. t vehicle collection And calculate the channel capacity C for each collaborator. j =log2(1+γ) j ); (2) For each power-constrained collaborator j, keep the model parameter β = β * Remain unchanged, generate a candidate ratio set (3) For each candidate proportion In the corresponding channel capacity C j The feasibility check step for reusing Algorithm 1 is performed; for the candidate proportions that pass the check, their optimal coding rate is calculated. And estimate the corresponding task losses. (4) Among all feasible ratios, select the ROI ratio that minimizes the task loss. (5) Set the model of collaborator j as follows: And adopt the corresponding optimal coding rate If no proportion meets the conditions, the data transmitted by that collaborator is discarded.

Images