Multi-vehicle cooperation method and system based on task driving

Abstract

The invention discloses a multi-vehicle cooperation method and system based on task driving, and belongs to the technical field of vehicle communication, and the method comprises the steps that a main vehicle initiates a cooperative perception request to candidate vehicles in a communication range; when the candidate vehicles agree to participate, collaborative assistance information is fed back; the host vehicle schedules the cooperative vehicle based on the information and constructs a sensing topology; the main vehicle adjusts parameters of the intelligent sensing model, moves a feature fusion position to the last layer of the model, and makes a calculation load balancing decision of model segmentation based on auxiliary information; the host vehicle jointly optimizes the bandwidth of the scheduling system and computing resources of the host vehicle; the host vehicle distributes the decision to the cooperative vehicle; and the master and cooperative vehicles execute tasks according to decisions, and the master vehicle side completes feature fusion to obtain a final perception reasoning result. The method is adaptive to a multi-vehicle cooperative task structure, sensing precision and real-time performance are guaranteed, the resource utilization rate is increased, and automatic driving efficient environment sensing is assisted.

Description

Technical Field

[0001] This invention belongs to the field of vehicle communication technology, specifically relating to a task-driven multi-vehicle collaborative method and system. Background Technology

[0002] With the development of intelligent transportation systems, autonomous driving technology has shown significant potential in areas such as traffic safety and traffic congestion management. Environmental perception, as a key basis for autonomous driving planning and decision-making, is crucial for safe autonomous driving due to its accuracy and real-time performance. However, the perception accuracy of a single vehicle is insufficient due to limitations in its field of view and sensor resolution. As a result, multi-vehicle collaborative perception technology has emerged to improve overall perception accuracy by expanding the perception range of a single vehicle.

[0003] However, existing technologies have the following drawbacks: First, existing model segmentation technologies are mostly designed for accelerating inference in single-data-source tasks, and are not adapted to the structural characteristics of distributed feature extraction and transmission and centralized feature fusion and prediction in multi-vehicle collaborative perception. Some technologies only optimize communication efficiency without coordinating the computing load, or fail to consider the heterogeneity of multi-vehicle computing resources, resulting in difficulties in real-time data transmission processing and low collaborative efficiency under computationally intensive tasks. Second, in multi-vehicle collaboration, vehicle topology and computing resources change dynamically, and the number of collaborative vehicles needs to balance perception accuracy and latency. However, existing technologies lack task-driven inference acceleration schemes and scheduling strategies, and cannot jointly schedule collaborative vehicles, task load, and computing resources. This can easily lead to wasted resources due to waiting latency, making it difficult to balance perception accuracy and collaborative latency, thus hindering the implementation of multi-vehicle collaborative perception technology. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a task-driven multi-vehicle cooperative method and system to solve or partially solve the problems mentioned in the background art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a task-driven multi-vehicle collaborative method, comprising the following steps: S1. The master vehicle initiates a cooperative perception request to at least one candidate vehicle within the communication range. S2. After receiving the collaborative perception request, the candidate vehicle determines whether to participate in collaborative perception. If it agrees to participate, it sends collaborative assistance information to the master vehicle. S3. Based on the received collaborative assistance information, the master vehicle performs collaborative vehicle scheduling operations and constructs a collaborative perception topology to determine the number of vehicles participating in collaborative perception and the specific vehicles. S4. The main vehicle, taking into account the structural characteristics of the multi-vehicle collaborative perception task, adjusts the parameters of the intelligent perception model, moves the feature fusion position of the intelligent perception model to the last layer of the model, and makes a computational load balancing decision based on model segmentation based on the collaborative auxiliary information. S5. Based on the scheduling results of the cooperating vehicles and the computing load balancing decision, the master vehicle performs a joint optimization scheduling decision on the system bandwidth resources and the computing resources on the master vehicle side, allocates bandwidth resources to the cooperating vehicles, and allocates computing resources to the model inference load corresponding to the cooperating vehicles on the master vehicle side. S6. The master vehicle distributes the computational load balancing decision and the joint optimization scheduling decision to the selected cooperative vehicles. S7. The master vehicle and the cooperating vehicle jointly execute the collaborative perception task according to the computational load balancing decision and the joint optimization scheduling decision, and complete the feature fusion operation on the master vehicle side to obtain the final collaborative perception reasoning result.

[0006] As a preferred option for a task-driven multi-vehicle cooperative method, in step S2, the cooperative auxiliary information specifically includes vehicle position coordinates, vehicle speed, channel environment status, computing power, and signal transmission power.

[0007] As a preferred solution for a task-driven multi-vehicle cooperative method, in step S4, the perception inference part of the intelligent perception model consists of multiple convolutional layers and batch normalization layers; wherein, the calculation expression for the area covered by the convolutional kernel in the convolutional layer is:

[0008] In the formula, The feature map corresponding to the input data channel. It is the convolution kernel corresponding to channel c. Represents the coordinates of the starting point of the convolution. Represents the element coordinates of the convolution kernel. For bias terms; The calculation expression for the batch normalization layer is:

[0009] In the formula, , The mean and variance are calculated from the training data, respectively. , These are trainable parameters. It is used to avoid the minimum value where the denominator is zero.

[0010] As a preferred solution for task-driven multi-vehicle collaboration, when adjusting the parameters of the PIXOR model to achieve consistent inference accuracy, for input data that satisfy the condition of additive separability, the bias terms of the convolutional layer are adjusted. and parameters of the batch normalization layer and This enables the incremental splitting of computation between convolutional layers and batch normalized layers.

[0011] As a preferred solution for task-driven multi-vehicle collaboration, in step S4, the load balancing decision based on model segmentation is implemented using a reinforcement learning algorithm based on graph neural networks, and the reward function expression is:

[0012] In the formula, To meet the maximum collaborative sensing latency constraint, For actual collaborative sensing latency. To improve the accuracy of collaborative sensing. , The parameter is non-negative, and the algorithm optimization objective is to minimize the collaborative sensing latency and maximize the collaborative sensing accuracy.

[0013] As a preferred solution for a task-driven multi-vehicle collaboration method, the topology of the graph neural network input is a star graph, with the master vehicle as the central node and the candidate vehicles as sub-nodes; The characteristic parameters of the central node are set as the computing power of the master vehicle, and the characteristic parameters of the sub-nodes are set as the computing power, signal transmission power and cooperative relationship with the master vehicle of the candidate vehicle. The cooperative relationship indicates whether the candidate vehicle was selected as a cooperative vehicle by the master vehicle before the current time. The characteristic parameters of the edge between the central node and the sub-nodes are set as the channel state between the master vehicle and the candidate vehicle.

[0014] As a preferred solution for a task-driven multi-vehicle cooperative method, the output of the graph neural network includes the cooperative vehicle number decision output by the master vehicle node, and the cooperative selection probability and model segmentation decision output by the candidate vehicle nodes; the neural network parameters of the graph nodes are iteratively updated through a reinforcement learning algorithm until the reward function is obtained. The convergence process yields the final load balancing decision.

[0015] As a preferred option for a task-driven multi-vehicle collaboration method, in step S5, the joint optimization scheduling of system bandwidth resources and main vehicle-side computing resources is achieved through a convex optimization algorithm. The optimization objective of the convex optimization algorithm is to reduce the waiting latency of collaborative vehicle task processing in order to minimize the overall collaborative perception latency.

[0016] As a preferred option for a task-driven multi-vehicle collaborative method, in step S7, the feature fusion adopts an additive fusion method, and the fusion formula is:

[0017] In the formula, to These are the processing results of the perception models corresponding to each cooperative vehicle. The main vehicle obtains the final cooperative perception inference result through this fusion formula.

[0018] Secondly, the present invention provides a task-driven multi-vehicle cooperative system, employing the aforementioned task-driven multi-vehicle cooperative method, including: The collaborative request initiation module is used by the master vehicle to initiate a collaborative perception request to at least one candidate vehicle within the communication range. The collaborative request response module is used to determine whether to participate in collaborative perception after the candidate vehicle receives the collaborative perception request. If it agrees to participate, it sends collaborative assistance information to the master vehicle. The collaborative scheduling and topology construction module is used by the master vehicle to perform collaborative vehicle scheduling operations and construct a collaborative perception topology based on the received collaborative assistance information, and to determine the number of vehicles and specific vehicles participating in collaborative perception. The model adjustment and load balancing decision module is used to adjust the parameters of the intelligent perception model based on the structural characteristics of the multi-vehicle collaborative perception task, move the feature fusion position of the intelligent perception model to the last layer of the model, and make a calculation load balancing decision based on the model segmentation based on the collaborative auxiliary information. The resource joint scheduling module is used by the master vehicle to perform joint optimization scheduling decisions on system bandwidth resources and master vehicle-side computing resources based on the scheduling results of the cooperating vehicles and the computing load balancing decision, allocate bandwidth resources to cooperating vehicles, and allocate computing resources to the model inference load corresponding to the cooperating vehicles on the master vehicle side. The decision distribution module is used by the master vehicle to distribute the computational load balancing decision and the joint optimization scheduling decision to the selected cooperative vehicles; The collaborative perception and reasoning module is used by the master vehicle and the collaborative vehicles to jointly execute the collaborative perception task according to the computational load balancing decision and the joint optimization scheduling decision, and to complete the feature fusion operation on the master vehicle side to obtain the final collaborative perception and reasoning result.

[0019] As a preferred solution for a task-driven multi-vehicle cooperative system, the cooperative request and response module includes cooperative auxiliary information specifically including vehicle position coordinates, vehicle speed, channel environment status, computing power, and signal transmission power.

[0020] As a preferred solution for a task-driven multi-vehicle cooperative system, the model adjustment and load balancing decision module comprises a perception inference part of the intelligent perception model consisting of multiple convolutional layers and batch normalization layers; wherein, the calculation expression for the region covered by the convolutional kernel in the convolutional layer is:

[0021] In the formula, The feature map corresponding to the input data channel. It is the convolution kernel corresponding to channel c. Represents the coordinates of the starting point of the convolution. Represents the element coordinates of the convolution kernel. For bias terms; The calculation expression for the batch normalization layer is:

[0022] In the formula, , The mean and variance are calculated from the training data, respectively. , These are trainable parameters. It is used to avoid the minimum value where the denominator is zero.

[0023] As a preferred solution for a task-driven multi-vehicle cooperative system, in the model adjustment and load balancing decision module, when adjusting the PIXOR model parameters to achieve consistent inference accuracy, for input data that satisfy the condition of additive separability, the bias terms of the convolutional layer are adjusted. and parameters of the batch normalization layer and This enables the incremental splitting of computation between convolutional layers and batch normalized layers.

[0024] As a preferred solution for a task-driven multi-vehicle cooperative system, the model adjustment and load balancing decision module implements load balancing decision based on model segmentation through a reinforcement learning algorithm based on a graph neural network. The reward function expression is as follows:

[0025] In the formula, To meet the maximum collaborative sensing latency constraint, For actual collaborative sensing latency. To improve the accuracy of collaborative sensing. , The parameter is non-negative, and the algorithm optimization objective is to minimize the collaborative sensing latency and maximize the collaborative sensing accuracy.

[0026] As a preferred solution for a task-driven multi-vehicle collaborative system, the model adjustment and load balancing decision module constructs a star-shaped topology for the graph neural network input, with the master vehicle as the central node and the candidate vehicles as sub-nodes. The characteristic parameters of the central node are set as the computing power of the master vehicle, and the characteristic parameters of the sub-nodes are set as the computing power, signal transmission power and cooperative relationship with the master vehicle of the candidate vehicle. The cooperative relationship indicates whether the candidate vehicle was selected as a cooperative vehicle by the master vehicle before the current time. The characteristic parameters of the edge between the central node and the sub-nodes are set as the channel state between the master vehicle and the candidate vehicle.

[0027] As a preferred solution for a task-driven multi-vehicle cooperative system, the output of the graph neural network in the model adjustment and load balancing decision module includes the cooperative vehicle number decision output by the master vehicle node, and the cooperative selection probability and model segmentation decision output by the candidate vehicle nodes; the neural network parameters of the graph nodes are iteratively updated through a reinforcement learning algorithm until the reward function... The convergence process yields the final load balancing decision.

[0028] As a preferred solution for a task-driven multi-vehicle collaborative system, the joint scheduling of system bandwidth resources and main vehicle-side computing resources in the resource joint scheduling module is achieved through a convex optimization algorithm. The optimization objective of the convex optimization algorithm is to reduce the waiting latency of collaborative vehicle task processing in order to minimize the overall collaborative perception latency.

[0029] As a preferred solution for a task-driven multi-vehicle cooperative system, the feature fusion in the cooperative perception and reasoning module adopts an additive fusion method, and the fusion formula is as follows:

[0030] In the formula, to These are the processing results of the perception models corresponding to each cooperative vehicle. The main vehicle obtains the final cooperative perception inference result through this fusion formula.

[0031] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the task-driven multi-vehicle cooperative method of the first aspect or any possible implementation thereof.

[0032] Fourthly, the present invention provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform steps in the task-driven multi-vehicle cooperative method of the first aspect or any possible implementation thereof.

[0033] The beneficial effects of the technical solution provided by this invention are as follows: First, this invention addresses the structural characteristics of distributed feature extraction and transmission and centralized feature fusion and prediction in multi-vehicle collaborative perception. By adjusting the parameters of the intelligent perception model, the feature fusion position is moved to the last layer of the model. While ensuring consistent inference accuracy, it adapts to the special architecture of multi-vehicle collaborative tasks, solving the problem that existing model segmentation techniques are only applicable to single data sources and cannot match the structure of multi-vehicle collaborative tasks, resulting in low efficiency. This significantly improves the efficiency of collaborative perception.

[0034] Second, the model segmentation decision is made based on the real-time status information of the collaborative vehicles, clarifying the computational load allocation between the master vehicle and the collaborative vehicles. At the same time, the collaborative vehicle scheduling is optimized by using a graph neural network reinforcement learning algorithm to avoid excessive load on a single vehicle or idle resources. This solves the problems of the existing technology not coordinating the computational load and low utilization of general computing resources, and improves the efficiency of computing resource utilization.

[0035] Third, by using convex optimization algorithms to jointly schedule system bandwidth resources and computing resources on the main vehicle side, appropriate bandwidth is allocated to cooperative vehicles to ensure data transmission efficiency, reasonable computing power is allocated to the inference load on the main vehicle side to reduce processing waiting latency, and the scheduling decision is optimized by combining the reward function, which effectively reduces the overall latency of multi-vehicle cooperative perception and meets the real-time perception requirements of autonomous driving.

[0036] Fourth, considering the dynamic changes in vehicle topology and available computing resources in multi-vehicle collaboration, the system dynamically adjusts collaborative vehicle scheduling, load allocation, and resource scheduling strategies by receiving real-time auxiliary information from collaborative vehicles online. This solves the problem of poor adaptability caused by the lack of consideration for dynamic environments in existing technologies, making the solution more suitable for the actual scenarios of autonomous vehicle movement and promoting the application of multi-vehicle collaborative perception technology.

[0037] Fifth, in the process of collaborative vehicle scheduling and load decision-making, based on the premise of satisfying perception accuracy, the feature fusion position is shifted to meet the consistency of inference accuracy by adjusting the model parameters. At the same time, communication and computing costs are reduced by resource joint scheduling and load balancing. This solves the problem that existing technologies cannot balance perception accuracy and collaborative latency / cost, and achieves the optimal balance between accuracy and cost. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in this invention or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the task-driven multi-vehicle collaborative method provided in an embodiment of the present invention; Figure 2 A schematic diagram of load balancing based on feature fusion and location movement provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a task-driven multi-vehicle collaborative system architecture provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0041] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this invention should have the ordinary meaning understood by those skilled in the art to which this invention pertains. The terms "comprising" or "including," or similar words used in the embodiments of this invention, mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0042] With the rapid development of intelligent transportation systems, autonomous driving technology has shown great potential in solving problems such as traffic safety and road congestion. Environmental perception, as a key basis for planning and decision-making in autonomous vehicles, is crucial for safe and reliable autonomous driving due to its accuracy and real-time performance. However, due to limitations in the field of vision and sensor resolution, the perception accuracy of a single vehicle is significantly insufficient in scenarios such as blind spots and complex road conditions, making it difficult to meet the requirements of high-precision autonomous driving.

[0043] To overcome the limitations of single-vehicle perception, multi-vehicle collaborative perception technology has emerged. This technology expands the perception range of a single vehicle by integrating the perception capabilities of multiple autonomous vehicles, potentially improving the overall accuracy of vehicle environmental perception and becoming an important direction for solving the bottleneck of single-vehicle perception. However, in the practice of multi-vehicle collaborative perception, existing technologies still have several shortcomings, making it difficult to fully leverage the advantages of collaboration: On the one hand, existing model segmentation techniques mostly focus on accelerating inference for intelligent tasks with single data sources, without designing adaptation solutions for the unique structural characteristics of multi-vehicle collaborative perception tasks, such as distributed feature extraction and transmission, and centralized feature fusion and prediction. This results in limited improvement in the efficiency of multi-vehicle collaborative perception. For example, some existing technologies only consider communication efficiency optimization without taking into account the overall design of computational load allocation. Other technologies, while segmenting multi-path models based on the structural characteristics of AI models, fail to consider the heterogeneity of computing and communication resources in multi-vehicle collaborative scenarios. When facing computationally intensive intelligent perception tasks, the real-time processing difficulty of data transmission between a large number of collaborative vehicles increases significantly, seriously affecting the real-time performance of multi-vehicle collaborative perception.

[0044] On the other hand, in a multi-vehicle collaborative environment, the vehicle topology changes dynamically with the movement of autonomous vehicles, and available computing resources also fluctuate in real time. The selection of the number of collaborative vehicles needs to strike a reasonable balance between perception accuracy and collaborative latency. How to efficiently schedule collaborative vehicles based on the dynamic environment has become an urgent challenge. At the same time, existing technologies generally lack task-driven multi-vehicle collaborative task inference acceleration solutions and corresponding scheduling strategies, making it impossible to achieve joint and efficient scheduling of collaborative vehicles, task load, and general computing resources. This easily leads to resource waste due to waiting latency between vehicles, making it difficult to reduce collaborative perception latency while ensuring perception accuracy, thus hindering the practical implementation and application of multi-vehicle collaborative perception technology.

[0045] In summary, the current field of multi-vehicle collaborative perception urgently needs a technical solution that can adapt to the characteristics of the task structure and comprehensively optimize collaborative scheduling and resource allocation to solve the problems of low efficiency in existing technologies; difficulty in efficiently scheduling vehicles in dynamic multi-vehicle scenarios; difficulty in balancing perception accuracy and latency; and low resource utilization due to the lack of joint optimization of computing load and general computing resources. The following are the specific contents of the embodiments of the present invention.

[0046] See Figure 1 This invention provides a task-driven multi-vehicle collaboration method, comprising the following steps: S1. The master vehicle initiates a cooperative perception request to at least one candidate vehicle within the communication range. S2. After receiving the collaborative perception request, the candidate vehicle determines whether to participate in collaborative perception. If it agrees to participate, it sends collaborative assistance information to the master vehicle. S3. Based on the received collaborative assistance information, the master vehicle performs collaborative vehicle scheduling operations and constructs a collaborative perception topology to determine the number of vehicles participating in collaborative perception and the specific vehicles. S4. The main vehicle, taking into account the structural characteristics of the multi-vehicle collaborative perception task, adjusts the parameters of the intelligent perception model, moves the feature fusion position of the intelligent perception model to the last layer of the model, and makes a computational load balancing decision based on model segmentation based on the collaborative auxiliary information. S5. Based on the scheduling results of the cooperating vehicles and the computing load balancing decision, the master vehicle performs a joint optimization scheduling decision on the system bandwidth resources and the computing resources on the master vehicle side, allocates bandwidth resources to the cooperating vehicles, and allocates computing resources to the model inference load corresponding to the cooperating vehicles on the master vehicle side. S6. The master vehicle distributes the computational load balancing decision and the joint optimization scheduling decision to the selected cooperative vehicles. S7. The master vehicle and the cooperating vehicle jointly execute the collaborative perception task according to the computational load balancing decision and the joint optimization scheduling decision, and complete the feature fusion operation on the master vehicle side to obtain the final collaborative perception reasoning result.

[0047] The process is structured as follows: S1 serves as the starting point, where the master vehicle scans candidate vehicles within its communication range using the vehicle-to-everything (V2X) communication protocol. When initiating a collaboration request, it carries basic task information, such as the perception target and task priority, ensuring that candidate vehicles clearly understand the collaboration direction. S2 is the two-way interaction stage, where candidate vehicles must first assess their own status, such as whether they currently have high-priority tasks and whether their remaining computing power is sufficient, before deciding whether to respond. The collaboration assistance information fed back during the response provides data support for subsequent decisions, preventing the master vehicle from blindly scheduling tasks. S3 addresses the question of which vehicle to collaborate with, S4 addresses how the model adapts to the collaboration scenario, and S5 addresses how resources are allocated to ensure efficiency. S6 is the decision implementation stage, where standardized decision instruction formats ensure that each collaborating vehicle accurately understands its own task and resource quota. S7 is the result generation stage, where the master vehicle centrally integrates the inference results from each collaborating vehicle, ultimately forming a complete and reliable perception inference result.

[0048] In this embodiment, in step S2, the cooperative assistance information specifically includes vehicle position coordinates, vehicle speed, channel environment status, computing power, and signal transmission power.

[0049] Specifically, the vehicle position coordinates are obtained using high-precision coordinates from the Global Navigation Satellite System. The master vehicle can use these coordinates to calculate the relative distance between itself and candidate vehicles, determining whether they are within a stable communication range and avoiding mission interruptions due to vehicles moving out of the communication range during subsequent coordination. Vehicle speed information can be used to predict the position changes of candidate vehicles in the near future, providing a basis for the master vehicle to dynamically adjust the coordination topology. For example, if a candidate vehicle is moving too fast and is far away from the master vehicle, the master vehicle can exclude it from being a coordination vehicle in advance, reducing ineffective decisions. The channel environment status is evaluated by detecting indicators such as signal-to-noise ratio (SNR) and bit error rate (BER). Candidate vehicles with good channel status have high data transmission stability and low latency. The signal transmission power is a crucial reference for allocating bandwidth resources to the master vehicle, preventing the assignment of critical data transmission tasks to vehicles with poor channel performance, which could lead to data loss or delay. Computational power is quantified by detecting indicators such as CPU / GPU computing power and memory utilization, directly determining whether candidate vehicles can handle the computational load after model segmentation. For example, if a vehicle with insufficient computing power undertakes complex computational tasks, it will lead to a significant increase in processing latency, affecting overall collaborative efficiency. Signal transmission power determines the coverage and anti-interference capability of candidate vehicle data transmission. Too low a transmission power may result in unstable data transmission to the master vehicle, while too high a power may cause channel interference. The master vehicle can adjust the communication parameters with candidate vehicles based on the signal transmission power to balance transmission stability and channel resource usage.

[0050] In this embodiment, in step S4, the intelligent perception model is a PIXOR model, and the perception inference part of the PIXOR model consists of multiple convolutional layers and batch normalization layers; wherein, the calculation expression for the area covered by the convolutional kernel in the convolutional layer is:

[0051] In the formula, The feature map corresponding to the input data channel. It is the convolution kernel corresponding to channel c. Represents the coordinates of the starting point of the convolution. Represents the element coordinates of the convolution kernel. This is a bias term.

[0052] Specifically, the calculation of the region covered by a single convolutional kernel in a convolutional layer is an affine transformation, for the input data... The calculation of the region covered by a single convolution kernel can be expressed as:

[0053] in, Therefore, it can be seen that the calculation of the coverage area of ​​a single convolution kernel is achieved by adjusting the bias term. Adjustments can be made to achieve additively separable computation. Furthermore, since the computation of a convolutional layer is achieved through the sliding of the convolutional kernel, the computation result of the convolutional layer is obtained by splicing the computation results of the regions covered by each convolutional kernel. Therefore, based on the above analysis of the computation of a single convolutional kernel's covered region, it can be shown that the convolutional layer can also perform split computation for additively separable inputs.

[0054] Among them, input data Break it down by channel dimension Each channel corresponds to a feature map, ensuring that features of different types of raw data can be extracted individually; convolution kernel These are pre-trained feature extraction templates. Different channels have different convolutional kernel parameters. For example, convolutional kernels for image channels are good at extracting edge and texture features, while convolutional kernels for point cloud channels are good at extracting spatial location features. These convolutional kernels are then used to extract features from the feature map. Starting point, according to Element coordinate sliding enables feature capture of local regions in the feature map; the double summation operation first multiplies the convolution kernel within each channel with the local region of the feature map element by element and then sums them to obtain the local feature response of that channel; then, the feature responses of all channels are summed to achieve preliminary fusion of multi-channel features; bias term It is a trainable parameter used to adjust the baseline value of the convolution output to compensate for biases in the feature extraction process of different channels. For example, when the overall feature extraction result of a certain channel is low, the bias term... This can improve the feature response of the channel and ensure that the features of each channel are weighted evenly in subsequent processing. When multiple convolutional layers are stacked, the output of the previous convolution is used as the input of the next layer, gradually realizing the extraction of low-level features to high-level features, providing a rich feature base for object detection.

[0055] In this embodiment, the calculation expression for the batch normalization layer is:

[0056] In the formula, , The mean and variance are calculated from the training data, respectively. , These are trainable parameters. It is used to avoid the minimum value where the denominator is zero.

[0057] The above parameters are fixed values ​​during the model inference phase, therefore, given the input data... The calculation of the batch normalization layer in the model inference stage can be expressed as:

[0058] in, , Therefore, it can be seen that by studying... and The adjustment can enable the additional splitting of the batch normalization layer for calculation.

[0059] This embodiment analyzes the case of composing multiple functions that satisfy the condition of being additively decomposable, considering the function... and For inputs that are summable and separable If they are additively decomposable, then the composite function of the two can be derived as follows:

[0060] Therefore, the function and Composite function Additively decomposable computations are still possible. The computations of the convolutional layers and batch normalization layers can be considered as functions. and The perceptual inference part of the PIXOR model consists of multiple convolutional layers and batch normalization layers, and can be viewed as multiple... and The PIXOR model's perception and inference part can be additively decomposed for computation. By adjusting the model parameters, the feature fusion multi-vehicle cooperative perception task based on the PIXOR model can move the feature fusion position from its original position (the layer before the perception and inference part) to the last layer of the model (the layer after the perception and inference part). After the perception and inference part of the PIXOR model, the inference results of each cooperative vehicle are added and fused to obtain an inference result consistent with that before the feature fusion position was moved. Based on this, computational load balancing based on model segmentation is performed. The intelligent perception model is segmented according to the cooperative auxiliary information to determine the computational load for processing on the cooperative vehicle side and the computational load for processing on the main vehicle side. Figure 2 A schematic diagram of load balancing based on feature fusion and location movement model segmentation is shown.

[0061] In this embodiment, in step S4, the load balancing decision based on model segmentation is implemented through a reinforcement learning algorithm based on graph neural networks, and the reward function expression is:

[0062] In the formula, To meet the maximum collaborative sensing latency constraint, For actual collaborative sensing latency. To improve the accuracy of collaborative sensing. , The parameter is non-negative, and the algorithm optimization objective is to minimize the collaborative sensing latency and maximize the collaborative sensing accuracy.

[0063] Specifically, the master vehicle and its candidate vehicles within its communication range are modeled as a star graph topology, with the master vehicle as the central node and candidate vehicles as sub-nodes. Based on the cooperative assistance information, the central node's feature parameters are set to the master vehicle's computing power, and the sub-nodes' feature parameters are set to the candidate vehicles' computing power, signal transmission power, and cooperative relationship with the master vehicle. The feature parameters of the edges between the central node and sub-nodes are set to the channel state between the master vehicle and the candidate vehicles. The cooperative relationship between the candidate vehicle and the master vehicle indicates whether the current candidate vehicle was previously selected as a cooperative vehicle by the master vehicle. This graph topology is used as input to a graph neural network to extract the embedded features of each node and obtain the scheduling decision results. The master vehicle node outputs the cooperative vehicle quantity decision, and the candidate vehicle nodes output the cooperative selection probability of the corresponding candidate vehicle and the model splitting decision. Based on this, a reinforcement learning algorithm is used to learn and update the neural network parameters of each graph node until the final reward function is obtained. The convergence yields the final collaborative vehicle scheduling and computational load balancing decision results based on model segmentation.

[0064] In one possible embodiment, the topology of the graph neural network input is a star graph, with the master vehicle as the central node and the candidate vehicles as sub-nodes. The feature parameters of the central node are set as the computing power of the master vehicle, and the feature parameters of the sub-nodes are set as the computing power, signal transmission power, and cooperative relationship with the master vehicle of the candidate vehicle. The cooperative relationship indicates whether the candidate vehicle was selected as a cooperative vehicle by the master vehicle before the current time. The feature parameters of the edge between the central node and the sub-nodes are set as the channel state between the master vehicle and the candidate vehicle.

[0065] Specifically, the central node (master vehicle) only has "computing power" as its feature parameter. Since the master vehicle primarily undertakes the final feature fusion and inference tasks in the collaboration, its computing power directly determines how many collaborative vehicles' processing results it can handle, forming the basis for load balancing. No additional parameters are needed to avoid parameter redundancy affecting decision-making efficiency. The feature parameters of the sub-nodes (candidate vehicles) are designed around "ability to undertake collaborative tasks": computing power determines the model segmentation levels it can handle; vehicles with strong computing power can handle complex workloads with many convolutional layers, while vehicles with weak computing power only handle simple feature preprocessing layers. Signal transmission power determines its data transmission... Reliability and power are key factors in evaluating candidate vehicles. High-power vehicles transmit data with low bit error rates, making them suitable for allocating critical loads. The cooperative relationship reflects the stability of the cooperative system, indicating whether a cooperative vehicle has established a communication link with the master vehicle at the current moment, thus reducing link instability caused by frequent switching between cooperative vehicles. These three parameters together constitute the "cooperative capability evaluation index" for candidate vehicles. The "channel state," a feature parameter of the edge between the central node and the branch nodes, quantifies the data transmission conditions between the master vehicle and candidate vehicles. Edges with good channel state correspond to lower data transmission latency, and the master vehicle will prioritize allocating load to candidate vehicles with good channel state to avoid data transmission delays caused by poor channel conditions. This topology and feature parameter design enable GNN to comprehensively and accurately model the state of multi-vehicle cooperative systems, providing high-quality input data for load balancing decisions.

[0066] In one possible embodiment, the output of the graph neural network includes the cooperative vehicle number decision output by the master vehicle node, and the cooperative selection probability and model splitting decision output by the candidate vehicle nodes; the neural network parameters of the graph nodes are iteratively updated through a reinforcement learning algorithm until the reward function R converges, and finally the load balancing decision result is obtained.

[0067] Specifically, the "cooperative vehicle quantity decision" output by the master vehicle node needs to find a balance between "accuracy improvement" and "resource consumption": too many vehicles will lead to increased master vehicle fusion latency and bandwidth resource shortage, while too few vehicles will not meet the high-precision perception requirements. The GNN will output the optimal quantity based on the overall cooperative capability (average computing power, average channel state) of the candidate vehicles. For example, 5-8 vehicles can be selected when the overall capability of the candidate vehicles is strong, and 2-3 vehicles can be selected when the capability is weak. The "cooperative selection probability" output by the candidate vehicle node is the quantitative evaluation of the cooperative adaptability of each candidate vehicle by the GNN (value range 0-1). The higher the probability, the more suitable the candidate vehicle is to participate in the cooperation. The master vehicle will select the top-N candidate vehicles as cooperative vehicles according to the probability ranking to avoid blind selection. The model splitting decision clarifies the range of model layers that each cooperative vehicle needs to process. The GNN will allocate the corresponding model layers according to the computing power of the candidate vehicles to ensure that the computing load of each vehicle matches its capability and avoid the imbalance of "vehicles with strong computing power have light loads and vehicles with weak computing power have heavy loads". In the initial stage, the GNN parameters are randomized, and the output decision may suffer from load imbalance. The reward function calculated at this time... The latency is relatively low; reinforcement learning algorithms adjust the neural network parameters of GNN nodes based on reward feedback to optimize output decisions; after each parameter adjustment, the collaborative task is re-executed and the actual latency is calculated. With accuracy Update the reward function Repeat this process until the reward function is in successive iterations. The change is less than the preset threshold, that is Convergence occurs when the decision is at its optimal level, ensuring that the decision remains optimal even in dynamic environments and preventing instability caused by non-convergence of parameters.

[0068] In this embodiment, in step S5, the joint optimization scheduling of system bandwidth resources and main vehicle-side computing resources is achieved through a convex optimization algorithm. The optimization objective of the convex optimization algorithm is to reduce the waiting latency of collaborative vehicle task processing in order to minimize the overall collaborative perception latency.

[0069] Specifically, in multi-vehicle collaboration, bandwidth and computing resources are interrelated. Scheduling them individually can lead to resource waste or increased latency; therefore, joint scheduling using convex optimization algorithms is necessary. Convex optimization algorithms have the characteristics of a convex objective function and convex constraints, enabling them to efficiently find the global optimum and avoid getting trapped in local optima, thus meeting the resource scheduling requirements of multi-vehicle collaboration. The algorithm's optimization objective, "reducing waiting latency," is achieved through the following two aspects: In bandwidth resource scheduling, the convex optimization algorithm takes "minimizing the data transmission latency of each cooperating vehicle" as its sub-objective. Based on the model-based partitioning of load data volume and channel state of each vehicle, a convex optimization model for bandwidth allocation is established, with the constraint that "total bandwidth resources do not exceed the system's maximum bandwidth." This solution yields the optimal bandwidth allocation scheme for each vehicle, avoiding data transmission waiting due to insufficient bandwidth. In master vehicle-side computing resource scheduling, the algorithm takes minimizing the computational result latency of each cooperating vehicle's processing tasks as its sub-objective. Based on the computing power requirements of each vehicle's processing tasks and the master vehicle's remaining computing power, a convex optimization model for computing resource allocation is established, with the constraint that total computing resources do not exceed the master vehicle's computing power. This solution yields the optimal computing power allocation scheme for each vehicle, avoiding computational waiting due to insufficient computing power. Through the joint scheduling of bandwidth and computing resources, the superposition effect of transmission waiting and computation waiting is eliminated, ultimately minimizing the overall collaborative perception latency, meeting the stringent real-time requirements of autonomous driving.

[0070] In this embodiment, in step S7, the feature fusion adopts an additive fusion method, and the fusion formula is:

[0071] In the formula, to These are the processing results of the perception models corresponding to each cooperative vehicle. The main vehicle obtains the final cooperative perception inference result through this fusion formula.

[0072] Specifically, in step S4, the convolutional layer bias term has been adjusted. Batch normalized layer parameters and This ensures the processing results of each cooperating vehicle. to These correspond to the split portions of the total input data. to The inference results, based on the additivity of the model calculations, Equal to total input data The centralized reasoning results ensure that the accuracy of the summation and fusion of the processing results of each collaborative vehicle is consistent with that of the centralized processing.

[0073] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and these multiple devices will interact with each other to complete the task-driven multi-vehicle collaborative method described above.

[0074] It should be noted that the above description describes some embodiments of the present invention. In some cases, the described actions or steps can be performed in a different order than that shown in the above embodiments and the desired result can still be achieved. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0075] See Figure 3 Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this embodiment of the invention also provides a task-driven multi-vehicle cooperative system, employing the task-driven multi-vehicle cooperative method of the above embodiments, including: The collaborative request initiation module 100 is used by the master vehicle to initiate a collaborative perception request to at least one candidate vehicle within the communication range. The collaborative request response module 200 is used to determine whether to participate in collaborative perception after the candidate vehicle receives the collaborative perception request. If it agrees to participate, it sends collaborative assistance information to the master vehicle. The collaborative scheduling and topology construction module 300 is used by the master vehicle to perform collaborative vehicle scheduling operations and construct a collaborative perception topology based on the received collaborative assistance information, and to determine the number of vehicles and specific vehicles participating in collaborative perception. The model adjustment and load balancing decision module 400 is used to adjust the parameters of the intelligent perception model based on the structural characteristics of the multi-vehicle collaborative perception task, move the feature fusion position of the intelligent perception model to the last layer of the model, and make a calculation load balancing decision based on model segmentation based on the collaborative auxiliary information. The resource joint scheduling module 500 is used by the master vehicle to perform joint optimization scheduling decisions on system bandwidth resources and master vehicle-side computing resources based on the scheduling results of the cooperating vehicles and the computing load balancing decision, allocate bandwidth resources to the cooperating vehicles, and allocate computing resources to the model inference load corresponding to the cooperating vehicles on the master vehicle side. The decision distribution module 600 is used by the master vehicle to distribute the computational load balancing decision and the joint optimization scheduling decision to the selected cooperative vehicles. The collaborative perception and reasoning module 700 is used by the master vehicle and the collaborative vehicles to jointly execute the collaborative perception task according to the computational load balancing decision and the joint optimization scheduling decision, and to complete the feature fusion operation on the master vehicle side to obtain the final collaborative perception and reasoning result.

[0076] In this embodiment, the collaborative assistance information in the collaborative request response module 200 specifically includes vehicle position coordinates, vehicle speed, channel environment status, computing power, and signal transmission power.

[0077] In this embodiment, in the model adjustment and load balancing decision module 400, the intelligent perception model is a PIXOR model, and the perception inference part of the PIXOR model consists of multiple convolutional layers and batch normalization layers; wherein, the calculation expression for the area covered by the convolutional kernel in the convolutional layer is:

[0078] In the formula, The feature map corresponding to the input data channel. It is the convolution kernel corresponding to channel c. Represents the coordinates of the starting point of the convolution. Represents the element coordinates of the convolution kernel. For bias terms; The calculation expression for the batch normalization layer is:

[0079] In the formula, , The mean and variance are calculated from the training data, respectively. , These are trainable parameters. It is used to avoid the minimum value where the denominator is zero.

[0080] In this embodiment, in the model adjustment and load balancing decision module 400, when adjusting the PIXOR model parameters to achieve consistent inference accuracy, for the case where the input data satisfies the condition of being additively separable, the bias terms of the convolutional layer are adjusted. and parameters of the batch normalization layer and This enables the incremental splitting of computation between convolutional layers and batch normalized layers.

[0081] In this embodiment, in the model adjustment and load balancing decision module 400, the load balancing decision based on model segmentation is implemented through a reinforcement learning algorithm based on a graph neural network, and the reward function expression is:

[0082] In the formula, To meet the maximum collaborative sensing latency constraint, For actual collaborative sensing latency. To improve the accuracy of collaborative sensing. , The parameter is non-negative, and the algorithm optimization objective is to minimize the collaborative sensing latency and maximize the collaborative sensing accuracy.

[0083] In this embodiment, the model adjustment and load balancing decision module 400 constructs a star graph as the topology of the graph neural network input, with the main vehicle as the central node and the candidate vehicles as sub-nodes. The characteristic parameters of the central node are set as the computing power of the master vehicle, and the characteristic parameters of the sub-nodes are set as the computing power, signal transmission power and cooperative relationship with the master vehicle of the candidate vehicle. The cooperative relationship indicates whether the candidate vehicle was selected as a cooperative vehicle by the master vehicle before the current time. The characteristic parameters of the edge between the central node and the sub-nodes are set as the channel state between the master vehicle and the candidate vehicle.

[0084] In this embodiment, the model adjustment and load balancing decision module 400 outputs the graph neural network, including the cooperative vehicle number decision from the master vehicle node and the cooperative selection probability and model splitting decision from the candidate vehicle nodes. The neural network parameters of the graph nodes are iteratively updated using a reinforcement learning algorithm until the reward function... The convergence process yields the final load balancing decision.

[0085] In this embodiment, in the resource joint scheduling module 500, the joint optimization scheduling of the system bandwidth resources and the computing resources on the main vehicle side is achieved through a convex optimization algorithm. The optimization objective of the convex optimization algorithm is to reduce the waiting latency of collaborative vehicle task processing in order to minimize the overall collaborative perception latency.

[0086] In this embodiment, the feature fusion in the collaborative perception reasoning module 700 adopts an additive fusion method, and the fusion formula is as follows:

[0087] In the formula, to These are the processing results of the perception models corresponding to each cooperative vehicle. The main vehicle obtains the final cooperative perception inference result through this fusion formula.

[0088] The apparatus of the above embodiments is used to implement a task-driven multi-vehicle collaborative method corresponding to any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0089] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a task-driven multi-vehicle collaborative method as described in any of the above embodiments.

[0090] Figure 4This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 810, a memory 820, an input / output interface 830, a communication interface 840, and a bus 850. The processor 810, memory 820, input / output interface 830, and communication interface 840 are interconnected internally via the bus 850.

[0091] The processor 810 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0092] The memory 820 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 820 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 820 and is called and executed by the processor 810.

[0093] The input / output interface 830 is used to connect input / output modules to enable information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0094] The communication interface 840 is used to connect the communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0095] Bus 850 includes a pathway for transmitting information between various components of the device, such as processor 810, memory 820, input / output interface 830, and communication interface 840.

[0096] It should be noted that although the above-described device only shows the processor 810, memory 820, input / output interface 830, communication interface 840, and bus 850, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0097] The electronic devices described in the above embodiments are used to implement a task-driven multi-vehicle collaboration method corresponding to any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0098] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute a task-driven multi-vehicle collaborative method as described in any of the above embodiments.

[0099] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0100] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute a task-driven multi-vehicle cooperative method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0101] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the embodiments of the invention as described above, which are not provided in detail for the sake of brevity.

[0102] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of the invention, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of the invention, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of the invention will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that the embodiments of the invention may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0103] Although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAMDRAM) may be used with the embodiments discussed.

[0104] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the scope of protection of this invention.

Claims

1. A task-driven multi-vehicle collaborative method, characterized in that, Includes the following steps: S1. The master vehicle initiates a cooperative perception request to at least one candidate vehicle within the communication range. S2. After receiving the collaborative perception request, the candidate vehicle determines whether to participate in collaborative perception. If it agrees to participate, it sends collaborative assistance information to the master vehicle. S3. Based on the received collaborative assistance information, the master vehicle performs collaborative vehicle scheduling operations and constructs a collaborative perception topology to determine the number of vehicles participating in collaborative perception and the specific vehicles involved. S4. The main vehicle, taking into account the structural characteristics of the multi-vehicle collaborative perception task, adjusts the parameters of the intelligent perception model, moves the feature fusion position of the intelligent perception model to the last layer of the model, and makes a computational load balancing decision based on model segmentation based on the collaborative auxiliary information. S5. Based on the scheduling results of the cooperating vehicles and the computing load balancing decision, the master vehicle performs a joint optimization scheduling decision on the system bandwidth resources and the computing resources on the master vehicle side, allocates bandwidth resources to the cooperating vehicles, and allocates computing resources to the model inference load corresponding to the cooperating vehicles on the master vehicle side. S6. The master vehicle distributes the computational load balancing decision and the joint optimization scheduling decision to the selected cooperative vehicles. S7. The master vehicle and the cooperating vehicle jointly execute the collaborative perception task according to the computational load balancing decision and the joint optimization scheduling decision, and complete the feature fusion operation on the master vehicle side to obtain the final collaborative perception reasoning result.

2. The task-driven multi-vehicle cooperative method according to claim 1, characterized in that, In step S2, the cooperative assistance information specifically includes vehicle position coordinates, vehicle speed, channel environment status, computing power, and signal transmission power.

3. The task-driven multi-vehicle cooperative method according to claim 1, characterized in that, In step S4, the perceptual inference part of the intelligent perception model consists of multiple convolutional layers and batch normalization layers; wherein, the calculation expression for the region covered by the convolutional kernel in the convolutional layer is: ； In the formula, The feature map corresponding to the input data channel. It is the convolution kernel corresponding to channel c. This represents the coordinates of the starting point of the convolution. Represents the element coordinates of the convolution kernel. For bias terms; The calculation expression for the batch normalization layer is: ； In the formula, , The mean and variance are calculated from the training data, respectively. , These are trainable parameters. It is used to avoid the minimum value where the denominator is zero.

4. The task-driven multi-vehicle cooperative method according to claim 3, characterized in that, When adjusting the parameters of the PIXOR model to achieve consistent inference accuracy, for input data that satisfy the condition of additive separability, the bias terms of the convolutional layers are adjusted. and parameters of the batch normalization layer and This enables the incremental splitting of computation between convolutional layers and batch normalized layers.

5. The task-driven multi-vehicle cooperative method according to claim 1, characterized in that, In step S4, the load balancing decision based on model segmentation is implemented using a reinforcement learning algorithm based on graph neural networks, and the reward function expression is: ； In the formula, To meet the maximum collaborative sensing latency constraint, For actual collaborative sensing latency. To improve the accuracy of collaborative sensing. , The parameter is non-negative, and the algorithm optimization objective is to minimize the collaborative sensing latency and maximize the collaborative sensing accuracy.

6. The task-driven multi-vehicle cooperative method according to claim 5, characterized in that, The topology of the graph neural network input is a star graph, with the main vehicle as the central node and the candidate vehicles as sub-nodes; The characteristic parameters of the central node are set as the computing power of the master vehicle, and the characteristic parameters of the sub-nodes are set as the computing power, signal transmission power and cooperative relationship with the master vehicle of the candidate vehicle. The cooperative relationship indicates whether the candidate vehicle was selected as a cooperative vehicle by the master vehicle before the current time. The characteristic parameters of the edge between the central node and the sub-nodes are set as the channel state between the master vehicle and the candidate vehicle.

7. The task-driven multi-vehicle cooperative method according to claim 5, characterized in that, The output of the graph neural network includes the cooperative vehicle number decision output by the master vehicle node, and the cooperative selection probability and model segmentation decision output by the candidate vehicle nodes; the neural network parameters of the graph nodes are iteratively updated through a reinforcement learning algorithm until the reward function is obtained. The convergence process yields the final load balancing decision.

8. The task-driven multi-vehicle cooperative method according to claim 1, characterized in that, In step S5, the joint optimization scheduling of system bandwidth resources and main vehicle side computing resources is achieved through a convex optimization algorithm. The optimization objective of the convex optimization algorithm is to reduce the waiting latency of collaborative vehicle task processing in order to minimize the overall collaborative perception latency.

9. The task-driven multi-vehicle cooperative method according to claim 1, characterized in that, In step S7, the feature fusion adopts an additive fusion method, and the fusion formula is: ； In the formula, to These are the processing results of the perception models corresponding to each cooperative vehicle. The main vehicle obtains the final cooperative perception inference result through this fusion formula.

10. A task-driven multi-vehicle cooperative system, employing the task-driven multi-vehicle cooperative method according to any one of claims 1 to 9, characterized in that, include: The collaborative request initiation module is used by the master vehicle to initiate a collaborative perception request to at least one candidate vehicle within the communication range. The collaborative request response module is used to determine whether to participate in collaborative perception after the candidate vehicle receives the collaborative perception request. If it agrees to participate, it sends collaborative assistance information to the master vehicle. The collaborative scheduling and topology construction module is used by the master vehicle to perform collaborative vehicle scheduling operations and construct a collaborative perception topology based on the received collaborative assistance information, and to determine the number of vehicles and specific vehicles participating in collaborative perception. The model adjustment and load balancing decision module is used to adjust the parameters of the intelligent perception model based on the structural characteristics of the multi-vehicle collaborative perception task, move the feature fusion position of the intelligent perception model to the last layer of the model, and make a calculation load balancing decision based on the model segmentation based on the collaborative auxiliary information. The resource joint scheduling module is used by the master vehicle to make joint optimization scheduling decisions on system bandwidth resources and master vehicle side computing resources based on the scheduling results of the cooperating vehicles and the computing load balancing decision, allocate bandwidth resources to cooperating vehicles, and allocate computing resources to the model inference load corresponding to the cooperating vehicles on the master vehicle side. The decision distribution module is used by the master vehicle to distribute the computational load balancing decision and the joint optimization scheduling decision to the selected cooperative vehicles; The collaborative perception and reasoning module is used by the master vehicle and the collaborative vehicles to jointly execute the collaborative perception task according to the computational load balancing decision and the joint optimization scheduling decision, and to complete the feature fusion operation on the master vehicle side to obtain the final collaborative perception and reasoning result.

Images