Vehicle infrastructure cooperative sensing method and device, computer equipment and medium

Abstract

The invention discloses a vehicle infrastructure cooperative sensing method and device, computer equipment and a medium, and the method comprises the steps: building personalized neural network elements through a differentiable architecture search, and obtaining a discretization architecture of a feature maintaining unit and a dimension reduction unit; determining the network depth based on the discretization architecture and the local data of the vehicle-mounted node; constructing a local personalized model based on the discretization architecture and the network depth, and establishing an operation alignment hierarchical aggregation mapping table M based on the unit sequence and the model parameter of each vehicle-mounted node; and finally, performing model aggregation of operation alignment driving based on the aggregation mapping table M, updating model parameters of each vehicle-mounted node to obtain a vehicle-road collaborative perception model, and performing vehicle-road collaborative perception by adopting the vehicle-road collaborative perception model. According to the invention, the cooperative sensing precision is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation, and in particular to a vehicle-road cooperative perception method, device, computer equipment, and medium. Background Technology

[0002] As urban intelligent transportation management evolves from traditional standardized approaches to deeply personalized customization, the demand for differentiated perception services from traffic participants is experiencing explosive growth. Traditional centralized decision-making based on a one-size-fits-all standardized perception model struggles to meet the personalized needs of different vehicle nodes in modern vehicle-road cooperative systems in terms of task characteristics, hardware configuration, and application scenarios. This fundamental shift has given rise to a key challenge in the field of vehicle-road cooperative perception—perception bias. Perception bias stems from a deep-seated contradiction between personalized needs and standardized models. From a traffic management perspective, different vehicle nodes undertake drastically different perception tasks (e.g., advanced autonomous driving requires complex environment understanding, while economical assisted driving only requires basic obstacle detection). This task heterogeneity prevents a unified perception model from accurately adapting to the needs of all nodes, resulting in bias. Specifically, standardized models optimized for average scenarios show a significant decrease in perception accuracy (e.g., false detections, missed detections) under specific tasks or special environments (e.g., complex urban environments). This not only affects the accuracy of individual vehicle decisions but also severely undermines the perception consistency of multi-vehicle cooperation, jeopardizing the coordination and safety of the entire traffic management system. From the perspective of personalized services, the root cause of perception bias lies in the structural conflict between the isomorphic assumption of traditional federated learning and the heterogeneous reality of the in-vehicle environment.

[0003] Traditional federated learning requires all participating nodes to adopt the same model architecture and optimize a unified objective function. This isomorphic assumption is clearly unsuitable for vehicle-to-infrastructure (V2I) scenarios. Different vehicle nodes exhibit high heterogeneity in hardware configuration (computing power differences can reach tens of times) and perception tasks (from lane keeping to automatic parking). This dual heterogeneity means that a unified model architecture cannot accommodate individualized needs, creating a gap between standardized perception accuracy and personalized requirements, fundamentally restricting service quality and decision reliability. Improving the accuracy of V2I perception has become a pressing problem. Summary of the Invention

[0004] This invention provides a vehicle-road cooperative sensing method, device, computer equipment, and storage medium to improve the accuracy of vehicle-road cooperative sensing.

[0005] To address the aforementioned technical problems, embodiments of this application provide a vehicle-road cooperative perception method, including: Personalized neural network primitives are constructed using a differentiable architecture search, resulting in a discretized architecture for feature-preserving units and dimensionality-reducing units; Based on the discretized architecture and local data from the vehicle nodes, the network depth d is determined.i This includes: computational task-data joint complexity; calculating the optimal network depth based on the joint complexity, local dataset size, and preset parameters; Based on the discretized architecture and network depth, a local personalized model is constructed, including: constructing a unit sequence with the network depth as the depth, wherein the unit types in the sequence consist of feature-preserving units and dimensionality-reducing units in the discretized architecture; determining the position of the dimensionality-reducing units according to the preset number of stages S, and generating model parameters; Based on the unit sequence and model parameters of each vehicle node, an operation-aligned hierarchical aggregation mapping table M is established; Based on the aggregation mapping table M, model aggregation driven by operation alignment is performed to update the model parameters of each vehicle node, thereby obtaining the vehicle-road cooperative perception model, and then using the vehicle-road cooperative perception model for vehicle-road cooperative perception.

[0006] Optionally, a differentiable architecture search is used to construct personalized neural network primitives, resulting in discretized architectures for feature-preserving units and dimensionality-reduction units, including: The optimal feature-preserving unit and dimensionality-reduction unit are searched using a two-layer optimization strategy; The architecture parameters are discretized to generate a discretized architecture.

[0007] Optionally, based on the unit sequence and model parameters Θi of each vehicle node, an operation-aligned hierarchical aggregation mapping table M is established, including: After each vehicle node is deployed locally, a corresponding layer operation mapping sequence is generated. , Let i be the operation mapping sequence of the i-th layer. ,in, Represents vehicle-mounted nodes In the The set of operations used by the layer; For each vehicle node's mapping sequence Generate an alignment sequence ; Alignment sequence based on all vehicle nodes Construct an aggregate mapping table.

[0008] Optionally, for each vehicle node's mapping sequence Generate an alignment sequence include: To align the layer counts of all vehicle nodes, the cloud server calculates the global maximum layer count. : For insufficient number of layers The vehicle-mounted node adopts an empty set Fill in the missing layers; For each vehicle node's mapping sequence The generated length is Alignment sequence : Among them, indicator function Defined as: .

[0009] Optionally, operation alignment-driven model aggregation based on the aggregation map M includes: Each vehicle node uses its local vehicle dataset for local training to obtain local model parameters; Each onboard node calculates gradient information that matches the specified operation in the aggregation mapping table; Aggregate the local model parameters corresponding to all gradient information that meet the aggregation conditions according to the aggregation mapping table.

[0010] To address the aforementioned technical problems, this application also provides a vehicle-road cooperative sensing device, comprising: The network building module is used to search and construct personalized neural network primitives using differentiable architectures, resulting in discretized architectures of feature-preserving units and dimensionality-reducing units; The depth determination module is used to determine the network depth d based on the discretized architecture and local data from the vehicle nodes. i This includes: the computational task-data joint complexity Ωi; and calculating the optimal network depth d based on Ωi, the local dataset size |Di|, and preset parameters ε0, β, κ, λO. i ; The model building module is used to build upon the discretized architecture and the network depth d. i Build a local personalized model, including: using d i To construct a deep unit sequence, the unit types in the sequence consist of feature-preserving units and dimensionality-reducing units in the discretization architecture; the position of the dimensionality-reducing unit is determined according to the preset number of stages S, and model parameters Θi are generated; The alignment mapping module is used to establish an operation alignment hierarchical aggregation mapping table M based on the unit sequence and model parameters Θi of each vehicle node; The aggregation update module is used to perform operation alignment-driven model aggregation based on the aggregation mapping table M, update the model parameters Θi of each vehicle node, obtain the vehicle-road cooperative perception model, and use the vehicle-road cooperative perception model to perform vehicle-road cooperative perception.

[0011] Optionally, the network building module includes: The search unit is used to search for the optimal feature-preserving unit and dimensionality-reduction unit through a two-layer optimization strategy. Discrete units are used to discretize the architecture parameters to generate the discretized architecture.

[0012] Optionally, the alignment mapping module includes: The operation mapping sequence generation unit is used to generate the corresponding layer's operation mapping sequence after each on-board node has been deployed locally. , Let i be the operation mapping sequence of the i-th layer. ,in, Represents vehicle-mounted nodes In the The set of operations used by the layer; Alignment sequence generation unit, used for mapping sequence for each vehicle node. Generate an alignment sequence ; Aggregate mapping table construction unit, based on the alignment sequence of all vehicle nodes. Construct an aggregate mapping table.

[0013] Optionally, the aggregation update module includes: The local training unit is used by each vehicle node to train the model locally using its local vehicle dataset and obtain local model parameters. The gradient upload unit is used for each vehicle node to calculate gradient information that matches the specified operation in the aggregation mapping table; The parameter aggregation unit is used to aggregate the local model parameters corresponding to all gradient information that meet the aggregation conditions according to the aggregation mapping table.

[0014] To address the aforementioned technical problems, this application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the aforementioned vehicle-road cooperative perception method.

[0015] To address the aforementioned technical problems, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned vehicle-road cooperative perception method.

[0016] The vehicle-road cooperative perception method, device, computer equipment, and storage medium provided in this invention construct personalized neural network primitives through differentiable architecture search, obtaining a discretized architecture of feature-preserving units and dimensionality-reduction units; based on the discretized architecture and local data from vehicle nodes, the network depth d is determined. i Based on discretized architecture and network depth d iA local personalized model is constructed, and an operation-aligned hierarchical aggregation mapping table M is established based on the unit sequence and model parameters of each vehicle node. Finally, operation-alignment-driven model aggregation is performed based on the aggregation mapping table M to update the model parameters of each vehicle node, resulting in a vehicle-road cooperative perception model. This model is then used for vehicle-road cooperative perception. This implementation achieves an aggregation mapping mechanism based on operation and hierarchical alignment strategies. A weighted fusion method is used to aggregate model parameters at matching operation and hierarchical positions, solving the model integration problem between heterogeneous vehicle nodes and thus improving the accuracy of cooperative perception. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an exemplary system architecture diagram to which this application can be applied; Figure 2 This is a flowchart of an embodiment of the vehicle-road cooperative perception method of this application; Figure 3 This is a schematic diagram of a structure of an embodiment of the vehicle-road cooperative sensing device according to this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0020] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 ,like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, and 103, a network 104, and a server 105. Network 104 serves as the medium for providing communication links between terminal devices 101, 102, and 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0023] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc.

[0024] Terminal devices 101, 102, and 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers, etc.

[0025] Server 105 can be a server that provides various services, such as a backend server that supports the pages displayed on terminal devices 101, 102, and 103.

[0026] It should be noted that the vehicle-road cooperative perception method provided in this application embodiment is executed by the server, and correspondingly, the vehicle-road cooperative perception device is set in the server.

[0027] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included. The terminal devices 101, 102, and 103 in this embodiment can specifically correspond to application systems in actual production.

[0028] Please see Figure 2 , Figure 2This invention illustrates a vehicle-road cooperative perception method provided by an embodiment of the present invention, which is applied to... Figure 1 Taking the server-side as an example, the details are as follows: S201: Differentiable architecture search is used to construct personalized neural network primitives, resulting in a discretized architecture of feature-preserving units and dimensionality-reducing units.

[0029] Optionally, in one specific implementation, a differentiable architecture search is used to construct personalized neural network primitives, resulting in a discretized architecture for feature-preserving units and dimensionality-reducing units, including: The optimal feature-preserving unit and dimensionality-reduction unit are searched using a two-layer optimization strategy; The architecture parameters are discretized to generate a discretized architecture.

[0030] Specifically, this step, based on the DARTS (Differentiable Architecture Search) framework, implements end-to-end differentiable architecture search through a two-layer optimization strategy to construct neural network units with structural adaptability. The network architecture consists of two types of complementary computational units: (1) The feature-preserving unit (Normal Cell) maintains the spatial dimension of the feature map through identity mapping; (2) Dimensionality reduction unit (Reduction Cell): Feature space downsampling is achieved using strided convolution. Each computational unit can be formalized as a directed acyclic graph (DAG). , where the node set Representation of intermediate features, edge set Define the connection relationships between candidate operations in the nodes. For the first... Nodes The characteristic evolution process of can be mathematically expressed as: ,in, This indicates a channel-level concatenation operation. For candidate operation set The first in Atom operators (such as 3×3 separable convolution, dilated convolution, etc.). These are learnable architecture weight parameters.

[0031] In this embodiment, to search for the optimal Normal Cell and Reduction Cell structures, the search process is modeled as including weight parameters. and architecture parameters The two-level optimization problem: Inner layer optimization (model weight update) is achieved by fixing the architecture parameters. Minimize empirical risk using the training set: .

[0032] Outer-layer optimization (architecture parameter update) is based on optimizing structural differentiability using a validation set, and calculating the architecture gradient through implicit function differentiation. : 。 go through After several alternating optimization iterations, the parameters of the continuous architecture are discretized using the Gumbel-Softmax reparameterization technique to determine the final cell structure and generate the discretized architecture. in, Random noise following a standard Gumbel distribution, (0,1): a Gumbel distribution with a location parameter of 0 and a scale parameter of 1, used for Gumbel-Softmax reparameterization to discretize the parameters of a continuous architecture.

[0033] S202: Determining network depth d based on discretized architecture and local data from vehicle nodes. i This includes: the computational task-data joint complexity Ωi; and calculating the optimal network depth d based on Ωi, the local dataset size |Di|, and preset parameters ε0, β, κ, λO. i .

[0034] Specifically, in step S201, the optimal Normal Cell and Reduction Cell are determined using DARTS, but the number of stacked cells (i.e., network depth) remains undetermined. This embodiment proposes an adaptive network depth determination method based on VC dimension theory to dynamically determine the number of stacked DARTS-generated units. This method models network depth optimization as a feasible solution search problem under the constraints of task-data joint complexity and model capacity, establishing a quantitative mapping relationship between task requirements, data distribution, and model structure. The specific implementation is as follows: (1) Local model and task-data joint complexity Each vehicle node They all have a local model ,in Represents a set of parameters. For model architecture. Through model depth To characterize its complexity, and assume the model architecture. Belongs to a depth of hypothesis space Vehicle-mounted node The model function on can be represented as .in, Indicates having depth The hypothesis space is defined. Meanwhile, to measure the fit between the task and the data, the joint task-data complexity is defined. It takes into account the relationship between task requirements and dataset characteristics: ,in The algorithm for estimating human manifold is adopted, and the KL divergence quantification model architecture is used. With data distribution The degree of compatibility.

[0035] (2) Optimal depth selection based on minimizing structural risk Suppose the complexity function has the following linear form: in, These are constant coefficients. Based on this, the objective of this step is to minimize the following structural risks: , ( For the model Loss of experience on local data; The structural risk coefficient balances empirical risk and complexity penalty; The linear coefficient constant of the complexity function; Network depth (number of layers); (This is the size of the local dataset for vehicle node i.) The Structural Risk Minimization (SRM) criterion is used to select the optimal model depth for each onboard node. And assume each vehicle node The learning process follows these objectives: ,in, For the model Empirical risk on local in-vehicle datasets measures the model's error on those datasets. Indicates depth The complexity of the time architecture (using the upper bound of the VC dimension) is used to penalize models with high complexity. To determine the structural risk coefficient and avoid overfitting of the model, This refers to the scale of local vehicle data.

[0036] To ensure the model's adaptability to different tasks and data environments, this invention sets two important constraints: (1) Lower limit of expressive ability constraint: in, Let's assume the space complexity is an inverse function. The task sensitivity coefficient. Represents the set of operations The average complexity.

[0037] (2) Upper limit of generalization constraint: in, This indicates a single-tier infrastructure. This is the preset generalization error threshold.

[0038] (3) Derivation of closed-loop solution By simultaneously applying the lower bound of the expressive power constraint and the upper bound of the generalization constraint, the optimal depth is obtained: The detailed derivation process is as follows: Due to the complexity of the model There exists a linear relationship, and its inverse function is... , will Substitute into the expression ability constraint formula The minimum depth can be obtained. Similarly, the complexity of a single-layer infrastructure is known to be... Substitute it into the generalization constraint. The maximum depth can be obtained. .

[0039] To balance expressive power and generalization, this invention uses the geometric mean method to solve for the optimal depth. Its closed-form solution is: Further simplification and integerization yield the final optimal depth: ,in Indicates the scale of local vehicle data Rounding operation to ensure network depth It is an integer.

[0040] S203: Based on Discretized Architecture and Network Depth i Build a local personalized model, including: using d i To construct a deep unit sequence, the unit types in the sequence consist of feature-preserving units and dimensionality-reducing units in the discretized architecture; the position of the dimensionality-reducing unit is determined according to the preset number of stages S, and the model parameters Θi are generated.

[0041] In a local personalized model building process, the vehicle node The local model consists of a depth of The unit sequence is composed of, denoted as The Normal Cells and Reduction Cells in the sequence are denoted as follows: and .

[0042] Because the model needs to downsample at specific layers to adjust the dimension and scale of features, deep neural network models are therefore divided into... Each stage has a boundary defined by [the following]. Location set Sure: , For the location index of the network layer; For nodes Total network depth; Number the current stage, where, This indicates the number of stages, which is usually set to the standard value. Used to determine the relative position of the Reduction Cell in each stage, symbol Indicates to The floor operation.

[0043] At the same time, in order to distinguish different types Unit, this step introduces indicator functions The element type at each location is dynamically determined, as defined below: Therefore, the first indivual The formal definition of an operational unit is: In the formula, For unit learnable parameters, and Sharing the same topological search space but employing different computational strategies, the former achieves feature map downsampling by introducing separable convolutions with a stride of 2. The local model... Reverse stacking of deep neural network models using function composition: Reverse stacking symbols are used here. Emphasizing the directionality of feature propagation and the composition order from right to left, the overall parameters of the local model are composed of the parameters of each unit, denoted as... , of which each Indicates the first The parameters of each unit, Let its parameter dimensions be defined. Therefore, the local model parameters are represented as: .

[0044] S204: Based on the unit sequence and model parameters Θi of each vehicle node, establish an operation-aligned hierarchical aggregation mapping table M.

[0045] Optionally, in one specific embodiment, establishing an operation-aligned hierarchical aggregation mapping table M based on the unit sequence and model parameters Θi of each vehicle node includes: After each vehicle node is deployed locally, a corresponding layer operation mapping sequence is generated. , Let i be the operation mapping sequence of the i-th layer. ,in, Represents vehicle-mounted nodes In the The set of operations used by the layer; For each vehicle node's mapping sequence Generate an alignment sequence ; Alignment sequence based on all vehicle nodes Construct an aggregate mapping table.

[0046] Optionally, in one specific implementation, for each vehicle node's mapping sequence Generate an alignment sequence include: To align the layer counts of all vehicle nodes, the cloud server calculates the global maximum layer count. : For insufficient number of layers The vehicle-mounted node adopts an empty set Fill in the missing layers; For each vehicle node's mapping sequence The generated length is Alignment sequence : . An empty set indicates "no operation" or "empty layer," used to fill insufficient layers. The missing layer of vehicle-mounted nodes.

[0047] Among them, indicator function Defined as: .

[0048] It should be noted that this step ensures that the layer-operation mapping sequence of all vehicle nodes has the same length, which facilitates subsequent unified processing.

[0049] Furthermore, based on the alignment sequence of all vehicle nodes Constructing the aggregation mapping table includes: Based on the mapping sequence after alignment of all vehicle nodes Cloud server builds aggregate mapping table : This mapping table records the set of operations selected by each vehicle node at each layer of the model.

[0050] In addition, each This can be further expanded to represent the vehicle-mounted node. In the The specific set of operations for a layer: ,in, Represents vehicle-mounted nodes In the The first layer One operation.

[0051] S205: Based on the aggregation mapping table M, perform operation alignment-driven model aggregation, update the model parameters Θi of each vehicle node, obtain the vehicle-road cooperative perception model, and use the vehicle-road cooperative perception model to perform vehicle-road cooperative perception.

[0052] Optionally, in one specific embodiment, operation alignment-driven model aggregation based on the aggregation mapping table M includes: Each vehicle node uses its local vehicle dataset for local training to obtain local model parameters; Each onboard node calculates gradient information that matches the specified operation in the aggregation mapping table; Aggregate the local model parameters corresponding to all gradient information that meet the aggregation conditions according to the aggregation mapping table.

[0053] Specifically, under the operation alignment hierarchical aggregation mapping mechanism in step S204, each vehicle node Perform personalized training locally, and then achieve global aggregation of model parameters through the following steps: (1) Local model training: In the first stage In each round of iteration, each on-board node Using its local vehicle dataset Update its model parameters The update rules are as follows: The loss function is defined as follows: , Represents the single-sample loss function. and These are vehicle-mounted nodes In the The input and output of each sample.

[0054] (2) Upload gradients of each local model: each vehicle node Calculation and Aggregation Mapping Specify gradient information for operation matching: ,in, Indicates vehicle-mounted node In the The gradient of the layer parameters.

[0055] (3) Global aggregation across heterogeneous models: The cloud server uses the aggregation mapping table Aggregate all vehicle node model parameters that satisfy the aggregation conditions. Specifically, for a given layer... and operation Aggregated in the vehicle node set Upward: ,in, Indicates the first Layer usage operations The number of vehicle-mounted nodes.

[0056] (4) Model deployment and local updates: The cloud server will aggregate the global model parameters Transmitted back to the corresponding vehicle node and update its local model parameters: .

[0057] In this embodiment, a differentiable architecture search is used to construct personalized neural network primitives, resulting in a discretized architecture of feature-preserving units and dimensionality-reduction units. Based on the discretized architecture and local data from the vehicle nodes, the network depth d is determined. i Based on discretized architecture and network depth d i A local personalized model is constructed, and an operation-aligned hierarchical aggregation mapping table M is established based on the unit sequence and model parameters Θi of each vehicle node. Finally, based on the aggregation mapping table M, model aggregation driven by operation alignment is performed to update the model parameters Θi of each vehicle node, resulting in a vehicle-road cooperative perception model. This model is then used for vehicle-road cooperative perception. An aggregation mapping mechanism based on operation and hierarchical alignment strategies is implemented. A weighted fusion method is used to aggregate model parameters at matching operation and hierarchical positions, solving the model integration problem between heterogeneous vehicle nodes and thus improving the accuracy of cooperative perception.

[0058] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0059] Figure 3 A schematic block diagram of a vehicle-road cooperative sensing device corresponding to the vehicle-road cooperative sensing method described in the above embodiments is shown. Figure 3 As shown, the vehicle-road cooperative perception device includes a network construction module 31, a depth determination module 32, a model construction module 33, an alignment and mapping module 34, and an aggregation and update module 35. Detailed descriptions of each functional module are as follows: Network building module 31 is used to search and build personalized neural network primitives using differentiable architectures to obtain the discretized architecture of feature-preserving units and dimensionality-reducing units; Depth determination module 32 is used to determine the network depth d based on the discretized architecture and local data from the vehicle nodes. i This includes: the computational task-data joint complexity Ωi; and calculating the optimal network depth d based on Ωi, the local dataset size |Di|, and preset parameters ε0, β, κ, λO. i ; Model building module 33 is used for models based on discretized architectures and network depth d. i Build a local personalized model, including: using d i To construct a deep unit sequence, the unit types in the sequence consist of feature-preserving units and dimensionality-reducing units in the discretized architecture; the position of the dimensionality-reducing unit is determined according to the preset number of stages S, and the model parameters Θi are generated; Alignment mapping module 34 is used to establish an operation alignment hierarchical aggregation mapping table M based on the unit sequence and model parameters Θi of each vehicle node; The aggregation update module 35 is used to perform operation alignment-driven model aggregation based on the aggregation mapping table M, update the model parameters Θi of each vehicle node, obtain the vehicle-road cooperative perception model, and use the vehicle-road cooperative perception model to perform vehicle-road cooperative perception.

[0060] Optionally, network building module 31 includes: The search unit is used to search for the optimal feature-preserving unit and dimensionality-reduction unit through a two-layer optimization strategy. Discrete units are used to discretize architectural parameters and generate discretized architectures.

[0061] Optionally, the alignment mapping module 34 includes: The operation mapping sequence generation unit is used to generate the corresponding layer's operation mapping sequence after each on-board node has been deployed locally. , Let i be the operation mapping sequence of the i-th layer. ,in, Represents vehicle-mounted nodes In the The set of operations used by the layer; Alignment sequence generation unit, used for mapping sequence for each vehicle node. Generate an alignment sequence ; Aggregate mapping table construction unit, based on the alignment sequence of all vehicle nodes. Construct an aggregate mapping table.

[0062] Optionally, the aggregation update module 35 includes: The local training unit is used by each vehicle node to train the model locally using its local vehicle dataset and obtain local model parameters. The gradient upload unit is used for each vehicle node to calculate gradient information that matches the specified operation in the aggregation mapping table; The parameter aggregation unit is used to aggregate the local model parameters corresponding to all gradient information that meet the aggregation conditions according to the aggregation mapping table.

[0063] Specific limitations regarding the vehicle-road cooperative sensing device can be found in the limitations of the vehicle-road cooperative sensing method described above, and will not be repeated here. Each module in the aforementioned vehicle-road cooperative sensing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0064] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference needed] for details. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0065] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components connected to the memory 41, processor 42, and network interface 43 is shown in the figure; however, it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0066] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0067] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 4. Of course, the memory 41 may include both the internal storage unit and its external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as the program code of the vehicle-road cooperative perception method. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0068] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run program code stored in the memory 41 or process data, for example, to run program code for a vehicle-to-everything (V2X) cooperative sensing method.

[0069] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0070] This application also provides another embodiment, namely, a computer-readable storage medium storing an interface display program that can be executed by at least one processor to cause the at least one processor to perform the steps of the vehicle-road cooperative perception method as described above.

[0071] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0072] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

Claims

1. A vehicle-road cooperative perception method, characterized in that, include: Personalized neural network primitives are constructed using a differentiable architecture search, resulting in a discretized architecture for feature-preserving units and dimensionality-reducing units; Based on the discretized architecture and local data of the vehicle nodes, the network depth is determined, including: the joint complexity of the computation task and the data; and the calculation of the network depth based on the joint complexity, the size of the local dataset, and preset parameters. Based on the discretization architecture and the network depth, a local personalized model is constructed, including: constructing a unit sequence with the network depth as the depth, wherein the unit types in the sequence consist of feature-preserving units and dimensionality-reduction units in the discretization architecture; determining the position of the dimensionality-reduction units according to a preset number of stages S, and generating model parameters; Based on the unit sequence and model parameters of each vehicle node, an operation-aligned hierarchical aggregation mapping table M is established; Based on the aggregation mapping table M, model aggregation driven by operation alignment is performed, the model parameters of each vehicle node are updated to obtain the vehicle-road cooperative perception model, and the vehicle-road cooperative perception model is used for vehicle-road cooperative perception.

2. The vehicle-road cooperative perception method as described in claim 1, characterized in that, The method of using differentiable architecture search to construct personalized neural network primitives, resulting in the discretized architecture of feature-preserving units and dimensionality-reducing units, includes: The optimal feature-preserving unit and dimensionality-reduction unit are searched using a two-layer optimization strategy; The architecture parameters are discretized to generate the discretized architecture.

3. The vehicle-road cooperative perception method as described in claim 1, characterized in that, The establishment of the operation-aligned hierarchical aggregation mapping table M based on the unit sequence and model parameters Θi of each vehicle node includes: After each vehicle node is deployed locally, a corresponding layer operation mapping sequence is generated. , Let i be the operation mapping sequence of the i-th layer. ,in, Represents vehicle-mounted nodes In the The set of operations used by the layer; For each vehicle node's mapping sequence Generate an alignment sequence ; Alignment sequence based on all vehicle nodes Construct an aggregate mapping table.

4. The vehicle-road cooperative perception method as described in claim 1, characterized in that, The mapping sequence for each vehicle node Generate an alignment sequence include: To align the layer counts of all vehicle nodes, the cloud server calculates the global maximum layer count. : For insufficient number of layers The vehicle-mounted node adopts an empty set Fill in the missing layers; For each vehicle node's mapping sequence The generated length is Alignment sequence : Among them, indicator function Defined as: .

5. The vehicle-road cooperative perception method as described in claim 1, characterized in that, The model aggregation driven by operation alignment based on the aggregation mapping table M includes: Each vehicle node uses its local vehicle dataset for local training to obtain local model parameters; Each onboard node calculates gradient information that matches the specified operation in the aggregation mapping table; Aggregate the local model parameters corresponding to all gradient information that meet the aggregation conditions according to the aggregation mapping table.

6. A vehicle-road cooperative sensing device, characterized in that, include: The network building module is used to search and construct personalized neural network primitives using differentiable architectures, resulting in discretized architectures of feature-preserving units and dimensionality-reducing units; The depth determination module is used to determine the network depth di based on the discretized architecture and local data of the vehicle nodes, including: calculating the joint complexity of the task and the data; and calculating the optimal network depth according to the joint complexity, the size of the local dataset, and preset parameters. The model building module is used to build a local personalized model based on the discretized architecture and the network depth, including: building a unit sequence based on the depth, wherein the unit types in the sequence consist of feature-preserving units and dimensionality-reducing units in the discretized architecture; determining the position of the dimensionality-reducing units according to a preset number of stages S; and generating model parameters. The alignment mapping module is used to establish an operation alignment hierarchical aggregation mapping table M based on the unit sequence and model parameters of each vehicle node; The aggregation update module is used to perform operation alignment-driven model aggregation based on the aggregation mapping table M, update the model parameters of each vehicle node, obtain the vehicle-road cooperative perception model, and use the vehicle-road cooperative perception model to perform vehicle-road cooperative perception.

7. The vehicle-road cooperative sensing device as described in claim 6, characterized in that, The network construction module includes: The search unit is used to search for the optimal feature-preserving unit and dimensionality-reduction unit through a two-layer optimization strategy. Discrete units are used to discretize the architecture parameters to generate the discretized architecture.

8. The vehicle-road cooperative sensing device as described in claim 6, characterized in that, The aggregation update module includes: The local training unit is used by each vehicle node to train the model locally using its local vehicle dataset and obtain local model parameters. The gradient upload unit is used for each vehicle node to calculate gradient information that matches the specified operation in the aggregation mapping table; The parameter aggregation unit is used to aggregate the local model parameters corresponding to all gradient information that meet the aggregation conditions according to the aggregation mapping table.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the vehicle-road cooperative perception method as described in any one of claims 1 to 5.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the vehicle-road cooperative perception method as described in any one of claims 1 to 5.

Images