Task-driven vehicle calculation unloading method

Abstract

The invention belongs to the technical field of Internet of Vehicles, and particularly relates to a task-driven vehicle calculation unloading method. The method comprises the following steps: firstly, training a precision neural network estimation model; establishing a time delay model and a calculation cost model of the vehicle target detection task in different unloading modes; secondly, according to the estimation model, estimating the precision of image acquisition of the vehicle end and the road end, according to the time delay model and the calculation cost model, calculating the time delay and the calculation cost corresponding to the service, and according to the estimation precision, the task completion time delay and the task completion cost, constructing an optimization problem; solving the optimization problem through a genetic algorithm to obtain an optimal solution; and finally, according to the obtained optimal solution, completing selection decision, resource allocation and calculation unloading of the vehicle-mounted service. According to the invention, a precision neural network prediction model is trained, and on the basis, a task-driven vehicle calculation unloading method is provided, so that the perception precision is maximized in an acceptable time delay and calculation cost range.

Description

Technical field:

[0001] The present invention belongs to the technical field of vehicle networking, and in particular relates to a task-driven vehicle computing offloading method. Background technology:

[0002] In the specific scenarios of the Internet of Vehicles (IoV), the large number of end users and the diverse service demands present significant challenges to network and computing resources. Various devices, including vehicles, road infrastructure, and smartphones supporting V2X services, must handle a wide range of tasks, from basic road safety services like emergency broadcasts to value-added services such as parking location and augmented reality provided by application developers and content providers. These services require low latency and high reliability, while also placing high demands on computing power. However, as mobile terminals, vehicles have relatively limited computing resources when processing large amounts of data or performing complex calculations, making it difficult to meet the needs of all applications.

[0003] To address these issues, Mobile Edge Computing (MEC) technology has been introduced into the Internet of Vehicles (IoV), demonstrating significant potential for improving network performance. By deploying MEC servers on base stations and roadside units (ROSs), vehicles offload computing tasks to the edge, alleviating onboard computing pressure. MEC servers analyze information provided by onboard applications and local data received by road sensors, prioritizing critical safety-critical tasks and critically important computational tasks. This not only helps avoid traffic congestion but also improves the user experience for drivers and passengers. Furthermore, MEC optimizes the allocation of computing and transmission resources through methods such as reinforcement learning, and optimizes latency for latency-sensitive applications, ensuring that vehicles receive the latest information. In multi-vehicle and multi-server environments, MEC can also coordinate task uploads and migrations to avoid network congestion and ensure service continuity. These applications demonstrate the enormous potential of MEC technology in resource-constrained, latency-sensitive, and dynamically changing network environments.

[0004] At present, research on computational offloading of integrated telematics and computing in the Internet of Vehicles mainly focuses on common evaluation indicators of communication systems, such as optimization of latency and computational cost. The document with application number "202311165393.7" discloses "A method and system for mobile edge task offloading based on the Internet of Vehicles", which provides a mobile edge task offloading method based on the Internet of Vehicles, jointly considers offloading decisions, latency and cost, and solves the optimal task offloading strategy through real-time cost prediction and genetic algorithm. The problem is that when processing computationally intensive services, the differentiated characteristics of the services themselves are often not fully considered, especially ignoring the specific impact of different offloading methods on various perception tasks (such as target detection). This deficiency may reduce the accuracy of perception tasks, and thus have a negative impact on the service experience of Internet of Vehicles users and the safety of autonomous driving. Summary of the invention:

[0005] The purpose of the present invention is to provide a task-driven vehicle computing offloading method to solve the problem in the prior art that different offloading methods do not fully consider the specific impact of various perception tasks.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows: a task-driven vehicle calculation offloading method, a task-driven calculation offloading method comprising the following steps:

[0007] Step 1: Obtain a precision neural network estimation model by using image data and YOLOv7 model for offline pre-training;

[0008] Step 2: Obtain the application scenarios of the Internet of Vehicles and establish the delay model and computational cost model for different offloading modes of the vehicle target detection task;

[0009] Step 3: Estimate the accuracy of images acquired by the vehicle and road using the trained precision neural network estimation model. Calculate the latency and computational cost corresponding to the service using the established latency model and computational cost model. Construct an optimization problem based on the estimated accuracy, task completion latency, and task completion cost.

[0010] Step 4: Solve the optimization problem by genetic algorithm to obtain the optimal solution;

[0011] Step 5: According to the obtained optimal solution, complete the selection decision, resource allocation and computation offloading of the in-vehicle service.

[0012] Furthermore, in the above step 1, the accuracy neural network estimation model is expressed as

[0013]

[0014] Among them, P represents the input image, n is the vehicle serial number, a nRepresents the estimated classification accuracy for the vehicle n target detection task.

[0015] Furthermore, the unloading mode in the above step 2 refers to the local unloading mode and the road-side unloading mode.

[0016] Furthermore, in the above step 2, the delay model in the local offloading mode is defined as

[0017]

[0018] Among them, b n Indicates the data size of the target detection task, represents the local computing resource size of vehicle n, Indicates the proportion of vehicle local unloading resources allocated to the object detection task.

[0019] The calculation cost model is

[0020]

[0021] Where μ1 represents the payment cost per unit of computing resources of the vehicle.

[0022] Furthermore, in the above step 2, the delay model in the road-side unloading mode is defined as:

[0023]

[0024] The computational cost model is defined as:

[0025]

[0026] Among them, μ2 represents the payment cost of RSU per unit computing resource, It represents the proportion of roadside unit computing resources allocated by RSU to the vehicle n target detection task.

[0027] Furthermore, in the above step 3, the task completion delay t final yes

[0028]

[0029] The task completion cost of the target detection task is

[0030]

[0031] Furthermore, in the above step 3, the optimization problem is:

[0032]

[0033] Furthermore, in the above step 3, the various resource constraints are

[0034]

[0035]

[0036] C4:t final ≤T (16)

[0037] C5:E≤E max (17)

[0038] Furthermore, the genetic algorithm in step 4 above takes as input the prediction accuracy of the YOLOv7 model on both the road side and the vehicle side, as well as the system's constraints on the maximum allowable delay and the maximum allowable computing cost, and outputs the optimal solution.

[0039] Furthermore, the resource allocation in the above step five includes communication resource allocation and computing resource allocation.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. The purpose of the present invention is to maximize the accuracy of target detection in order to provide highly interpretable auxiliary information to help the driver or the autonomous driving system make better decisions. To this end, before constructing the problem, the present invention trains a precision neural network estimation model, forming a mapping from image to estimated accuracy, and providing sufficient basis and interpretability for subsequent autonomous driving decisions. The model constructed in step 1 adopts a transfer learning method based on YOLOv7, and conducts targeted training through scene-related dedicated data sets to construct an optimized model that focuses on the core object categories of the target detection task, significantly improving the detection accuracy and feature extraction capabilities of the model in specific scenarios.

[0042] 2. This paper uses the task of object detection in high-dynamic scenarios within the connected vehicle (IoV) as a starting point, deriving a mapping from images to estimated accuracy. Based on this, a task-driven vehicle computation offloading method is proposed. This method builds on the traditional task offloading model in the IoV by incorporating a collaborative offloading model. This method further improves object detection accuracy while maintaining service latency requirements, thereby significantly enhancing the safety of autonomous driving.

[0043] 3. The present invention combines local offloading mode and roadside offloading mode in the same IoV application scenario, collaborating to achieve offloading. In this collaborative offloading mode, the present invention can always maintain the highest target detection accuracy. Description of the drawings:

[0044] Figure 1 Flowchart provided for the present invention;

[0045] Figure 2A schematic diagram of the training process of the precision neural network estimation model provided by the present invention;

[0046] Figure 3 This is a diagram of a task offloading model based on integrated synaesthesia and computing in the Internet of Vehicles scenario provided by the present invention;

[0047] Figure 4 A flow chart of the genetic algorithm provided by the present invention;

[0048] Figure 5 This is a comparison chart of the changes in accuracy with the number of users in different modes proposed by the present invention.

[0049] Figure 6 This is a comparison chart of the delay changes with the number of users in different modes provided by the present invention.

[0050] Figure 7 This is a comparison chart of how the total computing cost changes with the number of users in different modes provided by the present invention. Specific implementation method:

[0051] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0052] The framework of the present invention is as follows Figure 1 As shown, first, a precision neural network estimation model is trained offline on both the vehicle side and the road side. Next, the IoV application scenario is identified, and a latency model and computational cost model for the in-vehicle service are established. Second, the accuracy of the imagery acquired on the vehicle side and the road side is estimated based on the trained neural network estimation model. The latency and computational cost corresponding to the service are calculated based on the established latency and computational cost models. An optimization problem is constructed based on the estimated accuracy, maximum latency, and maximum computational cost. This optimization problem is then solved using a genetic algorithm to obtain the optimal solution. Finally, based on the optimal solution, in-vehicle service selection decisions, resource allocation, and computational offloading are completed.

[0053] Based on the above ideas, the present invention proposes a task-driven computing offloading method, which specifically includes the following steps:

[0054] Step 1: Use image data and the YOLOv7 model for offline pre-training to obtain a precision neural network estimation model and construct a mapping from image to estimated accuracy:

[0055] Before proposing a task offloading decision, the present invention first trains a precision neural network model on the vehicle side and the road side, and deploys it on the intelligent connected vehicle and RSU to assist the decision-making process.

[0056] This paper adopts supervised learning method, specifically, YOLOv7 is used as the basic model. The model takes image data as input and outputs an estimate of the target position. The precision neural network estimation model is expressed as

[0057] The accuracy neural network estimation model is expressed as:

[0058]

[0059] Among them, P represents the input image, a n represents the estimated classification accuracy for the vehicle n object detection task. The model is obtained by offline pre-training using simulated image data.

[0060] See also Figure 2 The steps for training a precision neural network model are divided into four steps: finding a suitable dataset; training a custom YOLOv7 (You Only Look Once Version 7) model; building a precision model dataset; and training a precision neural network model. Existing general-purpose precision models cannot meet the requirements of this research scenario. By following these steps, we can train a precision neural network model optimized for this scenario.

[0061] (1) Finding a suitable dataset:

[0062] This embodiment uses the DAIRV2X dataset as the dataset, and divides the dataset into a training set and a validation set at a ratio of 85% and 15%.

[0063] (2) Training a new YOLOv7 model:

[0064] First, in the DAIRV2X dataset used, the coordinate format of the data is: (x min ,y min ,x max ,y max ). Therefore, before formal training, the bounding box coordinates in the DAIRV2X dataset are converted to the YOLOv7 format (x center ,y center ,width,height), thus making it suitable for the input requirements of the YOLOv7 model.

[0065] Then, through training, the YOLOv7 model can perform specific target detection, and the detection category is customized by the user. After the training in this embodiment is completed, the YOLOv7 model can output text information and images with object detection locations and object categories on the image.

[0066] (3) Based on the image output from step (2), compare the detection locations identified by the YOLOv7 model with the original label locations in the dataset. Calculate the intersection and union of the object labels and detection locations on the image, then calculate the intersection-over-union ratio (IoU). Then, average the IoU ratios for all objects in the image as the accuracy evaluation metric. After performing this processing on each image, the corresponding accuracy for each image is obtained, thus completing the construction of the object detection accuracy model dataset.

[0067] (IV) Training precision neural network model:

[0068] First, data preparation and preprocessing are performed, dividing the target detection accuracy model dataset into a training set and a validation set at a ratio of 80% and 20%. Next, the target detection accuracy model is trained to obtain the final trained accuracy neural network estimation model.

[0069] Step 2: Obtain the delay model and computational cost model of the vehicle network application scenario and vehicle target detection task under different offloading modes.

[0070] See also Figure 3 The one-way two-lane road is covered by the perception area of ​​the RSU. In this area, there are multiple intelligent connected vehicles equipped with target detection models. The vehicles will detect objects in their area of ​​interest and classify the detected results into known categories.

[0071] The unloading modes mentioned refer to local unloading mode and road-side unloading mode.

[0072] The latency model includes data transmission latency and vehicle-borne service computation latency. Transmission latency can be calculated based on transmission rate and data size, while computation latency can be calculated based on computing power and task size.

[0073] The computational cost model takes into account the data transmission cost and the computational cost of processing the in-vehicle services.

[0074] The entire system has N intelligent connected vehicles, represented by N = {1, 2, ..., N}, and each vehicle is equipped with computing resources At the same time, RSU is also equipped with its own computing resources, using f RSU There are C categories in target detection, which are represented by C = {1, 2, …, C}.

[0075] For each object detection task, the considered integrated sensing and computing framework has two offloading modes: local offloading and roadside offloading. Specifically, local offloading means the vehicle performs object detection locally; roadside offloading means the vehicle transmits the task instruction to a nearby RSU, which then uses the MEC server to perform object detection.

[0076] In order to ensure the effective offloading and processing of vehicle target detection tasks and avoid repeated processing of tasks, the present invention introduces a binary variable d n To express the selection decision of the detection end. Specifically, when d n = 0, it means the local unloading mode is adopted; when d n =1, it means the road-side unloading mode is adopted.

[0077] The mission data packet mainly contains the mission input data (image data collected by the vehicle), the mission identifier (including the vehicle-side label and vehicle-side coordinates, used to uniquely identify and locate the mission source), and the quality of service requirements (such as latency, bandwidth, and computing resource requirements). The mission instruction contains the processing instruction (notifying the RSU of the target detection task to be processed), the mission identifier, and the quality of service requirements. It does not contain the mission input data to reduce the amount of data transmission and improve processing efficiency.

[0078] Next, we will show the latency and computational cost calculation of each part in detail.

[0079] (1) Latency model and computational cost model in local offloading mode:

[0080] In local offloading mode, since task data does not need to be transmitted, there is no communication delay and transmission cost. In this mode, task processing only involves computational delay and computational cost.

[0081] The calculation delay is expressed as follows:

[0082]

[0083] Among them, b n Indicates the data size of the target detection task, represents the local computing resource size of vehicle n, Indicates the proportion of vehicle local unloading resources allocated to the object detection task.

[0084] In the task processing scenario, the computational cost can be expressed as follows:

[0085] Where μ1 represents the payment cost per unit of computing resources of the vehicle.

[0086] (2) Delay model and computational cost model in road-side offloading mode:

[0087] When executing a task in the RSU, the vehicle sends a calculation instruction to the RSU. The delay includes the instruction transmission delay, the calculation delay and the downlink result return delay. Since the downlink result return delay is very small, the present invention ignores the downlink feedback delay.

[0088] First, calculate the Shannon upper bound of channel transmission according to the following formula:

[0089] Where B represents the overall bandwidth of all vehicles and RSU transmission signals, β n Indicates the ratio of bandwidth resources allocated to vehicle n target detection tasks, n is the transmission power of vehicle n, σ 2 is the Gaussian white noise power, g n is the channel gain of vehicle n, k n is the distance from vehicle n to RSU, F com is the communication frequency between the vehicle and the RSU (assuming that all vehicles have the same communication frequency), and h is the height of the RSU.

[0090] The instruction transmission delay is as follows:

[0091]

[0092] Among them, b n,comm Indicates the size of the transfer instruction.

[0093] When the computation instruction reaches the RSU, the computation delay at the RSU is as follows:

[0094]

[0095] Among them, f RSU Indicates the computing resource size of the roadside unit, It represents the proportion of roadside unit computing resources allocated by RSU to the vehicle n target detection task.

[0096] To ensure consistency between the RSU's calculation results and the vehicle's desired results, coordinate conversion is performed on the RSU's results. Specifically, matrix operations are performed using the vehicle coordinates included in the task instruction to convert the RSU data to the vehicle coordinate system. The results from different terminals are then compared, and the one with the highest confidence is retained.

[0097] Since the amount of task result data is small, the coordinate conversion delay and download delay are ignored. In summary, the delay model in the road-side unloading mode is defined as

[0098]

[0099] Compared with cooperative detection, the computational cost of the roadside unloading mode is very small and can be ignored. The computational cost model is expressed as follows:

[0100]

[0101] Where μ2 represents the payment cost of RSU per unit of computing resources.

[0102] Step 3: Estimate the accuracy of images acquired by the vehicle and road based on the trained neural network prediction model. Calculate the latency and computational cost corresponding to the service based on the established latency model and computational cost model. Construct an optimization problem based on the estimated accuracy, maximum latency, and maximum computational cost.

[0103] The task completion delay depends on the highest delay of different tasks in the two modes, i.e.

[0104]

[0105] The task completion cost is the sum of the individual costs, i.e.

[0106]

[0107] The present invention considers the optimal target detection accuracy under the condition of limited communication resources, computing resources and time delay. n , RSU bandwidth allocation variable β n , the computing resource variables of vehicles and RSUs and Optimize the target detection accuracy. The present invention hopes that the target detection accuracy is as high as possible. The final optimization problem is expressed as follows:

[0108]

[0109] C4:t final ≤T (16)

[0110] C5:E≤E max (17)

[0111] Among them, C1 represents the local computing resource constraint, C2 represents the bandwidth resource constraint, C3 represents the RSU computing resource constraint, C4 represents the delay constraint, and C5 represents the computing cost constraint. T represents the maximum acceptable delay, E max Indicates the maximum acceptable computational cost.

[0112] Step 4: Solve the optimization problem through genetic algorithm, input the decision variable set and the value of part of the objective function, and output the optimal solution.

[0113] like Figure 4 As shown in the figure, the genetic algorithm mainly follows the following steps to obtain the optimal solution.

[0114] The present invention will first explain the definition of the fitness function to facilitate the understanding of the subsequent content.

[0115] The fitness function primarily considers accuracy A, calculated based on the detection-side selection decisions and the prediction accuracy of the image after previous vehicle-side and road-side processing. To ensure optimal object detection accuracy while meeting latency and cost constraints, the algorithm incorporates latency and cost constraints as penalty terms in the fitness function.

[0116] Next, the present invention will explain in detail the specific steps of the solution algorithm according to the flow chart.

[0117] At the initial stage of the optimization process, known as the input phase, the system first receives and processes a series of key data, including the YOLOv7 model prediction accuracy for both the roadside and vehicle sides, as well as the system's constraints on latency and cost.

[0118] Entering the initialization phase, the algorithm begins to generate an original population consisting of J individuals, each of which represents a potential solution, covering the selection decision and resource allocation of the detection end, and the iteration counter j is set to 0.

[0119] After initialization, the algorithm enters an iterative process. At the beginning of each iteration, the system uses the optimal solution determined in iteration k as the starting point for iteration k+1 and sets the iteration counter j to 1. This step ensures that each iteration has a current optimal solution as the starting point, which helps accelerate convergence and improve optimization efficiency. In each iteration, the algorithm randomly selects a solution and the current optimal solution as parent chromosomes and performs selection, crossover, and mutation operations on these parent chromosomes. Using these genetic operations, the algorithm generates new candidate solutions.

[0120] In each iteration, the algorithm solves the proposed optimization problem of optimizing the prediction accuracy of object detection while meeting the latency and cost constraints. This step uses the fitness function proposed above to evaluate each newly generated candidate solution and adjust it according to the constraints.

[0121] Next, the algorithm determines whether the current number of iterations, j, is greater than or equal to J-1. If so, meaning a sufficient number of iterations have been completed, the algorithm proceeds to the next step. If not, j is incremented by 1 and the algorithm returns to the selection and crossover steps to generate new candidate solutions. This process repeats until the preset number of iterations is reached or another termination condition is met.

[0122] If j reaches the termination condition, the algorithm determines whether it has reached the last iteration. If so, the algorithm enters the output phase, outputting the optimal accuracy, the detection end selection decision, and the resource allocation. If not, the algorithm proceeds to the next iteration.

[0123] Step 5: Based on the optimal solution obtained in step 4, complete the selection decision, resource allocation, and computation offloading of the in-vehicle service:

[0124] The resource allocation includes communication resource allocation and computing resource allocation. Computing resource allocation refers to how much computing resources the MEC server allocates to the task vehicle, and communication resource allocation refers to how much bandwidth is allocated to each task vehicle.

[0125] To test the proposed method, we compared the performance of a coexisting local and roadside offloading modes, namely the collaborative offloading mode, with those of a single-end (vehicle-side) offloading mode and a single-end (road-side) offloading mode. The experimental parameters are shown in Table 1.

[0126] The basic configuration of the experimental equipment is NVIDIA GeForce RTX 3090.

[0127] Figure 5 This is a comparison of the changes in accuracy with the number of users in the different modes proposed by the present invention. The two-end collaborative offloading mode performs best in terms of accuracy and can achieve the highest perception accuracy. The single-end (road-end) offloading mode is second, and the single-end (vehicle-end) offloading mode has relatively low accuracy and may not be suitable for application scenarios with high precision requirements. The reason for this phenomenon may be that the road-end has a wider field of view, so it has higher perception accuracy in most cases; while the vehicle-end has a limited field of view and lower perception accuracy; but the two-end collaboration can always select the higher perception accuracy end for task offloading, so the perception accuracy can always be maintained at the highest.

[0128] Figure 6 This is a comparison of the latency changes with the number of users in the different modes proposed by the present invention. The latency of the single-end (vehicle-end) unloading mode is higher, and it rises rapidly with the increase in the number of users, which may not be suitable for application scenarios with high real-time requirements. The latency of the two-end collaborative unloading mode and the single-end (road-end) unloading mode is lower, which can better meet the real-time requirements. The reason for this phenomenon is that the computing resources in the road-end unloading mode are more sufficient, the computing power is stronger, and the computing latency will be relatively low. In contrast, the vehicle-end computing latency will increase. The collaborative unloading mode is the synergistic result of the two modes, and the latency is between the two. In particular, in the stage where the number of users is 10-14, the latency of individual users in the two-end unloading mode may be greater than that in the vehicle-end unloading mode. This is because the vehicle-end accuracy is greater in the case at that time, and the instruction transmission delay makes the latency slightly greater than that of the vehicle-end. However, in any case, the latency upper limit requirement is always met.

[0129] Table 1 Experimental parameters

[0130]

[0131] Figure 7 This is a comparison of the total computing cost as the number of users changes under the different modes proposed in the present invention. In the single-end (vehicle-end) unloading mode, the computing resource demand is extremely low, and it hardly changes with the increase in the number of users, and it has been maintained below 0.01. This shows that the single-end (vehicle-end) unloading mode has certain advantages in terms of computing resource demand, and is suitable for application scenarios with relatively limited computing resources. In the single-end (road-end) unloading mode, its computing resource demand will gradually increase with the increase in the number of users, from 0.01 to close to 0.07, and the growth trend is more obvious. In the two-end collaborative unloading mode, the computing resource demand corresponding to the blue line will also increase with the increase in the number of users, but the growth rate is relatively small, from 0.01 to close to 0.05, maintaining between the two single-end unloading modes.

[0132] For different application requirements, the following choices can be made: If real-time requirements are high, the single-end (road-side) offloading mode is an efficient and stable choice that can effectively control latency; if resources are very limited and accuracy requirements are not high, the single-end (vehicle-side) offloading mode is more suitable. For comprehensive performance requirements that require high accuracy, real-time performance, and efficient resource management, the two-end collaborative offloading mode is the best choice.

[0133] The above description is an explanation of the specific implementation of the present invention, rather than a limitation of the present invention. Those skilled in the relevant technical field can also make various equivalent technical solutions without departing from the scope of the present invention, so all equivalent technical solutions should be included in the scope of protection of the present invention.

Claims

1. A task-driven computing offloading method, characterized by: The following steps are involved: Step 1: Obtain a precision neural network estimation model by using image data and YOLOv7 model for offline pre-training; Step 2: Obtain the application scenarios of the Internet of Vehicles and establish the delay model and computational cost model for different offloading modes of the vehicle target detection task; Step 3: Estimate the accuracy of images acquired by the vehicle and road using the trained precision neural network estimation model. Calculate the latency and computational cost corresponding to the service using the established latency model and computational cost model. Construct an optimization problem based on the estimated accuracy, task completion latency, and task completion cost. Step 4: Solve the optimization problem by genetic algorithm to obtain the optimal solution; Step 5: According to the obtained optimal solution, complete the selection decision, resource allocation and computation offloading of the in-vehicle service.

2. The task-driven computing offloading method according to claim 1, characterized in that: In step 1, the precision neural network estimation model is expressed as Among them, P represents the input image, n is the vehicle serial number, a n Represents the estimated classification accuracy for the vehicle n target detection task.

3. The task-driven computing offloading method according to claim 1, characterized in that: The unloading mode in step 2 refers to the local unloading mode and the road-side unloading mode.

4. The task-driven computing offloading method according to claim 3, characterized in that: In step 2, the delay model in the local offloading mode is defined as Among them, b n Indicates the data size of the target detection task, represents the local computing resource size of vehicle n, Indicates the proportion of vehicle local unloading resources allocated to the object detection task. The calculation cost model is Where μ1 represents the payment cost per unit of computing resources of the vehicle.

5. The task-driven computing offloading method according to claim 3, characterized in that: In step 2, the delay model in the road-side unloading mode is defined as: The computational cost model is defined as: Among them, μ2 represents the payment cost of RSU per unit computing resource, It represents the proportion of roadside unit computing resources allocated by RSU to the vehicle n target detection task.

6. A task-driven computing offloading method according to any one of claims 4-5, characterized in that: In step 3, the task completion delay t final yes The task completion cost of the target detection task is 7. The task-driven computing offloading method according to claim 6, characterized in that: In step 3, the optimization problem is:

8. The task-driven computing offloading method according to claim 7, characterized in that: In step 3, the various resource constraints are: C4:t final ≤T (16) C5:E≤E max (17)。 9. The task-driven computing offloading method according to claim 8, characterized in that: The genetic algorithm in step 4 is inputted with the prediction accuracy of the YOLOv7 model on the road side and the vehicle side, as well as the system's constraints on the maximum allowable delay and the maximum allowable computing cost, and outputs the optimal solution.

10. The task-driven computing offloading method according to claim 9, characterized in that: The resource allocation in step five includes communication resource allocation and computing resource allocation.

Images