Decentralized collaborative sensing method and system based on zero-order query and local momentum

Abstract

The application relates to the technical field of Internet of Vehicles and edge intelligent computing, and discloses a decentralized collaborative perception method and system based on zero-order query and local momentum. The method comprises the following steps: initializing a local perception model and a local momentum of each vehicle node in the Internet of Vehicles and setting training parameters; estimating a local descending direction by applying random disturbance directions to the local perception model and performing function value query; updating the local momentum according to the local descending direction; performing singular value decomposition on the local momentum, and taking the product of left and right singular vector matrices as an orthogonalization updating direction; updating the local perception model along the direction; performing neighbor weighted mixing of the model and the momentum according to a current time-varying inter-vehicle communication topology, and cyclically iterating until collaborative convergence. The application does not require a central server, does not transmit local data, gradients and models to the center, adapts to dynamic changes of inter-vehicle connection relationships, and is efficient and stable in training.

Description

Technical Field

[0001] This invention belongs to the field of vehicle networking and edge intelligent computing technology, and specifically relates to a decentralized collaborative perception method and system based on zero-order query and local momentum. Background Technology

[0002] With the development of autonomous driving and intelligent connected vehicles, relying solely on onboard sensors for environmental perception in individual vehicles presents limitations such as obstructed view and limited sensing distance. Vehicle-to-everything (V2X) technology enables vehicles to share sensing information through vehicle-to-vehicle (V2V) communication, achieving multi-vehicle collaborative perception and improving the detection accuracy and coverage of targets such as pedestrians, vehicles, and obstacles.

[0003] However, in practical distributed vehicle-to-everything (V2X) applications, multi-vehicle collaborative training of perception models faces several challenges: First, there is usually a lack of a unified coordination center among vehicles, and the parameter aggregation, progress synchronization, and task scheduling during training lack a unified organizer; second, vehicles move at high speeds, and factors such as road environment obstruction and network switching cause the connection between vehicles to change dynamically and intermittently, resulting in limited communication bandwidth and link stability; third, the computing power, storage space, and energy consumption of onboard equipment are limited; fourth, onboard image data often contains sensitive information such as license plates, faces, driving routes, and location information, and parties are usually unwilling to hand over their local data or models to a central server or third party; fifth, different vehicles are in different environments, and the distribution of local data varies, so direct aggregation and use may affect the training effect; in addition, vehicles may join or leave the communication range at any time, and the scale of the network participating in training is also dynamically changing.

[0004] Currently, the training of vehicle-to-everything (V2X) cooperative perception models mainly adopts centralized or federated learning methods: a central server aggregates model updates or gradients uploaded by each vehicle, or multiple rounds of communication between vehicles and the server are used to complete the training of the global model. These methods typically require vehicles to upload gradients or model parameters to the server and rely on a central server or a fixed communication synchronization cycle. At the optimization algorithm level, optimizers used for deep network training mainly include stochastic gradient descent, momentum-based stochastic gradient descent, adaptive optimizers, and the Muon optimizer designed for matrix-form parameters. This optimizer uses momentum accumulation combined with orthogonalization of the update direction; the orthogonalization refers to Singular Value Decomposition (SVD) orthogonalization of the momentum matrix. The Muon optimizer handles the model parameter update direction through "momentum accumulation + orthogonalization," effectively alleviating the gradient rank collapse problem in deep network training. Furthermore, in scenarios where gradients are inconvenient to directly obtain or propagate, methods have been developed to estimate the update direction using only function value (zero-order) information.

[0005] The aforementioned existing technologies have the following problems in the context of vehicle-to-everything (V2X) collaborative perception: In scenarios where local models or gradients need to be uploaded to a central server, there are privacy concerns, and training is interrupted if the central node fails; In scenarios where vehicles move at high speeds and connection relationships change dynamically, existing methods often assume fixed topology or periodic synchronization, which is difficult to apply directly and can easily lead to training instability or failure; Existing distributed methods usually require globally synchronized momentum or auxiliary states, resulting in high communication overhead and making it difficult to deploy in a fully decentralized dynamic environment; Due to the different distribution of local data among vehicles and the different update directions of different nodes, direct mixing may cause mutual influence or even cancellation, affecting convergence speed and stability.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address or at least alleviate one or more of the above problems, a decentralized collaborative sensing method and system based on zero-order query and local momentum is provided. This method does not require a central server or transmit local models and gradients to a central server. It can adapt to dynamic changes in the connection relationships between vehicles and the training process is efficient and stable.

[0008] To achieve the above objectives, according to a first aspect of the present invention, a decentralized collaborative sensing method based on zero-order query and local momentum is provided, comprising the following steps:

[0009] S1. Initialize the local perception model and local momentum of each vehicle node in the vehicle network, and set the training parameters, including step size, momentum coefficient, smoothing parameter, mini-batch size, number of random directions, number of iterations, and time-varying hybrid matrix sequence of the vehicle network.

[0010] S2. Each vehicle node applies a random perturbation direction to the local perception model based on local data and queries the function value. It estimates the local descent direction based on the difference between the function value of the perturbation point and the current point, and obtains the zero-order descent direction estimate.

[0011] S3. Each vehicle node performs a weighted accumulation of the zero-order descent direction estimate and the historical local momentum, and updates the local momentum.

[0012] S4. Each vehicle node performs singular value decomposition on the updated local momentum, and takes the product of the left singular vector matrix and the right singular vector matrix as the orthogonal update direction to make the update amplitude of each singular value direction balanced.

[0013] S5. Each vehicle node updates its local perception model along the orthogonalization update direction to obtain an intermediate model.

[0014] S6. Each vehicle node performs a weighted mixing of its local intermediate model and the intermediate models of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain an updated local perception model.

[0015] S7. Each vehicle node weights and mixes its local momentum with that of its neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain the updated local momentum.

[0016] Repeat steps S2 to S7, iterating until all vehicle nodes converge collaboratively, to obtain a unified target perception model.

[0017] Furthermore, the local perception model and local momentum of each vehicle node in the vehicle network are initialized, and training parameters are set, including...

[0018] Set the initial perception model for each vehicle node.

[0019] The first zero-order descent direction estimate is calculated using the zero-order query estimation method and used as the initial local momentum value for each vehicle node.

[0020] The step size is set to be greater than zero.

[0021] The momentum coefficient is set, and its value is between zero and one, to weigh the historical momentum against the current zero-order descent direction estimate.

[0022] The smoothing parameter is set to be greater than zero, and is used to control the magnitude of the perturbation direction in the zero-order estimation.

[0023] Set the small batch size and the number of random directions.

[0024] Set the number of iteration rounds.

[0025] A time-varying dual random hybrid matrix is ​​determined for each round of the vehicle network, wherein the elements of the time-varying hybrid matrix reflect the communication connectivity and weights between vehicle nodes in the current round.

[0026] Furthermore, each vehicle node applies a random perturbation direction to the local perception model based on local data and queries the function value. It estimates the local descent direction based on the difference between the function value at the perturbation point and the current point, thus obtaining the zero-order descent direction estimate, including...

[0027] Each vehicle node retrieves its current local perception model from local storage and loads it into memory.

[0028] Sample small batches of data from the local dataset.

[0029] Generate several random perturbation directions.

[0030] For each random perturbation direction, the forward function value is queried for the perturbation point and the current point respectively to obtain the function value corresponding to the perturbation point and the function value corresponding to the current point, wherein the perturbation point is the sum of the current local perception model and the random perturbation direction.

[0031] The zero-order descent direction estimate is calculated by weighting and summing the differences between the function values ​​corresponding to the perturbation point and the function values ​​corresponding to the current point, according to the random perturbation direction.

[0032] Further, the step of calculating the zero-order descent direction estimate by weighted summation according to the random perturbation direction based on the difference between the function value corresponding to the perturbation point and the function value corresponding to the current point includes...

[0033] For each vehicle node, the difference between the disturbance point function value and the current point function value corresponding to each sample is multiplied by the ratio of the model parameter dimension to the smoothing parameter, and then multiplied by the random disturbance direction to obtain the directional component of the sample under the random disturbance direction.

[0034] Sum the directional components of all samples under the same random perturbation direction.

[0035] The summation of all random perturbation directions is divided by the product of the mini-batch size and the number of random directions to obtain the zero-order descent direction estimate.

[0036] Furthermore, each vehicle node performs a weighted accumulation of the zero-order descent direction estimate and the historical local momentum to update the local momentum, including...

[0037] Read the local momentum of each vehicle node in the previous round.

[0038] The zero-order descent direction estimate is weighted and accumulated with the historical local momentum according to the momentum coefficient to obtain the updated local momentum.

[0039] Furthermore, each vehicle node performs singular value decomposition on the updated local momentum, and takes the product of the left singular vector matrix and the right singular vector matrix as the orthogonalization update direction to make the update amplitude of each singular value direction balanced, including...

[0040] Singular value decomposition is performed on the updated local momentum to obtain the left singular vector matrix, the singular value matrix, and the right singular vector matrix.

[0041] The product of the left singular vector matrix and the transpose of the right singular vector matrix is ​​taken as the orthogonalization update direction to balance the update magnitude in each singular value direction.

[0042] Furthermore, each vehicle node updates its local perception model along the orthogonalization update direction to obtain an intermediate model, including: updating the local perception model along the orthogonalization update direction with the step size to obtain the intermediate model.

[0043] Furthermore, each vehicle node weighted mixes its local intermediate model with the intermediate models of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain an updated local perception model. Each vehicle node weighted mixes its local momentum with the momentum of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain an updated local momentum, including...

[0044] Each vehicle node calculates a weighted sum of its own intermediate model and the intermediate models of its neighboring vehicles based on the time-varying hybrid matrix corresponding to the current vehicle network communication topology, thus obtaining an updated local perception model.

[0045] Each vehicle node uses the same time-varying mixing matrix to perform a weighted summation of its own updated local momentum and the local momentum of its neighboring vehicles, thus obtaining the mixed local momentum.

[0046] The time-varying hybrid matrix is ​​a double random matrix, whose elements reflect the communication connectivity and weights between vehicle nodes in the current round.

[0047] Furthermore, the target perception model is a convolutional neural network target detection model; the method is applied to the collaborative training of road target detection models for autonomous vehicles, where each vehicle node holds a locally collected road image dataset, and the data distribution of each vehicle node is non-independent and identically distributed.

[0048] Repeat steps S2 to S7, iterating for a preset number of rounds until the local perception models of each vehicle node converge collaboratively to obtain the target perception model, and then use the target perception model for target detection, recognition and tracking tasks in vehicle network collaborative perception.

[0049] To achieve the above objectives, according to a second aspect of the present invention, a decentralized collaborative sensing system based on zero-order query and local momentum is provided, comprising:

[0050] The initialization module is used to initialize the local perception model and local momentum of each vehicle node in the vehicle network, and set the training parameters, including step size, momentum coefficient, smoothing parameter, mini-batch size, number of random directions, number of iterations, and time-varying hybrid matrix sequence of the vehicle network.

[0051] The zero-order query estimation module is used by each vehicle node to apply a random perturbation direction to the local perception model based on local data and perform function value queries. The local descent direction is estimated based on the difference between the function value of the perturbation point and the current point, thus obtaining the zero-order descent direction estimate.

[0052] The momentum update module is used by each vehicle node to perform a weighted accumulation of the zero-order descent direction estimate and the historical local momentum, and update the local momentum.

[0053] The orthogonalization module is used by each vehicle node to perform singular value decomposition on the updated local momentum. The product of the left singular vector matrix and the right singular vector matrix is ​​taken as the orthogonalization update direction to make the update amplitude of each singular value direction more balanced.

[0054] The local model update module is used to update the local perception model of each vehicle node along the orthogonal update direction to obtain an intermediate model.

[0055] The model fusion module is used by each vehicle node to weight and mix the local intermediate model with the intermediate models of neighboring vehicles according to the current time-varying inter-vehicle communication topology, so as to obtain the updated local perception model.

[0056] The momentum mixing module is used to weight and mix the local momentum of each vehicle node with the momentum of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain the updated local momentum.

[0057] The central control module is used to repeatedly trigger the zero-order query estimation module, momentum update module, orthogonalization module, local model update module, model mixing module, and momentum mixing module, iterating until all vehicle nodes converge collaboratively to obtain a unified target perception model.

[0058] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art.

[0059] In this invention, each vehicle node maintains only its local momentum, and model updates are achieved through information mixing among neighboring nodes. No central server is required, and local data, gradients, or models are not transmitted, effectively protecting data privacy and enhancing the feasibility and participation of practical deployments. This invention can adapt to dynamic topology changes caused by high-speed vehicle movement, ensuring training convergence and making it suitable for real-world vehicle-to-everything (V2X) environments. In terms of optimization strategy, by fusing momentum accumulation and singular value decomposition orthogonalization, the update magnitudes of each singular direction are balanced, mitigating gradient rank collapse and direction cancellation problems, thereby significantly improving convergence speed and training stability while reducing communication rounds. Furthermore, this invention has good scalability, adapting to changes in the number of nodes and network size, and its gradient estimation and orthogonalization implementation methods are flexible and diverse, facilitating large-scale application.

[0060] Specifically, this invention is completely decentralized, with each vehicle node maintaining its momentum locally and achieving collaborative training only through information mixing with neighboring nodes. It does not rely on a central server and does not transmit local data, gradients, or local models to the center. It is practical and feasible, and helps to dispel the data concerns of the participants, thus encouraging all parties to participate in the collaboration.

[0061] Specifically, this invention adapts to real-world scenarios where the connectivity between vehicles changes over time. Training can still converge even when vehicle movement causes changes in the link, making it usable in real-world vehicle networking environments.

[0062] Specifically, this invention updates the model by combining momentum accumulation with singular value decomposition orthogonalization, making the update magnitude more balanced in each singular value direction, alleviating gradient rank collapse, reducing directional cancellation caused by differences in local data distribution, thereby accelerating convergence, improving training stability, and achieving the same effect with fewer communication rounds.

[0063] Specifically, this invention has good scalability, can still work normally when the number of nodes and the network size change, is easy to deploy on a large scale, and has diverse gradient estimation methods and orthogonalization implementation methods, making it highly scalable.

[0064] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0065] The accompanying drawings, as part of this invention, are provided to further illustrate the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation thereof. Clearly, the drawings described below are merely some embodiments; those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0066] Figure 1 This is a logical schematic diagram of the decentralized collaborative sensing method based on zero-order query and local momentum in this specific embodiment.

[0067] Figure 2 This is a flowchart illustrating the decentralized collaborative sensing method based on zero-order query and local momentum in this specific embodiment.

[0068] Figure 3 This is a schematic diagram of the evolution of the time-varying communication topology of the vehicle network in this specific embodiment, wherein (a) is a schematic diagram of the connection relationship between vehicles at time t; (b) is a schematic diagram of the connection relationship between vehicles at time t+1; and (c) is a schematic diagram of the connection relationship between vehicles at time t+2.

[0069] Figure 4 This is a schematic diagram of the decentralized collaborative sensing system based on zero-order query and local momentum in this specific embodiment. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0071] In the following description, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0072] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0073] Please see Figure 1 This invention provides a decentralized collaborative sensing method based on zero-order query and local momentum, comprising the following steps.

[0074] S1. Initialize the local perception model and local momentum of each vehicle node in the vehicle network, and set the training parameters, including step size, momentum coefficient, smoothing parameter, mini-batch size, number of random directions, number of iterations, and the time-varying hybrid matrix sequence of the vehicle network.

[0075] S2. Each vehicle node applies a random perturbation direction to the local perception model based on local data and queries the function value. It estimates the local descent direction based on the difference between the function value of the perturbation point and the current point, and obtains the zero-order descent direction estimate.

[0076] S3. Each vehicle node weights and accumulates the zero-order descent direction estimate with the historical local momentum to update the local momentum.

[0077] S4. Each vehicle node performs singular value decomposition on the updated local momentum, and takes the product of the left singular vector matrix and the right singular vector matrix as the orthogonal update direction to make the update amplitude of each singular value direction balanced.

[0078] S5. Each vehicle node updates its local perception model along the orthogonalization update direction to obtain an intermediate model.

[0079] S6. Each vehicle node performs a weighted mixing of its local intermediate model and the intermediate models of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain an updated local perception model.

[0080] S7. Each vehicle node weights and mixes its local momentum with that of its neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain the updated local momentum.

[0081] Repeat steps S2 to S7, iterating until all vehicle nodes converge collaboratively, to obtain a unified target perception model.

[0082] In a fully decentralized vehicle network, each vehicle node independently maintains its local perception model and local momentum state. It estimates the local descent direction by querying the zero-order function value, obtains the update direction through singular value decomposition of the momentum term, and updates the local model. The model and momentum are then mixed using a neighbor-weighted method according to the current time-varying inter-vehicle communication method, and the process is iterated until convergence.

[0083] Further research by the inventors revealed that directly applying zero-order estimation based on function value lookup, momentum accumulation, singular value decomposition orthogonalization, and neighbor-weighted hybrid methods to collaborative perception training in decentralized, time-varying vehicle networks will generate the following new technical problems, making it difficult to function properly: First, the numerical amplitude and random noise of the zero-order estimator increase significantly with the dimensionality of the model parameters. The zero-order descent direction estimator based on function value difference contains an amplification factor for the ratio of parameter dimensionality to smoothing parameter, and requires successive function value lookups for multiple random directions. In cases where the deep perception model has a large parameter scale and the smoothing parameter has a small value, the amplitude of the estimator is greatly amplified, and its noise variance also increases with the dimensionality. If this type of estimator is used directly to update the deep model, the training step size must be reduced accordingly with the dimensionality; otherwise, the model update will diverge and the training will oscillate. In decentralized scenarios where the amount of local data for each vehicle is limited and the data distribution is not independent and identically distributed, it is difficult to determine a step size that simultaneously adapts to all nodes and remains stable. Secondly, zero-order query noise is approximately isotropic in the direction spectrum, which can mask the inherent low-rank structure and main directions of deep network update directions. Furthermore, update directions in deep network training generally suffer from spectral degradation and imbalanced update amplitudes across directions. These two problems overlap, making it difficult to obtain stable and balanced effective update directions based solely on zero-order estimation. Thirdly, local momentum and orthogonalization are sensitive to the consistency of directions between nodes. Different vehicles have different local data distributions, resulting in systematic differences in their local momentum directions. If only neighbor mixing is performed on the model without momentum mixing, momentum differences will accumulate across wheels and propagate to the update directions obtained through orthogonalization, causing the update directions of neighboring vehicles to cancel each other out. Moreover, the update direction output by orthogonalization has a constant amplitude, so the cancellation effect is amplified by the full magnitude. Fourth, singular value decomposition is sensitive to noise, and orthogonalization must be based on momentum that has been smoothed over time and whose true direction is dominant. The order of momentum accumulation cannot be arbitrarily changed. At the same time, momentum, as a cross-wheel state, is repeatedly weighted by a series of time-varying hybrid matrices on a network with time-varying connections. Its divergence has cross-wheel accumulation and memory effects, and existing designs under fixed topology assumptions cannot be directly used.

[0084] Based on this, the present invention integrates zero-order query estimation, local momentum accumulation, momentum singular value decomposition orthogonalization, and neighbor-weighted mixing of the model and momentum as a whole, and coordinates each step in a specific order: Local momentum is used to accumulate the zero-order estimation direction containing query noise over time, thereby enhancing the coherence of the true descent direction, canceling out query noise, suppressing amplitude instability, and improving the spectral structure upon which subsequent orthogonalization depends; Singular value decomposition is performed on the accumulated momentum instead of the instantaneous zero-order estimation, and the product of the left and right singular vector matrices is taken as the update direction, normalizing the update direction to a bounded direction with balanced amplitude that is independent of parameter dimensions, smoothing parameters, and the magnitude of query noise, thus eliminating the amplitude amplification introduced by zero-order estimation and allowing the training step size to be stably selected at a conventional magnitude; By mixing local momentum and the local model together using a unified time-varying dual random mixing matrix for neighbor-weighted mixing, each vehicle node maintains consistency in the momentum state upon which the model and orthogonalization depend, suppressing directional divergence and cross-wheel accumulation caused by non-independent identically distributed data and time-varying connections.

[0085] Please see Figure 1 , Figure 1 This is a schematic diagram of the overall system logic of the decentralized collaborative sensing method. Figure 1 The upper part shows the decentralized connection relationship between multiple vehicle nodes (vehicle 1 to vehicle K) in the Internet of Vehicles (IoV) through V2V communication. Each vehicle node only exchanges information with neighboring vehicles within its communication range. Figure 1 The lower part takes vehicle node k as an example to illustrate its internal processing chain, which includes: local image dataset; zero-order query estimation, corresponding to formula (1), used to obtain the update direction estimate. Momentum update, corresponding to formula (2), is used to maintain momentum M locally. (k) Momentum SVD orthogonalization, corresponding to equations (3a) and (3b), is used to obtain the orthogonalization update direction Q. (k) Local model update, corresponding to equation (4), is used to update model X. (k) ; and model mixing and momentum mixing with surrounding vehicles, corresponding to equations (5) and (6). Figure 1 The update and hybridization links of model X are shown on the left, while the update and hybridization links of momentum M are shown on the right.

[0086] Please see Figure 2 The overall process of this method includes: initialization, zero-order query estimation, momentum update, orthogonalization, local model update, model neighbor mixing, momentum neighbor mixing, and iterative loop.

[0087] For example, this embodiment illustrates the collaborative training of a convolutional neural network (CNN) model for road object detection by K autonomous vehicles. Each vehicle node does not need to upload original image data or analyze gradients throughout the process; collaborative training is completed solely through function value queries and momentum mixing with its neighbors. Assume there are K vehicle nodes in the vehicle network. The k-th vehicle holds its locally collected road image dataset, and the data distribution of each vehicle is non-independent and identically distributed (Non-IID) (e.g., different vehicles travel on different road sections and at different times, resulting in differences in the image categories and scenes). The perception model to be trained for each vehicle is a CNN object detection or classification network with the same structure. The weights of one (or more) layers are organized in matrix form, denoted as X∈ .

[0088] (a) Initialization.

[0089] In some embodiments, the local perception model and local momentum of each vehicle node in the vehicle network are initialized, and training parameters are set, including: setting the initial perception model of each vehicle node; calculating the first zero-order descent direction estimate using the zero-order query estimation method, and using it as the initial value of the local momentum of each vehicle node; setting the step size, wherein the step size is greater than zero; setting the momentum coefficient, wherein the momentum coefficient takes a value between zero and one, used to balance the proportion between historical momentum and the current zero-order descent direction estimate; setting the smoothing parameter, wherein the smoothing parameter is greater than zero, used to control the amplitude of the perturbation direction in the zero-order estimation; setting the mini-batch size and the number of random directions; setting the number of iteration rounds; determining the time-varying double random mixing matrix for each round of the vehicle network, wherein the elements of the time-varying mixing matrix reflect the communication connectivity and weights between each vehicle node in the current round.

[0090] Specifically, each vehicle node initializes its local perception model matrix and local momentum matrix, and sets the training parameters: step size η, momentum coefficient β, smoothing parameter μ, and mini-batch size. Random direction number The number of iteration rounds T, and the time-varying hybrid matrix sequence of the vehicle network { }

[0091] Specifically, the initialization steps include: setting the initial model matrix for each vehicle node. The subscript 0 represents the initial wheel; each vehicle node can use the same or different initialization methods; the first zero-order estimator is calculated using the zero-order estimation method. And use it as the initial value of the local momentum matrix. Set the training step size η, where η > 0; set the momentum coefficient β, where β ∈ (0, 1), to balance the weight between historical momentum and the current zero-order estimate; set the smoothing parameter μ, where μ > 0, to control the magnitude of the perturbation direction in the zero-order estimate; set the mini-batch size. With random direction number Set the number of iteration rounds T; determine the time-varying double random mixture matrix for each round of the vehicle-to-everything (V2X) network. Its elements This reflects the communication connectivity and weights between vehicle k and vehicle j in round t. For example, each vehicle node may adopt one or more of the following initialization strategies, but not limited to: initialization using the same pre-trained model parameters; independent random initialization of each vehicle node according to a standard normal distribution; or initialization by a unified baseline model issued by the roadside unit.

[0092] For example, the step size η = 0.01, momentum coefficient β = 0.9, smoothing parameter μ = 0.001, and mini-batch size can be set. =32, random direction number =20. The above values ​​are only examples and can be adjusted by those skilled in the art according to actual scenarios.

[0093] (ii) Zero-order query estimation.

[0094] In some embodiments, each vehicle node applies a random perturbation direction to the local perception model based on local data and performs function value lookup. It then estimates the local descent direction based on the difference between the function value at the perturbation point and the current point, obtaining a zero-order descent direction estimate. This includes: each vehicle node retrieving the current local perception model from local storage and loading it into memory; sampling a small batch of data from the local dataset; generating several random perturbation directions; for each random perturbation direction, performing a forward function value lookup on the perturbation point and the current point to obtain the function value corresponding to the perturbation point and the function value corresponding to the current point, wherein the perturbation point is the sum of the current local perception model and the random perturbation direction; and calculating the zero-order descent direction estimate by weighting the sum of the difference between the function value corresponding to the perturbation point and the function value corresponding to the current point according to the random perturbation direction.

[0095] Specifically, the zero-order descent direction estimate is calculated by weighting and summing the difference between the function value corresponding to the disturbance point and the function value corresponding to the current point according to the random disturbance direction. This includes: for each vehicle node, multiplying the difference between the function value of the disturbance point and the function value of the current point corresponding to each sample by the ratio of the model parameter dimension to the smoothing parameter, and then multiplying by the random disturbance direction to obtain the directional component of the sample under the random disturbance direction; summing the directional components of all samples under the same random disturbance direction; summing the summation results of all random disturbance directions and dividing by the product of the mini-batch size and the number of random directions to obtain the zero-order descent direction estimate.

[0096] Specifically, each vehicle node estimates its local update direction based on local data. In practice, there are various gradient estimation methods (such as analytical gradient, function value lookup, random direction sampling, etc., among which zero-order difference estimation based on function value lookup can be further divided into forward difference, bidirectional difference, coordinate direction difference, simultaneous perturbation, and single-point estimation, etc.). In this embodiment, the forward difference method based on function value lookup is used as an example. The k-th node constructs the update direction estimate in the t-th round according to the following formula.

[0097] (1)

[0098] in, This is a mini-batch of data (size: ) sampled from the k-th vehicle node in round t. ), For random perturbation direction (total) (where D is the dimension of the model parameters, and μ is the smoothing parameter). Let ξ be the local loss function for the k-th vehicle node, and ξ be a single sample in the mini-batch. Through this zero-order query estimation, each vehicle node does not need to compute or upload analytical gradients during collaborative training.

[0099] In some other embodiments, the zero-order descent direction estimator can also be constructed using a two-way difference (i.e., central difference) approach: in addition to querying the function value at the perturbation point, the function value at the point in the opposite direction of the perturbation is also queried simultaneously. The zero-order descent direction estimator is constructed using the difference between the function value at the perturbation point and the function value at the point in the opposite direction, calculated as follows:

[0100] (2)

[0101] Compared to the forward difference form, the two-way difference form requires two function value queries for each random perturbation direction, and the function value query cost is twice that of the forward difference form. However, it can further reduce the difference truncation bias and improve the direction estimation accuracy. In practice, the forward difference and two-way difference forms can be selected according to the available function value query budget and accuracy requirements, without departing from the protection scope of this invention.

[0102] For example, the generation step of the random perturbation direction includes: each vehicle node generates the random perturbation direction in a Gaussian-spherical manner, that is, first generates a random vector with the same dimension as the local perception model, and then normalizes the random vector into a unit direction vector. Specifically, the k-th vehicle node generates a random vector for each of the random directions in the t-th round, and normalizes it into a unit perturbation direction, as shown in the following formula.

[0103] (3)

[0104] in, Let J be the j-th random vector of the same dimension as the local perception model generated by the k-th vehicle node in round t. Each of its components is independent and identically distributed, and all follow a standard normal distribution (Gaussian distribution) with a mean of 0 and a variance of 1. This represents the Euclidean norm. It is the unit direction vector obtained by normalizing a standard normal random vector whose components are independent and identically distributed, and which is uniformly distributed on a unit sphere in D-dimensional space. obey The random perturbation direction is uniformly distributed on the model matrix, where D is the dimension of the model parameters (i.e., the product of the number of rows and columns of the model matrix). Each vehicle node can independently generate the random perturbation direction based on a local random seed, or it can be generated by a shared random seed determined by the round number and the node number, so as to ensure reproducibility while taking into account randomness.

[0105] In some other embodiments, the zero-order estimation method of differential along the coordinate direction can also be adopted: after expanding the parameter matrix of the local sensing model into a D-dimensional parameter vector, query the function values ​​at the disturbance point and the opposite direction point one by one along each coordinate direction and perform central difference, and then synthesize the zero-order descent direction estimate by weighting according to each coordinate direction, and calculate it according to the following formula.

[0106] (4)

[0107] in, Let be the D-dimensional parameter vector obtained by expanding the parameter matrix of the local perception model of the k-th vehicle node in round t, i.e. =vec( ); vec(·) represents the operator that expands the matrix column-wise into a D-dimensional parameter vector, where D is the model parameter dimension; is the p-th standard basis direction in the D-dimensional space, that is, the unit direction vector where the p-th element is 1 and the rest are 0 (p=1,2,…,D); the meanings of the other symbols are the same as in equation (1). Compared with equation (2), this method performs central difference along each coordinate direction one by one, and requires about 2D function value queries for one estimation. The obtained estimation does not contain random direction sampling variance and the direction is determined. It is especially suitable for scenarios with low parameter dimension or scenarios where the dimension can be compressed first by combining matrix sparsity and low-rank structure before use.

[0108] In some other embodiments, the zero-order descent direction estimator can also be constructed by simultaneous perturbation: by performing a function value query only once for the perturbation point and the opposite direction point, all parameter dimensions can be excited simultaneously by a single random perturbation and the difference information of each coordinate direction can be obtained, and the following formula can be used for calculation.

[0109] (5)

[0110] Where ⊙ represents the element-wise multiplication between two matrices or vectors of the same dimension; is a random perturbation matrix of the same dimension as the local perception model, whose elements independently follow a Rademacher distribution, i.e., each element takes a probability of +1 or -1 with a probability of 1 / 2. Since the absolute value of each element is 1, its reciprocal is equal to itself. Therefore, the element-wise multiplication in the formula is equivalent to performing central difference estimation for each parameter direction, and weighting the difference results with the perturbation sign of each coordinate. This method only requires two function value queries for each random perturbation direction, and the query cost is comparable to that of formula (2), but it can simultaneously perturb all parameter dimensions. In practice, multiple perturbation matrices can be generated, estimated separately, and then averaged to reduce variance.

[0111] In some other embodiments, when the overhead of function value lookup is limited, a single-point zero-order estimation method can also be used: for each random perturbation direction, only the function value at the perturbation point is queried, and the function value at the current point is no longer queried, and the following formula is used for calculation.

[0112] (6)

[0113] The meanings of the symbols are consistent with those in equation (1); represents the unit perturbation direction generated by equation (3) and uniformly distributed on the D-dimensional unit sphere. Since it satisfies the zero-mean symmetry (its expectation is the zero vector), omitting the function value term at the current point will not change the expectation of the estimator, but will increase the variance of a single estimate. This method requires only one function value query for each random perturbation direction, and the query cost is only half that of equation (1). The variance can be compensated by appropriately increasing the number of random perturbation directions, making it suitable for application scenarios where the function value query budget is limited.

[0114] (iii) Momentum renewal.

[0115] In some other embodiments, each vehicle node updates its local momentum by weighting and accumulating the zero-order descent direction estimate with its historical local momentum, including: reading the local momentum of each vehicle node in the previous round; and weighting and accumulating the zero-order descent direction estimate with its historical local momentum according to the momentum coefficient to obtain the updated local momentum.

[0116] Each vehicle node updates its local momentum based on the zero-order descent direction estimate, calculated using the following formula.

[0117] (7)

[0118] Specifically, the momentum update steps include:

[0119] Read the local momentum of the k-th vehicle node in the previous round. .

[0120] The zeroth-order estimator is obtained according to formula (7). With historical momentum By performing weighted accumulation, the updated momentum is obtained. .

[0121] (iv) Muon orthogonalization.

[0122] In other embodiments, each vehicle node performs singular value decomposition on the updated local momentum, and takes the product of the left singular vector matrix and the right singular vector matrix as the orthogonalized update direction to make the update magnitude of each singular value direction more balanced. This includes: performing singular value decomposition on the updated local momentum to obtain a left singular vector matrix, a singular value matrix, and a right singular vector matrix; and taking the product of the transpose of the left singular vector matrix and the right singular vector matrix as the orthogonalized update direction to make the update magnitude of each singular value direction more balanced.

[0123] Specifically, each vehicle node performs singular value decomposition on the updated local momentum matrix and takes the product of its left and right singular vector matrices as the orthogonalization update direction.

[0124] (8a)

[0125] (8b)

[0126] in, , , Let be the left singular vector matrix, singular value matrix, and right singular vector matrix of the singular value decomposition of the momentum matrix of the k-th vehicle node, respectively, and ⊤ denotes transpose. This orthogonalization operation flattens the singular value spectrum of the update direction into an approximately semi-orthogonal matrix, making the update amplitude in each singular value direction more balanced, thereby alleviating gradient rank collapse and accelerating the convergence of the deep sensing network.

[0127] Specifically, the orthogonalization steps include...

[0128] For the updated local momentum matrix Perform singular value decomposition to obtain =U·S· ,in, It is a left singular vector matrix. It is a singular value matrix. Let be a right singular vector matrix, where the superscript k denotes the vector matrix of the k-th client.

[0129] Take the product of the left and right singular vector matrices. =U· As shown in Equations (3a) and (3b), the orthogonalization update direction makes the update magnitude more balanced in each singular value direction.

[0130] (v) Local model update.

[0131] In some other embodiments, each vehicle node updates its local perception model along the orthogonalization update direction to obtain an intermediate model, including: updating the local perception model along the orthogonalization update direction with the step size to obtain the intermediate model.

[0132] Each vehicle node updates its local perception model along the orthogonalization update direction.

[0133] (9)

[0134] Specifically, the local model update steps include: following formula (4) along the orthogonalization direction. Update the local model with a step size η to obtain the intermediate model. .

[0135] (vi) Model neighbor mixing.

[0136] Each vehicle node performs a weighted mixing of its local model and neighbor vehicle models according to the current time-varying inter-vehicle communication topology.

[0137] (10)

[0138] (vii) Momentum neighbor mixing.

[0139] In other embodiments, each vehicle node weightedly mixes its local intermediate model with the intermediate models of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain an updated local perception model. Each vehicle node also weightedly mixes its local momentum with the momentum of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain an updated local momentum. This includes: each vehicle node weightedly summing its own intermediate model with the intermediate models of neighboring vehicles according to the time-varying mixing matrix corresponding to the current vehicle network communication topology to obtain an updated local perception model; and each vehicle node weightedly summing its updated local momentum with the local momentum of neighboring vehicles according to the same time-varying mixing matrix to obtain a mixed local momentum. The time-varying mixing matrix is ​​a double random matrix, whose elements reflect the communication connectivity and weights between vehicle nodes in the current wheel.

[0140] Specifically, each vehicle node weights and mixes its local momentum with the momentum of its neighboring vehicles.

[0141] (11)

[0142] in, The time-varying double random mixture matrix corresponding to the vehicle-to-everything (V2X) communication topology in round t. The (k,j)th element is given, where K is the total number of vehicle nodes. The momentum of each vehicle node evolves independently, exchanging information only through neighbor mixing, without the need for global synchronization.

[0143] For example, in a real-world vehicle-to-everything (V2X) network, each vehicle node periodically broadcasts heartbeat messages to obtain neighbor relationships through vehicle-to-vehicle communication. The heartbeat message includes at least the vehicle identifier, location, speed, timestamp, and the number (degrees) of the vehicle's currently communicable neighbors. At the start of each round, each vehicle node records vehicles from which it can still receive heartbeat messages within the most recent time window as its current neighbors, forming a current neighbor set. If a neighbor vehicle's heartbeat message is not received within the specified time window, that neighbor vehicle is removed from the neighbor set, thus reflecting the establishment and disconnection of communication connections caused by vehicle movement.

[0144] For example, the time-varying mixing matrix does not need to be centrally generated by any central node, nor does it require any vehicle node to know its complete form; each vehicle node only needs to calculate its corresponding mixing weight row vector according to the following local rules, and the weight rows calculated by each vehicle node according to the local rules constitute the double random mixing matrix of the t-th round. The sum of the elements in each row and the sum of the elements in each column of the double random mixing matrix are both 1, and all elements are non-negative, that is, the following formula holds for any k and j.

[0145] (12a)

[0146] (12b)

[0147] in, Let K be the (k,j)th element of the double random mixing matrix in the t-th round, where K is the total number of vehicle nodes.

[0148] Specifically, set Let be the number of current neighbors of the k-th vehicle node in round t (i.e., the degree of the vehicle node). Let J be its current neighbor set. For any neighbor vehicle j in the current neighbor set, the vehicle node calculates its corresponding mixed weight according to the following Metropolis weighting rule.

[0149] (13)

[0150] The hybrid weight of the vehicle node itself is calculated using the following formula.

[0151] = (14)

[0152] Where max(·,·) represents taking the larger of the two values ​​within the parentheses. When the k-th vehicle node cannot communicate with any vehicle in round t, its current neighbor set is empty, and its own weight is 1, meaning that this round is updated only based on local information; for vehicle j that is neither itself nor belongs to its current neighbor set, the corresponding mixed weight is 0. The rows of vehicle weights calculated according to the above local rules together constitute the double random mixing matrix. The entire process does not require any central node to maintain the global mixing matrix, and each vehicle node does not need to know the complete form of the global mixing matrix. It only needs to know its own current neighbor set and the degree of each neighbor vehicle in it to complete the neighbor mixing for this round.

[0153] After completing their local update, each vehicle node broadcasts the intermediate model and updated momentum to vehicles in its current neighbor set. It then weights and sums the broadcast information according to the calculated mixing weights, completing the neighbor mixing of the model and momentum. Each round can be aligned via onboard satellite positioning and timing, ensuring that each vehicle node executes the above process according to the same round cycle.

[0154] Specifically, the model neighbor mixing step and the momentum neighbor mixing step include: each vehicle node according to the time-varying mixing matrix corresponding to the current vehicle network communication topology. , and its own intermediate model Intermediate model with neighboring vehicles After performing a weighted summation, the updated local model is obtained according to formula (10). Each vehicle node is based on the same time-varying hybrid matrix. , will update its momentum Momentum of neighboring vehicles By performing a weighted summation, the mixed local momentum is obtained according to formula (11). .

[0155] Please see Figure 3 The time-varying hybrid matrix The communication topology of the vehicle network changes dynamically. Due to the high speed of vehicles, the communication links between vehicles are constantly being established and disconnected at different times. Therefore, the communication topology and the hybrid matrix corresponding to different rounds change accordingly. This invention uses the time-varying double random hybrid matrix to ensure that training can still converge under dynamic topology changes.

[0156] Please see Figure 3This illustrates the changing relationship between vehicles in a vehicle-to-everything (V2X) network over time: (a) represents the connection at time t, (b) represents the connection at time t+1, and (c) represents the connection at time t+2. As vehicles move, some vehicles enter or leave the communication range, and the communication connection between vehicles is correspondingly broken or established. The neighbor set and calculated hybrid weights obtained by each vehicle node in each round change accordingly. Based on the dynamic neighbor relationship, this invention performs weighted mixing of the model and momentum of each vehicle node according to formulas (5) and (6), and theoretically ensures that training can still converge under dynamic changes in the connection relationship.

[0157] Optionally, to ensure consistency in the rounds of each vehicle node in actual deployment, each vehicle node can align the round boundaries based on the timing signal of the vehicle-mounted satellite positioning system (GNSS) and perform a round of training at a fixed period (e.g., every 200 milliseconds to 1 second). The heartbeat message sending period is consistent with the round period, and the time window of the neighbor table is taken as several times the round period (e.g., 2 to 5 times) to suppress frequent neighbor jitter.

[0158] Optionally, in addition to reaching the preset number of iterations T, the iteration termination condition can also be set as follows: the average change between the local models of each vehicle node is less than a preset threshold for several consecutive iterations, or the change norm of the local model of each vehicle node compared with the previous iteration is less than a preset threshold. At this time, it is determined that each vehicle node has converged collaboratively, and a unified target perception model is output.

[0159] Repeat steps (ii) to (vii) above for T iterations until the local models of each vehicle node converge to the highest quality target perception model. The training objective is then complete, and the resulting model can be used for target detection, recognition, and tracking tasks in vehicle-to-everything (V2X) collaborative perception.

[0160] Based on the same inventive concept, please see Figure 4 The present invention also provides a decentralized collaborative sensing system based on zero-order query and local momentum, comprising:

[0161] The initialization module is used to initialize the local perception model and local momentum of each vehicle node in the vehicle network, and set the training parameters, including step size, momentum coefficient, smoothing parameter, mini-batch size, number of random directions, number of iterations, and time-varying hybrid matrix sequence of the vehicle network.

[0162] The zero-order query estimation module is used by each vehicle node to apply a random disturbance direction to the local perception model based on local data and perform function value queries. The local descent direction is estimated based on the difference between the function value of the disturbance point and the current point, thus obtaining the zero-order descent direction estimate.

[0163] The momentum update module is used by each vehicle node to perform a weighted accumulation of the zero-order descent direction estimate and the historical local momentum, and update the local momentum.

[0164] The orthogonalization module is used by each vehicle node to perform singular value decomposition on the updated local momentum. The product of the left singular vector matrix and the right singular vector matrix is ​​taken as the orthogonalization update direction to make the update amplitude of each singular value direction more balanced.

[0165] The local model update module is used to update the local perception model of each vehicle node along the orthogonal update direction to obtain an intermediate model.

[0166] The model fusion module is used by each vehicle node to weight and mix the local intermediate model with the intermediate models of neighboring vehicles according to the current time-varying inter-vehicle communication topology, so as to obtain the updated local perception model.

[0167] The momentum mixing module is used to weight and mix the local momentum of each vehicle node with the momentum of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain the updated local momentum.

[0168] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the decentralized collaborative sensing method based on zero-order query and local momentum as described above.

[0169] The program product of the present invention for implementing the above-described method may employ a portable compact disk read-only memory and include program code, and may run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto. In the present invention, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0170] It should be noted that a computer-readable storage medium may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0171] Unless otherwise defined, the technical or scientific terms used in this invention shall have the general meaning understood by one of ordinary skill in the art to which this invention pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this invention do not indicate quantitative limitations and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this invention are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “A plurality” used in this invention refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. In this invention, the terms "first," "second," "third," etc., are used only to distinguish similar objects and do not represent a specific order of objects.

[0172] Obviously, the accompanying drawings are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, certain design, manufacturing, or production modifications made according to the technical content disclosed in the present invention are merely conventional technical means and should not be considered as insufficient disclosure of the present invention.

[0173] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A decentralized collaborative sensing method based on zero-order query and local momentum, characterized in that, Includes the following steps: S1. Initialize the local perception model and local momentum of each vehicle node in the vehicle network, and set the training parameters, including step size, momentum coefficient, smoothing parameter, mini-batch size, number of random directions, number of iterations, and time-varying hybrid matrix sequence of the vehicle network. S2. Each vehicle node applies a random disturbance direction to the local perception model based on local data and queries the function value. It estimates the local descent direction based on the difference between the function value of the disturbance point and the current point, and obtains the zero-order descent direction estimate. S3. Each vehicle node performs a weighted accumulation of the zero-order descent direction estimate and the historical local momentum, and updates the local momentum. S4. Each vehicle node performs singular value decomposition on the updated local momentum, and takes the product of the left singular vector matrix and the right singular vector matrix as the orthogonal update direction to make the update amplitude of each singular value direction balanced. S5. Each vehicle node updates its local perception model along the orthogonalization update direction to obtain an intermediate model; S6. Each vehicle node performs a weighted mixing of the local intermediate model and the neighbor vehicle intermediate model according to the current time-varying inter-vehicle communication topology to obtain the updated local perception model. S7. Each vehicle node weights and mixes its local momentum with that of its neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain the updated local momentum. Repeat steps S2 to S7, iterating until all vehicle nodes converge collaboratively, to obtain a unified target perception model.

2. The method according to claim 1, characterized in that, Initialize the local perception model and local momentum of each vehicle node in the vehicle-to-everything (V2X) network, and set the training parameters, including: Set the initial perception model for each vehicle node; The first zero-order descent direction estimate is calculated using the zero-order query estimation method and used as the initial value of the local momentum for each vehicle node. Set the step size, and the step size is greater than zero; The momentum coefficient is set, and the momentum coefficient takes a value between zero and one, to weigh the proportion between historical momentum and the current zero-order descent direction estimate; The smoothing parameter is set to be greater than zero, and is used to control the magnitude of the perturbation direction in the zero-order estimation. Set the small batch size and the number of random directions; Set the number of iteration rounds; A time-varying dual random hybrid matrix is ​​determined for each round of the vehicle network, wherein the elements of the time-varying hybrid matrix reflect the communication connectivity and weights between vehicle nodes in the current round.

3. The method according to claim 1, characterized in that, Each vehicle node applies a random perturbation direction to its local perception model based on local data and queries the function value. It then estimates the local descent direction based on the difference between the function value at the perturbation point and the current point, obtaining a zero-order descent direction estimate, including: Each vehicle node retrieves its current local perception model from local storage and loads it into memory; Sample small batches of data from the local dataset; Generate several random perturbation directions; For each random perturbation direction, the forward function value is queried for the perturbation point and the current point respectively to obtain the function value corresponding to the perturbation point and the function value corresponding to the current point, wherein the perturbation point is the sum of the current local perception model and the random perturbation direction; The zero-order descent direction estimate is calculated by weighting and summing the differences between the function values ​​corresponding to the perturbation point and the function values ​​corresponding to the current point, according to the random perturbation direction.

4. The method according to claim 3, characterized in that, The step of calculating the zero-order descent direction estimate by weighted summation according to the random perturbation direction based on the difference between the function value corresponding to the perturbation point and the function value corresponding to the current point includes: For each vehicle node, the difference between the disturbance point function value and the current point function value corresponding to each sample is multiplied by the ratio of the model parameter dimension to the smoothing parameter, and then multiplied by the random disturbance direction to obtain the directional component of the sample under the random disturbance direction. Sum the directional components of all samples under the same random perturbation direction; The summation of all random perturbation directions is divided by the product of the mini-batch size and the number of random directions to obtain the zero-order descent direction estimate.

5. The method according to claim 3, characterized in that, Each vehicle node updates its local momentum by weighting and accumulating the zero-order descent direction estimate with its historical local momentum, including: Read the local momentum of each vehicle node in the previous round; The zero-order descent direction estimate is weighted and accumulated with the historical local momentum according to the momentum coefficient to obtain the updated local momentum.

6. The method according to claim 5, characterized in that, Each vehicle node performs singular value decomposition on the updated local momentum, taking the product of the left and right singular vector matrices as the orthogonalized update direction to ensure balanced update magnitudes across singular value directions, including: Singular value decomposition is performed on the updated local momentum to obtain the left singular vector matrix, the singular value matrix, and the right singular vector matrix; The product of the left singular vector matrix and the transpose of the right singular vector matrix is ​​taken as the orthogonalization update direction to balance the update magnitude in each singular value direction.

7. The method according to claim 6, characterized in that, Each vehicle node updates its local perception model along the orthogonalization update direction to obtain an intermediate model, including: updating the local perception model along the orthogonalization update direction with the step size to obtain the intermediate model.

8. The method according to claim 7, characterized in that, Each vehicle node, based on the current time-varying inter-vehicle communication topology, weights and mixes its local intermediate model with the intermediate models of neighboring vehicles to obtain an updated local perception model. Each vehicle node, based on the current time-varying inter-vehicle communication topology, weights and mixes its local momentum with the momentum of neighboring vehicles to obtain an updated local momentum, including: Each vehicle node calculates a weighted sum of its own intermediate model and the intermediate models of its neighboring vehicles based on the time-varying hybrid matrix corresponding to the current vehicle network communication topology, thus obtaining an updated local perception model. Each vehicle node, based on the same time-varying mixing matrix, performs a weighted summation of its own updated local momentum with the local momentum of its neighboring vehicles to obtain the mixed local momentum. The time-varying hybrid matrix is ​​a double random matrix, whose elements reflect the communication connectivity and weights between vehicle nodes in the current round.

9. The method according to claim 1, characterized in that, The target perception model is a convolutional neural network target detection model; the method is applied to the collaborative training of road target detection models for autonomous vehicles, where each vehicle node holds a locally collected road image dataset, and the data distribution of each vehicle node is non-independent and identically distributed. Repeat steps S2 to S7, iterating for a preset number of rounds until the local perception models of each vehicle node converge collaboratively to obtain the target perception model, and then use the target perception model for target detection, recognition and tracking tasks in vehicle network collaborative perception.

10. A decentralized collaborative sensing system based on zero-order query and local momentum, characterized in that, include: The initialization module is used to initialize the local perception model and local momentum of each vehicle node in the vehicle network, and set the training parameters, including step size, momentum coefficient, smoothing parameter, mini-batch size, number of random directions, number of iterations, and time-varying hybrid matrix sequence of the vehicle network. The zero-order query estimation module is used by each vehicle node to apply a random disturbance direction to the local perception model based on local data and perform function value query. The local descent direction is estimated based on the difference between the function value of the disturbance point and the current point, and the zero-order descent direction estimate is obtained. The momentum update module is used by each vehicle node to perform a weighted accumulation of the zero-order descent direction estimate and the historical local momentum, and update the local momentum. The orthogonalization module is used by each vehicle node to perform singular value decomposition on the updated local momentum. The product of the left singular vector matrix and the right singular vector matrix is ​​taken as the orthogonalization update direction to make the update amplitude of each singular value direction more balanced. The local model update module is used to update the local perception model of each vehicle node along the orthogonal update direction to obtain an intermediate model. The model fusion module is used by each vehicle node to weight and blend the local intermediate model with the intermediate models of neighboring vehicles according to the current time-varying inter-vehicle communication topology, so as to obtain the updated local perception model. The momentum mixing module is used to weight and mix the local momentum of each vehicle node with the momentum of neighboring vehicles according to the current time-varying inter-vehicle communication topology to obtain the updated local momentum. The central control module is used to repeatedly trigger the zero-order query estimation module, momentum update module, orthogonalization module, local model update module, model mixing module, and momentum mixing module, iterating until all vehicle nodes converge collaboratively to obtain a unified target perception model.

Images