An efficient domain adaptation method and device for cooperative perception parameters of internet of vehicles and a medium

Abstract

The application belongs to the technical field of vehicle networking cooperative perception, and relates to a vehicle networking cooperative perception parameter efficient domain self-adaptation method, device and medium; feature extraction is performed on environment data collected by each vehicle-mounted node in a target domain to obtain a space-time feature vector of each vehicle-mounted node in the target domain; based on Euclidean distances between the space-time feature vectors of each vehicle-mounted node, an optimal transmission distance between each two vehicle-mounted nodes is obtained; taking minimization of a maximum value of the optimal transmission distance between each vehicle-mounted node in the target domain and a selected target vehicle-mounted node set as an objective, a vehicle-mounted node selection function is constructed and solved to obtain the target vehicle-mounted node set; and environment data collected by each vehicle-mounted node in the target vehicle-mounted node set is used to perform parameter efficient fine-tuning on a vehicle networking cooperative perception model carried on each vehicle-mounted node.

Description

Technical Field

[0001] This invention relates to the field of vehicle-to-everything (V2X) cooperative sensing technology, and in particular to a method, apparatus, and computer-readable storage medium for efficient domain adaptation of V2X cooperative sensing parameters. Background Technology

[0002] Traditional single-vehicle perception systems are limited by issues such as sensor field of view obstruction and limited perception range, making it difficult to obtain comprehensive environmental information and severely restricting the safety and reliability of autonomous driving systems. With the rapid development of intelligent transportation systems and autonomous driving technologies, Vehicle-to-Everything (V2X) collaborative perception technology has become a key technological direction for solving the limitations of single-vehicle intelligent perception. V2X collaborative perception technology breaks through the physical limitations of single-vehicle perception by sharing and integrating information between vehicles (V2V) and between vehicles and infrastructure (V2I), enabling vehicles to achieve a wider range and more comprehensive environmental perception during operation.

[0003] In recent years, deep learning-based vehicle cooperative perception models have made significant progress. Each onboard node is equipped with a cooperative perception model, which collects environmental data around the node through onboard sensors and uses the cooperative perception model to extract and filter features, obtaining feature information strongly related to the cooperative task. After the onboard nodes exchange feature information through vehicle-to-everything (V2X) communication protocols, each node fuses the received feature information to achieve visual feature complementarity among multiple nodes, ultimately generating a global, blind-spot-free environmental feature map for decision-making tasks such as target detection, trajectory prediction, and vehicle control. The performance of existing cooperative perception models is highly dependent on the data distribution during the model training phase. When the deployment environment changes (e.g., adjustments to the sensor configuration of onboard nodes, changes in data style due to weather fluctuations, differences in traffic scenarios, etc.), the model struggles to adapt to the new data distribution, resulting in a significant performance degradation. Traditional solutions focus on retraining the model using the new environmental data distribution for full parameter fine-tuning, but this requires a large amount of labeled data and computational resources, as well as significant time costs, failing to achieve fast and lightweight model updates.

[0004] Parameter-Efficient Fine-Turning (PEFT) technology offers a novel solution for rapidly updating models to adapt to new domains. This technique freezes most of the parameters of a pre-trained model, inserting only a small number of trainable lightweight modules or updating only a few parameters to achieve model adaptation to new scenarios. It has already achieved success in natural language processing and single-agent vision tasks. However, directly applying PEFT to multi-agent (multi-vehicle node) collaborative perception scenarios presents the following problems: 1. Due to the high redundancy of spatiotemporal data from multiple agents, the difficulty of effectively selecting features and removing redundant information during the small parameter update process increases, leading to unstable model adaptation in new scenarios. Specifically, the vehicle collaborative perception model updated with the new scenario cannot effectively extract, filter, and fuse feature information from vehicle nodes, thus making it difficult to provide accurate environmental data for subsequent vehicle control decisions; 2. The core of vehicle-to-everything (V2X) collaborative perception is to capture vehicle position and attitude through the fusion of perception data between vehicles. This feature information highly depends on the fine-grained features extracted by shallow networks. When applying efficient parameter fine-tuning techniques to multi-agent (multi-vehicle node) collaborative perception scenarios, the parameters of the backbone network are frozen, preventing the effective transfer of fine-grained spatial semantic information extracted by the model in shallow layers to deep networks. As the network depth increases, these important semantic features gradually decay or are lost, leading to a decline in the model's deep feature representation capability and ultimately affecting the model's target detection accuracy. Furthermore, the lack of effective information interaction and knowledge transfer mechanisms between different stages after inserting adaptation modules into the model also prevents the domain adaptation knowledge learned in early stages from being utilized in subsequent stages, thus affecting the domain adaptation performance of the vehicle-to-everything (V2X) collaborative perception model.

[0005] In summary, when applying the existing parameter fine-tuning technology to the vehicle-to-everything (V2X) cooperative perception model for domain adaptation, there is spatiotemporal data redundancy among multiple vehicle nodes. This makes it difficult to select effective features and remove redundant information during the updating of a small number of parameters. The updated V2X cooperative perception model cannot effectively extract, filter, and fuse feature information from vehicle nodes. At the same time, the feature information extracted at the shallow layer of the model cannot be transmitted to the deep network, and the adaptation modules inserted at different stages of the model cannot interact with each other. This results in poor domain adaptation accuracy of the model, making it difficult to provide accurate environmental data for subsequent vehicle control decisions. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the problem that when the parameter fine-tuning technology in the prior art is applied to the vehicle network cooperative perception model for domain adaptation, there is spatiotemporal data redundancy of multiple vehicle nodes, which makes it difficult to select effective features and remove redundant information during the update of a small number of parameters. The updated vehicle cooperative perception model cannot effectively extract, filter and fuse the feature information of vehicle nodes. At the same time, the feature information extracted by the shallow layer of the model cannot be transmitted to the deep network, and the adaptation modules inserted at different stages of the model cannot interact with each other, resulting in poor model domain adaptation accuracy and difficulty in providing accurate environmental data for subsequent vehicle control decisions.

[0007] To address the aforementioned technical problems, this invention provides an efficient domain adaptive method for cooperative perception parameters in vehicle-to-everything (V2X) networks, comprising:

[0008] Feature extraction is performed on the environmental data collected by each vehicle-mounted node in the target domain to obtain the spatiotemporal feature vector of each vehicle-mounted node in the target domain;

[0009] Based on the Euclidean distance between the spatiotemporal feature vectors of each vehicle node, the optimal transmission distance between any two vehicle nodes is obtained.

[0010] With the objective of minimizing the maximum optimal transmission distance between each vehicle node in the target domain and the selected target vehicle node set, a vehicle node selection function is constructed and solved to obtain the target vehicle node set. The solution to the vehicle node selection function includes: each time a vehicle node is selected, the vehicle node in the target domain with the largest optimal transmission distance to the current target vehicle node set is added to the target vehicle node set, until the number of vehicle nodes in the target vehicle node set reaches a preset value.

[0011] By utilizing the environmental data collected by each vehicle node in the target vehicle node cluster, the parameters of the vehicle-to-everything (V2X) collaborative perception model mounted on each vehicle node are efficiently fine-tuned.

[0012] Preferably, the environmental data collected by each vehicle node in the target vehicle node cluster is used to efficiently fine-tune the parameters of the vehicle-to-everything (V2X) cooperative perception model mounted on each vehicle node, including:

[0013] Connect a dual-path adapter, a collaborative intelligent prompter, and a knowledge compressor in series to the output of the voxel encoder of the vehicle-to-everything (V2X) collaborative perception model.

[0014] A series of dual-path adapters, collaborative intelligent prompters, knowledge compressors, and knowledge injectors are connected to the output of the intermediate feature layer of the backbone network of the vehicle-to-everything (V2X) cooperative perception model.

[0015] Connect a series of dual-path adapters, a collaborative intelligent prompter, a knowledge compressor, and a knowledge injector at the output end of the backbone network of the vehicle-to-everything (V2X) collaborative perception model.

[0016] The parameters of the backbone network are frozen, and the environmental data collected by each vehicle node in the target vehicle node set are input into the vehicle network cooperative perception model. The dual-path adapter, cooperative intelligent prompter, knowledge compressor, knowledge injector and detection head are trained until the target detection loss value is minimized, and the vehicle network cooperative perception model with efficient parameter fine-tuning is obtained.

[0017] Preferably, three first dual-path adapters connected in series are connected to the output of the voxel encoder of the vehicle-to-everything (V2X) collaborative perception model, and multiple first dual-path adapters are used for early feature adaptation in the domain.

[0018] A first collaborative intelligent prompter is connected to the output of the last first dual-path adapter to average and project the features output by the last first dual-path adapter in the vehicle network collaborative perception model of different vehicle nodes, and generate an early collaborative intelligent agent prompt vector.

[0019] A first knowledge compressor is connected to the output of the first collaborative intelligent prompter to compress and reduce the dimensionality of the features output by the last first dual-path adapter.

[0020] A second dual-path adapter is connected to the output of the intermediate feature layer of the backbone network, and the second dual-path adapter is used for mid-domain feature adaptation.

[0021] A second cooperative intelligent prompter is connected to the output end of the second dual-path adapter to average and project the features output by the second dual-path adapter in the vehicle network cooperative perception model of different vehicle nodes, and generate a mid-term cooperative intelligent agent prompt vector.

[0022] A second knowledge compressor is connected to the output of the second collaborative agent prompter to compress and reduce the dimensionality of the features output by the second dual-path adapter.

[0023] The first knowledge injector is connected to the output of the second knowledge compressor to convert the compressed and dimensionality-reduced features into an attention map.

[0024] A third dual-path adapter is connected to the output of the backbone network, and the third dual-path adapter is used for domain post-feature adaptation.

[0025] A third collaborative intelligent prompter is connected to the output end of the third dual-path adapter to average and project the features output by the third dual-path adapter in the vehicle network collaborative perception model of different vehicle nodes, and generate the subsequent collaborative intelligent agent prompt vector.

[0026] A third knowledge compressor is connected to the output of the third collaborative agent prompter to compress and reduce the dimensionality of the features output by the third dual-path adapter;

[0027] The output of the third knowledge compressor is connected to the second knowledge injector, which converts the compressed and dimensionality-reduced features into an attention map and fuses it with the early, mid-term, and late-term collaborative agent cue vectors. This fusion is then input into the detection head for target detection.

[0028] Preferably, the compression ratio of the first knowledge compressor is less than that of the second knowledge compressor, and the compression ratio of the second knowledge compressor is less than that of the third knowledge compressor.

[0029] Preferably, each time an on-board node is selected, the on-board node with the largest optimal transmission distance from the current target on-board node set is added to the target on-board node set, until the number of on-board nodes in the target on-board node set reaches a preset value, including:

[0030] Step 1: Based on the Euclidean distance between the spatiotemporal feature vector of each vehicle node and the mean of the spatiotemporal feature vectors of all vehicle nodes, obtain the optimal transmission distance of each vehicle node from the center of the vehicle node and the average optimal transmission distance of all vehicle nodes from the center of the vehicle node.

[0031] Step 2: Add the vehicle node with the smallest absolute value of the difference between the optimal transmission distance from the center of the vehicle node and the average optimal transmission distance to the target vehicle node set.

[0032] Step 3: Among the remaining vehicle nodes, select the vehicle node with the largest optimal transmission distance to the target vehicle node set and add it to the target vehicle node set. Then return to Step 3 and continue until the number of vehicle nodes in the target vehicle node set reaches the preset value.

[0033] Preferably, feature extraction is performed on the environmental data collected by each vehicle-mounted node in the target domain to obtain the spatiotemporal feature vector of each vehicle-mounted node in the target domain, including:

[0034] The time data in the environmental data collected by each vehicle node is normalized to obtain time feature values;

[0035] By using a spatial projection function to map the spatial location coordinates in the environmental data collected by each vehicle node, two spatial feature values ​​are obtained.

[0036] The node index of each vehicle node in the target domain is normalized to obtain the sequence feature value;

[0037] Based on the temporal feature values, two spatial feature values, and sequence feature values ​​of each vehicle node, the spatiotemporal feature vector of each vehicle node is obtained.

[0038] Preferably, the formula for calculating the optimal transmission distance between every two vehicle-mounted nodes is:

[0039] ,

[0040] in, This represents the optimal transmission distance between the i-th vehicle node and the j-th vehicle node; This represents the spatiotemporal feature vector of the i-th vehicle node; This represents the spatiotemporal feature vector of the j-th vehicle node; The dimension of the spatiotemporal feature vector; Represents time characteristic values; Represents the eigenvalues ​​of the first space; Represents the eigenvalues ​​of the second space; Represents the sequence feature value; This represents the weight of the k-th eigenvalue in the spatiotemporal eigenvector; express The k-th eigenvalue in the equation; express The k-th eigenvalue in the equation;

[0041] Onboard node selection function Represented as:

[0042] ,

[0043] in, Represents the target set of vehicle-mounted nodes; This represents the set of vehicle-mounted nodes in the target domain; This indicates that the number of vehicle nodes in the target vehicle node set is m; Represent the feature space; This represents the selected set of spatiotemporal feature vectors.

[0044] Preferably, the distribution of vehicle-mounted nodes and the environmental data they collect in the target domain. Represented as:

[0045] ,

[0046] in, Indicates the number of vehicle-mounted nodes in the target domain; This represents the spatiotemporal feature vector of the i-th vehicle node; Indicates the Dirac measure;

[0047] Distribution of vehicle-mounted nodes and their collected environmental data in the target vehicle-mounted node set Represented as:

[0048] ,

[0049] in, Indicates the number of vehicle nodes in the target vehicle node set; Represents the target set of vehicle-mounted nodes; This represents the spatiotemporal feature vector of the j-th vehicle node;

[0050] ,

[0051] in, express and The optimal transmission distance, i.e. The mean of the spatiotemporal feature vectors of all vehicle nodes and The Euclidean distance between the mean spatiotemporal feature vectors of all vehicle nodes; express All vehicle-mounted nodes can cover The minimum radius of all vehicle-mounted nodes; express All vehicle-mounted nodes can cover The theoretical minimum radius of all vehicle-mounted nodes.

[0052] The present invention also provides a high-efficiency domain adaptive device for vehicle-to-everything (V2X) cooperative sensing parameters, comprising:

[0053] The spatiotemporal feature construction module is used to extract features from the environmental data collected by each vehicle node in the target domain, and obtain the spatiotemporal feature vector of each vehicle node in the target domain.

[0054] The transmission distance acquisition module is used to obtain the optimal transmission distance between any two vehicle nodes based on the Euclidean distance between the spatiotemporal feature vectors of each vehicle node.

[0055] The vehicle-mounted node selection module is used to construct and solve a vehicle-mounted node selection function with the objective of minimizing the maximum value of the optimal transmission distance between each vehicle-mounted node in the target domain and the selected target vehicle-mounted node set, thereby obtaining the target vehicle-mounted node set. The solution to the vehicle-mounted node selection function includes: each time a vehicle-mounted node is selected, the vehicle-mounted node in the target domain with the largest optimal transmission distance to the current target vehicle-mounted node set is added to the target vehicle-mounted node set, until the number of vehicle-mounted nodes in the target vehicle-mounted node set reaches a preset value.

[0056] The model domain adaptive module is used to efficiently fine-tune the parameters of the vehicle-to-everything (V2X) collaborative perception model mounted on each V2X node by utilizing the environmental data collected by each V2X node in the target V2X node set.

[0057] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described efficient domain adaptive method for vehicle-to-everything (V2X) cooperative perception parameters.

[0058] The efficient domain adaptive method for cooperative perception parameters in vehicle-to-everything (V2X) networks provided in this application has the following advantages:

[0059] 1. Perform unified spatiotemporal feature extraction on the environmental data of all vehicle-mounted nodes in the target domain. Define the Euclidean distance between the spatiotemporal feature vectors of the vehicle-mounted nodes as the optimal transmission distance. The smaller the distance, the more similar the spatiotemporal features collected by the two vehicle-mounted nodes, and the higher the data redundancy. The larger the distance, the stronger the complementarity of the spatiotemporal features of the two vehicle-mounted nodes, and the higher the data value. Transform the problem of eliminating redundant information into a computable distance optimization problem. When selecting the target vehicle-mounted node set that can represent the distribution of environmental data of all vehicle-mounted nodes in the target domain, first add the vehicle-mounted node that is closest to the global features to the target vehicle-mounted node set to cover the core environmental information of the target domain. Then, in each selection, select the unselected vehicle-mounted nodes in the target domain that are closest to the current target vehicle-mounted node. The vehicle node with the largest optimal transmission distance between point sets is added to the target vehicle node set, thereby selecting the vehicle node with the strongest feature complementarity to ensure the comprehensive feature coverage of the target vehicle node set. The final target vehicle node set is used to efficiently fine-tune the parameters of the vehicle network cooperative perception model mounted on each vehicle node. Since the redundancy of the input data has been greatly reduced, a small number of parameter updates of the model can accurately focus on the core spatiotemporal features of the target domain. This solves the problem that when PEFT technology is directly used to update the vehicle cooperative perception model, the spatiotemporal data redundancy causes the updated vehicle cooperative perception model to be unable to effectively extract, filter and fuse the feature information of vehicle nodes, thus making it difficult to provide accurate environmental data for subsequent vehicle control decisions.

[0060] 2. This application divides the vehicle-to-everything (V2X) cooperative perception model into three stages: the early stage refers to the period after the voxel encoder and before the backbone network; the mid-stage refers to the intermediate feature layer of the backbone network; and the late stage refers to the period after the backbone network and before the detector head. In addition to introducing dual-path adapters for domain adaptation at each stage, cooperative intelligent agent prompts are also introduced at each stage. At each stage, the features of each vehicle node are aggregated to capture common patterns among the vehicle nodes. The aggregated features are mapped to prompt vectors. Finally, the prompt vectors generated at each stage are concatenated with the feature information of the vehicle nodes input to the detector head, and the prompt vectors are used as... To provide additional contextual information, the model is guided to focus on the collaborative patterns between vehicle nodes. In addition, to ensure efficient transmission of feature information to later stages, this application introduces a knowledge compressor at each stage to compress and reduce the dimensionality of feature information at each stage. A progressive knowledge transfer mechanism is used to establish an information transmission path across network stages, enabling deep networks to acquire fine-grained spatial semantic features extracted from shallow layers. This avoids the gradual decay of feature information during forward propagation, requiring only about 1% of the parameters to be updated to achieve model domain adaptation. Compared with the full parameter fine-tuning method, this reduces the amount of parameter updates by about 99% and lowers the computational resource requirements. Attached Figure Description

[0061] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0062] Figure 1 The flowchart of the efficient domain adaptive method for vehicle-to-everything (V2X) cooperative perception parameters provided in this application is shown below.

[0063] Figure 2 This is a schematic diagram of the vehicle-to-everything (V2X) cooperative perception model structure provided in this application;

[0064] Figure 3 This is a schematic diagram of the structure of the high-efficiency domain adaptive device for vehicle network cooperative perception parameters provided in this application. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0066] Please see Figure 1 , Figure 1 The flowchart shown is a method for efficient domain adaptive sensing of vehicle network cooperative perception parameters provided in this application. The method specifically includes S10~S40:

[0067] S10: Extract features from the environmental data collected by each vehicle-mounted node in the target domain to obtain the spatiotemporal feature vectors of each vehicle-mounted node in the target domain.

[0068] It should be noted that the vehicle nodes and their environmental data in the source domain refer to the vehicle nodes and data in the scenario used to train the vehicle-to-everything (V2X) cooperative perception model, that is, the vehicle nodes and data in the scenario originally equipped with the V2X cooperative perception model; the vehicle nodes and environmental data in the target domain refer to the vehicle nodes and environmental data in the new scenario after the environment changes.

[0069] For example, the target domain dataset , This represents the sensor observation data deployed on the i-th vehicle-mounted node in the target domain; The timestamp of environmental data collected by the sensors deployed on the i-th vehicle node; This represents the spatial coordinates collected by the sensors deployed on the i-th vehicle node.

[0070] S20: Based on the Euclidean distance between the spatiotemporal feature vectors of each vehicle node, the optimal transmission distance between any two vehicle nodes is obtained.

[0071] S30: With the objective of minimizing the maximum optimal transmission distance between each vehicle node in the target domain and the selected target vehicle node set, construct and solve the vehicle node selection function to obtain the target vehicle node set; wherein, solving the vehicle node selection function includes: each time a vehicle node is selected, adding the vehicle node in the target domain with the largest optimal transmission distance to the current target vehicle node set to the target vehicle node set, until the number of vehicle nodes in the target vehicle node set reaches a preset value.

[0072] S40: Utilize the environmental data collected by each vehicle node in the target vehicle node cluster to efficiently fine-tune the parameters of the vehicle-to-everything (V2X) collaborative perception model mounted on each vehicle node.

[0073] Specifically, feature extraction is performed on the environmental data collected by each vehicle-mounted node in the target domain to obtain the spatiotemporal feature vectors of each vehicle-mounted node in the target domain, including S100~S103:

[0074] S100: Normalize the time data in the environmental data collected by each vehicle-mounted node to obtain time feature values. For example, the time feature value of the i-th vehicle-mounted node... The calculation formula is:

[0075] ,

[0076] in, This represents the timestamp of the environmental data collected by the sensors deployed on the i-th vehicle node; This represents the time normalization coefficient.

[0077] S101: The spatial location coordinates in the environmental data collected by each vehicle node are mapped using a spatial projection function to obtain two spatial feature values. For example, the two spatial feature values ​​of the i-th vehicle node... and The calculation formula is:

[0078] ,

[0079] in, Represents the spatial projection function. .

[0080] S102: Normalize the node index of each vehicle node in the target domain to obtain the sequence feature value. For example, the sequence feature value of the i-th vehicle node... The calculation formula is:

[0081] ,

[0082] in, This indicates the number of vehicle-mounted nodes in the target domain.

[0083] S103: Based on the temporal feature values, two spatial feature values, and sequence feature values ​​of each vehicle node, the spatiotemporal feature vector of each vehicle node is obtained. For example, the spatiotemporal feature vector of the i-th vehicle node... Represented as:

[0084] ,

[0085] in, This indicates transpose.

[0086] Furthermore, the formula for calculating the optimal transmission distance between every two vehicle-mounted nodes is as follows:

[0087] ,

[0088] in, This represents the optimal transmission distance between the i-th vehicle node and the j-th vehicle node; This represents the spatiotemporal feature vector of the i-th vehicle node; This represents the spatiotemporal feature vector of the j-th vehicle node; The dimension of the spatiotemporal feature vector; Represents time characteristic values; Represents the eigenvalues ​​of the first space; Represents the eigenvalues ​​of the second space; Represents the sequence feature value; This represents the weight of the k-th eigenvalue in the spatiotemporal eigenvector; express The k-th eigenvalue in the equation; express The k-th eigenvalue in the matrix.

[0089] In one embodiment of this application, the weight vector This weighting design prioritizes the diversity of temporal distribution, followed by spatial coverage, and finally sequence continuity. This ensures that the selected vehicle nodes are widely distributed in the temporal dimension while guaranteeing comprehensive spatial coverage.

[0090] Furthermore, this application models the vehicle node selection problem as finding the minimum dominating set that minimizes the coverage radius. By ensuring that the maximum distance from the unselected vehicle nodes in the target domain to the current target vehicle node set is minimized after each selection of vehicle nodes, the vehicle nodes in the target vehicle node set can cover the vehicle nodes in the entire target domain with the minimum radius, thereby minimizing the maximum transmission cost.

[0091] Specifically, the on-board node selection function Represented as:

[0092] ,

[0093] in, Represents the target set of vehicle-mounted nodes; This represents the set of vehicle-mounted nodes in the target domain; This indicates that the number of vehicle nodes in the target vehicle node set is m; Represent the feature space; This represents the selected set of spatiotemporal feature vectors.

[0094] Specifically, solving the on-board node selection function includes steps 1 through 3:

[0095] Step 1: Based on the Euclidean distance between the spatiotemporal feature vector of each vehicle node and the mean of the spatiotemporal feature vectors of all vehicle nodes, obtain the optimal transmission distance of each vehicle node from the center of the vehicle node and the average optimal transmission distance of all vehicle nodes from the center of the vehicle node.

[0096] Step 2: Add the vehicle node with the smallest absolute value of the difference between the optimal transmission distance from the center of the vehicle node and the average optimal transmission distance to the target vehicle node set.

[0097] For example, the vehicle node with the smallest absolute value of the difference between the optimal transmission distance from the center of the vehicle node and the average optimal transmission distance. , This represents the optimal transmission distance between the i-th vehicle node and the center of the vehicle node; This represents the average optimal transmission distance from all vehicle-mounted nodes to the center of the vehicle-mounted node.

[0098] Step 3: Among the remaining vehicle nodes, select the vehicle node with the largest optimal transmission distance to the target vehicle node set and add it to the target vehicle node set. Then return to Step 3 and continue until the number of vehicle nodes in the target vehicle node set reaches the preset value.

[0099] Specifically, the optimal transmission distance between the vehicle-mounted node and the target set of vehicle-mounted nodes is the Euclidean distance between the spatiotemporal feature vector of the vehicle-mounted node and the mean of the spatiotemporal feature vectors of all vehicle-mounted nodes in the target set of vehicle-mounted nodes.

[0100] Specifically, the distribution of vehicle-mounted nodes and the environmental data they collect in the target domain. Represented as:

[0101] ,

[0102] in, Indicates the number of vehicle-mounted nodes in the target domain; This represents the spatiotemporal feature vector of the i-th vehicle node; This represents the Dirac measure.

[0103] Distribution of vehicle-mounted nodes and their collected environmental data in the target vehicle-mounted node set Represented as:

[0104] ,

[0105] in, Indicates the number of vehicle nodes in the target vehicle node set; Represents the target set of vehicle-mounted nodes; This represents the spatiotemporal feature vector of the j-th vehicle node;

[0106] ,

[0107] in, express and The optimal transmission distance, i.e. The mean of the spatiotemporal feature vectors of all vehicle nodes and The Euclidean distance between the mean spatiotemporal feature vectors of all vehicle nodes; express All vehicle-mounted nodes can cover The minimum radius of all vehicle-mounted nodes; express All vehicle-mounted nodes can cover The theoretical minimum radius of all vehicle-mounted nodes.

[0108] Specifically, all vehicle-mounted nodes can cover The minimum radius of all vehicle-mounted nodes in the target domain refers to the minimum radius of a circle centered on each vehicle-mounted node in the target vehicle-mounted node set, such that all vehicle-mounted nodes in the target domain are contained within at least one circle. By limiting the optimal transmission distance between the target vehicle-mounted node set and the target domain vehicle-mounted node set, a uniform distribution of the target vehicle-mounted node set can be ensured, thus representing the distribution of vehicle-mounted nodes in the target domain with a bounded transmission cost.

[0109] During the efficient fine-tuning of parameters, the parameters of the backbone network are frozen, which prevents the fine-grained spatial semantic information extracted by the model in the shallow layers from being effectively transferred to the deep layers. As the network layers deepen, these important semantic features gradually decay or are lost, resulting in a decrease in the model's deep feature representation ability and ultimately affecting the model's target detection accuracy. At the same time, after inserting adaptation modules at different stages of the model, there is a lack of effective information interaction and knowledge transfer mechanisms between the stages, which also prevents the domain adaptation knowledge learned in the early stages from being used in subsequent stages, thus affecting the model's domain adaptation performance.

[0110] To address the aforementioned issues, this application aims to establish a knowledge transfer pathway across network stages when efficiently fine-tuning the parameters of the vehicle-to-everything (V2X) cooperative perception model. This pathway transfers fine-grained semantic features from earlier stages to deeper networks, such as... Figure 2 As shown, the vehicle-to-everything (V2X) cooperative perception model is first divided into three stages: the early stage. This refers to the intermediate stage, after the voxel encoder and before the backbone network. This refers to the intermediate feature layer of the backbone network, the later stage. This refers to the process after the backbone network and before the detection head.

[0111] Specifically, by utilizing the environmental data collected by each vehicle-mounted node in the target vehicle-mounted node cluster, the parameters of the vehicle-to-everything (V2X) cooperative perception model mounted on each vehicle-mounted node are efficiently fine-tuned, including S400~S403:

[0112] S400: Connects a dual-path adapter, a collaborative intelligent prompter, and a knowledge compressor in series to the output of the voxel encoder of the vehicle-to-everything (V2X) collaborative perception model.

[0113] S401: Connects a series of dual-path adapters, collaborative intelligent prompters, knowledge compressors, and knowledge injectors at the output of the intermediate feature layer of the backbone network of the vehicle-to-everything (V2X) cooperative perception model.

[0114] S402: Connect a series of dual-path adapters, a collaborative intelligent prompter, a knowledge compressor, and a knowledge injector to the output of the backbone network of the vehicle-to-everything (V2X) collaborative perception model.

[0115] S403: Freeze the parameters of the backbone network, input the environmental data collected by each vehicle node in the target vehicle node set into the vehicle network cooperative perception model, train each dual-path adapter, cooperative intelligent prompter, knowledge compressor, knowledge injector and detection head until the target detection loss value is minimized, and obtain the vehicle network cooperative perception model with efficient parameter fine-tuning.

[0116] Specifically, the object detection loss function adopts the standard 3D object detection loss, which is obtained by a weighted combination of classification loss, localization loss, and orientation loss. The AdamW optimizer is used during model training, and the learning rate is set to... The cosine annealing learning rate scheduling strategy is adopted. At the same time, since the number of trainable parameters is significantly reduced to about 1% of the total parameters, the training process of this application has lower requirements for hardware resources. Training can be completed on a single NVIDIA RTX 40800 GPU. Compared with the full parameter fine-tuning valve, the GPU memory ratio is greatly reduced, the backpropagation computation is significantly reduced, the time per iteration is shorter, the model converges faster, and fewer iterations are required.

[0117] Specifically, this application also provides another implementation to further explain the above steps S400~S403, which specifically includes S500~S510:

[0118] S500: Connect three first dual-path adapters in series to the output of the voxel encoder of the vehicle-to-everything (V2X) cooperative perception model, and use multiple first dual-path adapters to perform early feature adaptation in the domain.

[0119] Specifically, using multiple dual-path adapters in the early stages can fully capture the domain characteristics of the original features.

[0120] S501: Connect the first cooperative intelligent prompter to the output of the last first dual-path adapter to average and project the features output by the last first dual-path adapter in the vehicle network cooperative perception model of different vehicle nodes to generate an early cooperative intelligent agent prompt vector.

[0121] S502: Connect the first knowledge compressor to the output of the first collaborative intelligent prompter to compress and reduce the dimensionality of the features output by the last first dual-path adapter.

[0122] S503: Connect a second dual-path adapter to the output of the intermediate feature layer of the backbone network, and use the second dual-path adapter to perform mid-domain feature adaptation.

[0123] S504: Connect the second collaborative intelligent prompter to the output end of the second dual-path adapter to average and project the features output by the second dual-path adapter in the vehicle network collaborative perception model of different vehicle nodes to generate a mid-term collaborative intelligent agent prompt vector.

[0124] S505: Connect the second knowledge compressor to the output of the second collaborative agent prompter to compress and reduce the dimensionality of the features output by the second dual-path adapter.

[0125] S506: Connect the first knowledge injector to the output of the second knowledge compressor to convert the compressed and dimensionality-reduced features into an attention map.

[0126] S507: Connect a third dual-path adapter to the output of the backbone network and use the third dual-path adapter for late-stage domain feature adaptation.

[0127] S508: Connect the third collaborative intelligent prompter to the output end of the third dual-path adapter. It is used to average and project the features output by the third dual-path adapter in the vehicle network collaborative perception model of different vehicle nodes to generate the subsequent collaborative intelligent agent prompt vector.

[0128] S509: Connect the third knowledge compressor to the output of the third collaborative agent prompter to compress and reduce the dimensionality of the features output by the third dual-path adapter.

[0129] S510: Connect the second knowledge injector to the output of the third knowledge compressor, convert the compressed and dimensionality-reduced features into an attention map and fuse it with the early collaborative agent cue vector, the mid-term collaborative agent cue vector and the late-term collaborative agent cue vector, and input it into the detection head for target detection.

[0130] Furthermore, the dual-path adapter performs domain feature adaptation, including steps 2-1 to 2-4:

[0131] Step 2-1: Use spatial paths to extract features from the input features to obtain local spatial features.

[0132] Specifically, the spatial path includes multiple grouped convolutions, each grouped convolution performing a 3*3 convolution operation on the input features to capture local features of the input features, and finally activating them with BatchNorm and ReLU.

[0133] Step 2-2: Perform pointwise convolution on the input features using the channel path to obtain global semantic features.

[0134] Specifically, the channel path uses 1*1 convolution to perform dimensionality reduction and dimensionality increase operations on the channel dimension, with a dimensionality reduction ratio of 1 / 4, thereby modeling the semantic association across channels.

[0135] Steps 2-3: Perform weighted fusion of local spatial features and global semantic features to obtain fused features.

[0136] Specifically, the outputs of the two paths are adaptively combined using learnable fusion weights to obtain fusion features. Represented as:

[0137] ,

[0138] in, Represents local spatial features; express The weights; Represents global semantic features; express The weight. Specifically, and Normalization via Softmax , and These are learnable parameters.

[0139] Steps 2-4: Add the fused features to the input features through residual connections to obtain the output features of the dual-path adapter.

[0140] Specifically, the output characteristics of the dual-path adapter Represented as:

[0141] ,

[0142] in, This represents the input characteristics of a two-path adapter; This represents the learnable scaling factor, initialized to 0.1.

[0143] Furthermore, this application introduces a collaborative agent prompting strategy at each stage. At each stage, the features of each vehicle node are aggregated to capture common patterns among vehicle nodes. The aggregated features are mapped into prompt vectors. Finally, the prompt vectors generated at each stage are concatenated with the feature information of the vehicle nodes input to the detection head. The prompt vectors are used as additional contextual information to guide the model to focus on the collaborative patterns among vehicle nodes.

[0144] Specifically, the collaborative agent cue vectors generated at each stage Represented as:

[0145] ,

[0146] in, This represents the vehicle node group used for feature aggregation; The characteristics of the i-th on-board node output by the dual-path adapter at each stage are represented. This indicates a feature aggregation operation; It is a lightweight projection layer used to map aggregated features into cue vectors.

[0147] The cue vectors generated at each stage are concatenated with the feature information of the vehicle-mounted nodes in the input detection head, and represented as follows:

[0148] ,

[0149] in, The feature of the i-th vehicle node input to the detection head is represented by the concatenation of feature information and prompt vector. This indicates a channel-level concatenation operation; This represents the feature information of the i-th vehicle-mounted node in the input detection head.

[0150] Furthermore, to ensure efficient transmission of feature information to later stages, this application also introduces a knowledge compressor to compress and reduce the dimensionality of feature information at each stage. Each knowledge compressor uses an adaptive average pooling layer to compress the input feature map to 1 / r of the original size, where r represents the compression ratio. Then, a convolutional layer is used to compress the channel dimension. Finally, after passing through a BN batch normalization layer and a ReLU activation function, the compressed and dimensionality-reduced feature map is output.

[0151] Specifically, the compression ratio of the first knowledge compressor is less than that of the second knowledge compressor, and the compression ratio of the second knowledge compressor is less than that of the third knowledge compressor, thereby ensuring that more refined feature representations are gradually transmitted as the model depth increases.

[0152] Furthermore, for the mid- and late-stages, this application also introduces a knowledge injector to convert the compressed and dimensionality-reduced features into attention maps. :

[0153] ,

[0154] in, , All are 1*1 convolutional layers; This represents the Sigmoid activation function; This represents the features after compression and dimensionality reduction.

[0155] Furthermore, attention maps It can also be used to selectively enhance features of the current stage:

[0156] ，

[0157] in, Indicates the enhanced features; This represents the characteristics of the vehicle-mounted node at the current stage (i.e., the characteristics output by the intermediate feature layer of the backbone network in the mid-term stage or the characteristics output by the backbone network in the late stage). This represents element-wise multiplication; This represents the learnable enhancement coefficient, initialized to 0.1~0.5 to avoid overmodulation.

[0158] Alternatively, to stabilize training and avoid the computational overhead of gradient backpropagation, a decoupled feature memory mechanism can be used at the output of each stage. , This represents the feature of the i-th vehicle node output by the dual-path adapter in the early stage. The gradient flow is cut off using the detach operation, making the optimization in the early stage independent of the mid- and late-stages. The mid- and late-stages can directly access the gradient flow from the detached path. Features are read from the graph, avoiding the memory overhead of maintaining a long-term computation graph.

[0159] To verify the effectiveness of the above method, this application also provides a process for rapidly adapting a vehicle-to-everything (V2X) cooperative perception model pre-trained on the simulation dataset OPV2V to the real-world scenario dataset DAIR-V2X:

[0160] First, the CoAlign collaborative perception model was pre-trained using the OPV2V dataset. OPV2V contains 11,464 frames of multi-vehicle node volumetric collaborative scenes, with each frame containing 2 to 7 vehicle nodes. The model uses PointPillar as the backbone network, converting LiDAR point cloud voxels into bird's-eye view features with a grid size of 0.4m*0.4m. Pre-training was performed using full-parameter training for 30 epochs, ultimately achieving an AP@50 of 68.5% and an AP@70 of 59.3% on the OPV2V test set.

[0161] The DAIR-V2X dataset contains approximately 9,000 real-world vehicle-to-infrastructure (V2I) scenarios, each scenario comprising one vehicular agent and one road test infrastructure. The dataset is divided into training, validation, and testing sets. Approximately 900 vehicular nodes are selected from the training set using the vehicular node selection method provided in this application, representing 10% of the total, for domain adaptation.

[0162] When performing domain adaptation of the cooperative perception model using selected vehicle nodes, the AdamW optimizer is employed, with a learning rate set to [value missing]. The weight decay coefficient is Cosine annealing learning rate scheduling is used, with a minimum learning rate of [missing information]. The training cycle is 30 epochs, the maximum gradient norm of gradient clipping is 10.0, mixed precision training is accelerated using FP16, and the weight of classification loss in the object detection loss function is 1, the weight of localization loss is 2, and the weight of orientation loss is 0.2.

[0163] The domain-adaptive collaborative sensing model was tested on the DAIR-V2X test set, with a sensing range of: x [-100m, 100m], y [-40m, 40m], AP@IoU=0.5, AP@IoU=0.7, AP@50=65.9%, AP@70=49.9%.

[0164] This application finds that when a collaborative perception model trained on a source domain dataset is directly used on a target domain dataset, its AP@50=42.9% and AP@70=21.7%. However, after training with the domain adaptation method provided in this application, the model's performance on the target domain is improved by 23% and 28.2%, respectively, effectively bridging the gap between the simulation and the real domain. Compared with the performance of the traditional full-parameter fine-tuning method under the same amount of data (AP@50=59.9% and AP@70=39.5%), this application improves by 6% and 10.4%, respectively, proving the effectiveness of the solution provided in this application.

[0165] Based on the efficient domain adaptive method for vehicle-to-everything (V2X) cooperative perception parameters provided in the above embodiments, this application also provides an efficient domain adaptive device for V2X cooperative perception parameters, such as... Figure 3 As shown, the device specifically includes:

[0166] The spatiotemporal feature construction module 10 is used to extract features from the environmental data collected by each vehicle node in the target domain to obtain the spatiotemporal feature vector of each vehicle node in the target domain.

[0167] The transmission distance acquisition module 20 is used to obtain the optimal transmission distance between every two vehicle nodes based on the Euclidean distance between the spatiotemporal feature vectors of each vehicle node.

[0168] The vehicle-mounted node filtering module 30 is used to construct and solve a vehicle-mounted node selection function with the objective of minimizing the maximum value of the optimal transmission distance between each vehicle-mounted node in the target domain and the selected target vehicle-mounted node set, thereby obtaining the target vehicle-mounted node set. The solution of the vehicle-mounted node selection function includes: each time a vehicle-mounted node is selected, the vehicle-mounted node in the target domain with the largest optimal transmission distance to the current target vehicle-mounted node set is added to the target vehicle-mounted node set, until the number of vehicle-mounted nodes in the target vehicle-mounted node set reaches a preset value.

[0169] The model domain adaptive module 40 is used to efficiently fine-tune the parameters of the vehicle network collaborative perception model mounted on each vehicle node by utilizing the environmental data collected by each vehicle node in the target vehicle node set.

[0170] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described efficient domain adaptive method for vehicle network cooperative perception parameters.

[0171] The method provided in this application is the first to apply optimal transport theory to domain adaptation in vehicle-mounted node cooperative perception. By modeling with optimal transport theory, a target set of vehicle-mounted nodes that can fully represent the distribution of the original data is selected from multiple vehicle-mounted nodes in the target domain. After using 10% of the vehicle-mounted nodes to efficiently fine-tune the parameters of the vehicle-to-everything (V2X) cooperative perception model, the performance of the model can exceed that of the model fine-tuned by randomly selecting vehicle-mounted nodes, thus reducing the parameter fine-tuning time. During the efficient parameter fine-tuning, a progressive knowledge transfer mechanism is used to establish an information transmission path across network stages, enabling the deep network to acquire fine-grained spatial semantic features extracted from the shallow layers. This avoids the gradual attenuation of feature information during forward propagation, and only about 1% of the parameters need to be updated to achieve model domain adaptation. Compared with the full parameter fine-tuning method, the amount of parameter updates is reduced by about 99%, thus reducing the computational resource requirements. It is applicable to different cooperative perception architectures, and the V2X cooperative perception model after efficient parameter fine-tuning exhibits good robustness under different localization noise levels.

[0172] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0173] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0174] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0175] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0176] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A highly efficient domain adaptive method for cooperative sensing parameters in vehicle-to-everything (V2X) communication, characterized in that, include: Feature extraction is performed on the environmental data collected by each vehicle-mounted node in the target domain to obtain the spatiotemporal feature vector of each vehicle-mounted node in the target domain; Based on the Euclidean distance between the spatiotemporal feature vectors of each vehicle node, the optimal transmission distance between any two vehicle nodes is obtained. With the objective of minimizing the maximum optimal transmission distance between each vehicle node in the target domain and the selected target vehicle node set, a vehicle node selection function is constructed and solved to obtain the target vehicle node set. The solution to the vehicle node selection function includes: each time a vehicle node is selected, the vehicle node in the target domain with the largest optimal transmission distance to the current target vehicle node set is added to the target vehicle node set, until the number of vehicle nodes in the target vehicle node set reaches a preset value. By utilizing environmental data collected from various vehicle-mounted nodes in the target vehicle-mounted node cluster, the parameters of the vehicle-to-everything (V2X) cooperative perception model mounted on each vehicle-mounted node are efficiently fine-tuned. Specifically, this includes: Connect a dual-path adapter, a collaborative intelligent prompter, and a knowledge compressor in series to the output of the voxel encoder of the vehicle-to-everything (V2X) collaborative perception model. A series of dual-path adapters, collaborative intelligent prompters, knowledge compressors, and knowledge injectors are connected to the output of the intermediate feature layer of the backbone network of the vehicle-to-everything (V2X) cooperative perception model. Connect a series of dual-path adapters, a collaborative intelligent prompter, a knowledge compressor, and a knowledge injector at the output end of the backbone network of the vehicle-to-everything (V2X) collaborative perception model. The parameters of the backbone network are frozen, and the environmental data collected by each vehicle node in the target vehicle node set are input into the vehicle network cooperative perception model. The dual-path adapter, cooperative intelligent prompter, knowledge compressor, knowledge injector and detection head are trained until the target detection loss value is minimized, and the vehicle network cooperative perception model with efficient parameter fine-tuning is obtained.

2. The efficient domain adaptive method for collaborative perception parameters in vehicle networking according to claim 1, characterized in that, Three first dual-path adapters connected in series are connected to the output of the voxel encoder of the vehicle-to-everything (V2X) cooperative perception model, and multiple first dual-path adapters are used to perform early feature adaptation in the domain. A first collaborative intelligent prompter is connected to the output of the last first dual-path adapter to average and project the features output by the last first dual-path adapter in the vehicle network collaborative perception model of different vehicle nodes, and generate an early collaborative intelligent agent prompt vector. A first knowledge compressor is connected to the output of the first collaborative intelligent prompter to compress and reduce the dimensionality of the features output by the last first dual-path adapter. A second dual-path adapter is connected to the output of the intermediate feature layer of the backbone network, and the second dual-path adapter is used for mid-domain feature adaptation. A second cooperative intelligent prompter is connected to the output end of the second dual-path adapter to average and project the features output by the second dual-path adapter in the vehicle network cooperative perception model of different vehicle nodes, and generate a mid-term cooperative intelligent agent prompt vector. A second knowledge compressor is connected to the output of the second collaborative agent prompter to compress and reduce the dimensionality of the features output by the second dual-path adapter. The first knowledge injector is connected to the output of the second knowledge compressor to convert the compressed and dimensionality-reduced features into an attention map. A third dual-path adapter is connected to the output of the backbone network, and the third dual-path adapter is used for domain post-feature adaptation. A third collaborative intelligent prompter is connected to the output end of the third dual-path adapter to average and project the features output by the third dual-path adapter in the vehicle network collaborative perception model of different vehicle nodes, and generate the subsequent collaborative intelligent agent prompt vector. A third knowledge compressor is connected to the output of the third collaborative agent prompter to compress and reduce the dimensionality of the features output by the third dual-path adapter; The output of the third knowledge compressor is connected to the second knowledge injector, which converts the compressed and dimensionality-reduced features into an attention map and fuses it with the early, mid-term, and late-term collaborative agent cue vectors. This fusion is then input into the detection head for target detection.

3. The efficient domain adaptive method for collaborative perception parameters in vehicle networking according to claim 2, characterized in that, The compression ratio of the first knowledge compressor is less than that of the second knowledge compressor, and the compression ratio of the second knowledge compressor is less than that of the third knowledge compressor.

4. The efficient domain adaptive method for collaborative perception parameters in vehicle networking according to claim 1, characterized in that, Each time an on-board node is selected, the on-board node with the largest optimal transmission distance from the current target on-board node set is added to the target on-board node set, until the number of on-board nodes in the target on-board node set reaches a preset value, including: Step 1: Based on the Euclidean distance between the spatiotemporal feature vector of each vehicle node and the mean of the spatiotemporal feature vectors of all vehicle nodes, obtain the optimal transmission distance of each vehicle node from the center of the vehicle node and the average optimal transmission distance of all vehicle nodes from the center of the vehicle node. Step 2: Add the vehicle node with the smallest absolute value of the difference between the optimal transmission distance from the center of the vehicle node and the average optimal transmission distance to the target vehicle node set. Step 3: Among the remaining vehicle nodes, select the vehicle node with the largest optimal transmission distance to the target vehicle node set and add it to the target vehicle node set. Then return to Step 3 and continue until the number of vehicle nodes in the target vehicle node set reaches the preset value.

5. The efficient domain adaptive method for collaborative perception parameters in vehicle networking according to claim 1, characterized in that, Feature extraction is performed on the environmental data collected by each vehicle-mounted node in the target domain to obtain the spatiotemporal feature vectors of each vehicle-mounted node in the target domain, including: The time data in the environmental data collected by each vehicle node is normalized to obtain time feature values; By using a spatial projection function to map the spatial location coordinates in the environmental data collected by each vehicle node, two spatial feature values ​​are obtained. The node index of each vehicle node in the target domain is normalized to obtain the sequence feature value; Based on the temporal feature values, two spatial feature values, and sequence feature values ​​of each vehicle node, the spatiotemporal feature vector of each vehicle node is obtained.

6. The efficient domain adaptive method for cooperative sensing parameters in vehicle networking according to claim 5, characterized in that, The formula for calculating the optimal transmission distance between any two vehicle-mounted nodes is: ， in, This represents the optimal transmission distance between the i-th vehicle node and the j-th vehicle node; This represents the spatiotemporal feature vector of the i-th vehicle node; This represents the spatiotemporal feature vector of the j-th vehicle node; The dimension of the spatiotemporal feature vector; Represents time characteristic values; Represents the eigenvalues ​​of the first space; Represents the eigenvalues ​​of the second space; Represents the sequence feature value; This represents the weight of the k-th eigenvalue in the spatiotemporal eigenvector; express The k-th eigenvalue in the equation; express The k-th eigenvalue in the equation; Onboard node selection function Represented as: ， in, Represents the target set of vehicle-mounted nodes; This represents the set of vehicle-mounted nodes in the target domain; This indicates that the number of vehicle nodes in the target vehicle node set is m; Represent the feature space; This represents the selected set of spatiotemporal feature vectors.

7. The efficient domain adaptive method for collaborative perception parameters in vehicle networking according to claim 1, characterized in that, Distribution of vehicle-mounted nodes and their collected environmental data in the target domain Represented as: ， in, Indicates the number of vehicle-mounted nodes in the target domain; This represents the spatiotemporal feature vector of the i-th vehicle node; Indicates the Dirac measure; Distribution of vehicle-mounted nodes and their collected environmental data in the target vehicle-mounted node set Represented as: ， in, Indicates the number of vehicle nodes in the target vehicle node set; Represents the target set of vehicle-mounted nodes; This represents the spatiotemporal feature vector of the j-th vehicle node; ， in, express and The optimal transmission distance, i.e. The mean of the spatiotemporal feature vectors of all vehicle nodes and The Euclidean distance between the mean spatiotemporal feature vectors of all vehicle nodes; express All vehicle-mounted nodes can cover The minimum radius of all vehicle-mounted nodes; express All vehicle-mounted nodes can cover The theoretical minimum radius of all vehicle-mounted nodes.

8. A high-efficiency domain adaptive device for collaborative sensing parameters in vehicle networking, characterized in that, include: The spatiotemporal feature construction module is used to extract features from the environmental data collected by each vehicle node in the target domain, and obtain the spatiotemporal feature vector of each vehicle node in the target domain. The transmission distance acquisition module is used to obtain the optimal transmission distance between any two vehicle nodes based on the Euclidean distance between the spatiotemporal feature vectors of each vehicle node. The vehicle-mounted node selection module is used to construct and solve a vehicle-mounted node selection function with the objective of minimizing the maximum value of the optimal transmission distance between each vehicle-mounted node in the target domain and the selected target vehicle-mounted node set, thereby obtaining the target vehicle-mounted node set. The solution to the vehicle-mounted node selection function includes: each time a vehicle-mounted node is selected, the vehicle-mounted node in the target domain with the largest optimal transmission distance to the current target vehicle-mounted node set is added to the target vehicle-mounted node set, until the number of vehicle-mounted nodes in the target vehicle-mounted node set reaches a preset value. The model domain adaptive module is used to efficiently fine-tune the parameters of the vehicle-to-everything (V2X) cooperative perception model mounted on each V2X node using environmental data collected from various V2X nodes in the target V2X node set. Specifically, it includes: Connect a dual-path adapter, a collaborative intelligent prompter, and a knowledge compressor in series to the output of the voxel encoder of the vehicle-to-everything (V2X) collaborative perception model. A series of dual-path adapters, collaborative intelligent prompters, knowledge compressors, and knowledge injectors are connected to the output of the intermediate feature layer of the backbone network of the vehicle-to-everything (V2X) cooperative perception model. Connect a series of dual-path adapters, a collaborative intelligent prompter, a knowledge compressor, and a knowledge injector at the output end of the backbone network of the vehicle-to-everything (V2X) collaborative perception model. The parameters of the backbone network are frozen, and the environmental data collected by each vehicle node in the target vehicle node set are input into the vehicle network cooperative perception model. The dual-path adapter, cooperative intelligent prompter, knowledge compressor, knowledge injector and detection head are trained until the target detection loss value is minimized, and the vehicle network cooperative perception model with efficient parameter fine-tuning is obtained.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the efficient domain adaptive method for cooperative perception parameters in vehicle networking as described in any one of claims 1 to 7.

Images