Cooperative perception and calculation unloading joint optimization method in Internet of Vehicles

Abstract

The invention belongs to the field of Internet of Vehicles, and relates to a cooperative sensing and calculation unloading joint optimization method, which comprises the following steps of: constructing a cooperative sensing model and a cooperative calculation stage model; on the premise that constraint conditions are met, the perception quality is improved by reducing redundancy, and an objective function with the purpose of minimizing the maximum task completion time is established; in order to solve the proposed optimization problem, firstly, the problem is decoupled into an optimization problem 1 in a cooperative sensing stage and an optimization problem 2 in a cooperative computing stage; in order to solve the optimization problem 1, a vehicle sensing block selection algorithm is designed; a DDPG-HA algorithm is designed to solve the optimization problem 2; according to the method, the perceptual performance and the task completion time delay are taken as indexes, the effectiveness of the method is verified in a simulation result, and the superior perceptual performance and the perceptual task processing capability are shown.

Description

Technical Field

[0001] The present invention belongs to the field of Internet of Vehicles, and specifically relates to a method for joint optimization of collaborative perception and computation offloading. Background Art

[0002] The statements in this section only provide background information related to the present disclosure, and these statements may constitute prior art. In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art.

[0003] Autonomous driving technology is a core area of ​​development for intelligent transportation systems. Within this technological framework, intelligent vehicle perception systems play a crucial role. Their core mission is to collect environmental data through a variety of sensors, enabling accurate identification and understanding of surrounding objects, road conditions, and other traffic participants. Accurate and comprehensive vehicle perception helps autonomous vehicles adapt to complex traffic environments, supporting decisions such as obstacle avoidance, lane selection, speed control, and traffic signal response, thereby improving the feasibility and safety of autonomous driving. However, the perception range of vehicle sensors is limited by physical properties and environmental interference, and blind spots are prone to occur when obstructions are present. Furthermore, extreme weather and poor lighting conditions can significantly impact sensor performance, leading to significant limitations in vehicle perception systems. To address the challenges of single-vehicle perception caused by these limitations, the concept of collaborative perception has been proposed. By enabling heterogeneous data sharing and fusion between vehicles and between vehicles and infrastructure, collaborative perception can extract more advanced features or lightweight sensing results, thereby jointly constructing a more comprehensive and accurate environmental model. Collaborative perception not only expands the perception range but also improves perception accuracy, effectively alleviating perception challenges in various driving scenarios. Since perception information is time-sensitive, the amount of perception task data is large, and the computing resources of vehicles and RSUs are limited, computational offloading technology can be used to offload perception tasks to the most appropriate target for processing to ensure the timeliness of perception task processing, thereby improving driving experience and safety.

[0004] Existing research on improving perception performance is primarily categorized into optimal viewpoint perception, random perception, and global perception. The optimal viewpoint perception method selects a single viewpoint for perception through disjoint spatial partitioning, and transmits and processes the corresponding data. While this reduces communication and computing resource consumption, it fails to leverage the complementary perspectives of multiple vehicles, potentially leading to reduced perception performance. Global perception, on the other hand, integrates sensor data from all vehicles through data fusion, improving perception accuracy and confidence. However, as the number of vehicles increases, the amount of data increases linearly, leading to increased data redundancy and resource consumption, and limited improvement amidst diminishing marginal returns. In contrast, the proposed random perception method rotates data from a number of existing perception vehicles for perception, reducing data redundancy. However, this method can lead to inconsistent perception results and potentially suboptimal perception quality. Therefore, how to rationally select perception vehicles and reduce data volume while ensuring perception performance remains a critical challenge that needs to be addressed.

[0005] In terms of sensory data processing, some existing research focuses on the problem of complete offloading in mobile edge computing (MEC)-assisted connected vehicles, where all computational tasks are offloaded to MEC servers. While these solutions address computational resource bottlenecks to some extent, when all tasks are offloaded to remote devices, particularly under poor network conditions, the large amount of data transmitted at once can lead to data congestion, transmission delays, and even data loss, compromising the timeliness and accuracy of the overall task processing. A fixed-ratio offloading approach has been proposed, which, compared to complete offloading, can efficiently utilize the computational resources of multiple offloading nodes and reduce reliance on remote resources. However, this approach fails to account for differences in computing power, resource utilization, load, and network status across devices. For example, some devices with stronger computational resources may not be fully utilized, while others with heavier loads may be overloaded with tasks, resulting in longer task completion times and reduced system efficiency. This can lead to resource waste or device overload. Furthermore, when vehicle density is high or the RSU is already burdened with a large number of computational tasks, the RSU's computational resources can become severely limited. In this case, continuing to offload tasks to the RSU may result in it being unable to process new tasks in a timely manner, thereby increasing task queue waiting time and prolonging processing delays, which in turn affects the quality and response speed of related services in the Internet of Vehicles system and fails to meet the vehicle's demand for task processing timeliness under different traffic conditions. A study proposed a method to partially offload perception tasks to the RSU and surrounding collaborative computing vehicles with computing capabilities. The strategy of selecting collaborative computing vehicles with the shortest distance is adopted for selecting collaborative computing vehicles. Although this method effectively utilizes the computing resources of vehicles and RSUs, it only considers the distance between vehicles and ignores the differences in computing power, which may not minimize the task completion time.

[0006] Furthermore, existing research mainly focuses on improving sensing performance or accelerating the processing of sensing tasks generated by cooperative sensing, ignoring the internal relationship between cooperative sensing and sensing task processing, resulting in inefficient cooperation between sensing, communication and computing, and thus reducing resource allocation efficiency.

[0007] In summary, in terms of improving sensing performance, optimal-view sensing reduces resource consumption but lacks multi-vehicle view complementarity, global sensing improves accuracy but easily leads to data redundancy, and random sensing reduces redundancy but affects consistency; in terms of sensing task processing, MEC completely offloads to alleviate the computing bottleneck but may cause congestion and delay, fixed proportion offloading reduces remote dependence but resource allocation is uneven, and partial offloading methods improve computing utilization but do not consider device differences; overall, the sensing and computing are not fully coordinated, leading to the problem of decreased resource allocation efficiency.

[0008] For example, patent application number 202411163721.4, entitled "Optimization method for edge computing offloading decision and resource allocation in Internet of Vehicles", proposes an optimization method based on partial task offloading, jointly considers task offloading ratio, bandwidth allocation and edge computing resource allocation, to minimize system completion time, and uses an improved genetic algorithm to solve the optimization scheme. For example, patent application number 202411532260.3, entitled "Generation method and device for heterogeneous group cooperative sensing task allocation scheme", proposes a cooperative sensing task allocation method for heterogeneous agents, combining local observation and global state, and making action decisions and value evaluations based on a multi-agent reinforcement learning framework. However, both of them do not model the sensing task and communication resources together, ignore the dependency between sensing and communication, lead to inefficient resource utilization and increased task delay, and cannot achieve real communication-sensing fusion optimization.

[0009] Therefore, how to effectively optimize the cooperation of sensing, communication and computing resources, find the best balance between sensing performance and sensing data volume, and thus improve the overall resource allocation efficiency of the system, has become a problem to be solved in the current cooperative sensing and computing offloading research. SUMMARY

[0010] To solve the above problems, the purpose of the present application is to solve some of the problems in the prior art, or at least alleviate these problems.

[0011] A method for joint optimization of cooperative sensing and computing offloading in Internet of Vehicles, comprising the following steps:

[0012] A one-way three-lane cooperative sensing scenario covered by a Roadside Unit (RSU) is constructed, which includes a requesting vehicle, multiple cooperative sensing vehicles and a cooperative computing vehicle; the sensing vehicle has a rectangular area of interest, which is divided into several blocks;

[0013] Cooperative perception stage modeling: including single vehicle perception model, multi-vehicle cooperative perception model, block selection model of perception vehicle;

[0014] Cooperative computing stage modeling: including transmission model, time delay model and total delay model;

[0015] Based on the perception stage model and the computing stage model, a system optimization model is established to minimize the total task completion delay under the condition of meeting the perception performance constraints; the optimization problem is expressed as:

[0016]

[0017] s.t.C1:d i,k ≤R sen ,

[0018] C2:a i,k ∈{0,1},

[0019]

[0020] C4:ω i,i +ω i,rsu +ω i,j =1,

[0021] C5:d i,j ≤R com ,

[0022] C6:f j ≥f0,

[0023]

[0024] Wherein, a is the association matrix of perception vehicle and block, c is the pairing matrix of perception vehicle and cooperative computing vehicle, ω is the partial offloading allocation matrix, β is the bandwidth allocation matrix, λ is the RSU computing resource allocation matrix; T represents the total task completion time; d i,k represents the distance between perception vehicle i and block k; R sen represents the perception radius of perception vehicle; a i,k ∈{0,1} represents whether perception vehicle i perceives block k; c i,j ∈{0,1} represents whether perception vehicle i offloads tasks to cooperative computing vehicle j; V s represents the set of perception vehicles; ω i,i represents the task allocation ratio of perception vehicle i to the tasks on the vehicle, ω i,rsu represents the task allocation ratio of perception vehicle i to the tasks on the RSU, ω i,j represents the task allocation ratio of perception vehicle i to the tasks on the cooperative computing vehicle j; d i,j represents the distance between perception vehicle i and cooperative computing vehicle j; Rcom represents the communication radius between vehicles; f j is the computation frequency on collaborative computing vehicle j; f0 represents the minimum computation frequency limit that the collaborative computing vehicle must meet; β i represents the bandwidth resource allocation coefficient allocated by the system to the perception vehicle i; i is the computing resource allocation coefficient of the RSU used to process the partially unloaded tasks from the sensing vehicle i;

[0025] Decouple the optimization problem into the collaborative perception phase optimization problem 1 and the collaborative calculation phase optimization problem 2;

[0026] A block selection algorithm for perception vehicles is designed to solve optimization problem 1, and a DDPG-HA algorithm is designed to solve optimization problem 2. Finally, the perception solution and perception task processing solution are obtained.

[0027] Furthermore, the collaborative perception stage is modeled by first defining the reliability of the perception results as a function of distance, then deriving the occupancy probability of a single-vehicle perception block based on this function, and further introducing the occupancy probability of multiple-vehicle perception blocks, thereby constructing a spatial-quality adaptively coupled perception control architecture. Finally, combined with the block selection algorithm of the perception vehicle, it is clear which blocks in the area of ​​interest each perception vehicle should perceive, and then the amount of tasks formed by each perception vehicle is determined.

[0028] The collaborative sensing stage modeling specifically includes the following steps:

[0029] S21: Use λ i,k represents the reliability of the perception block k of the sensor of vehicle i, and defines the reliability formula of the perception result as a function related to the distance:

[0030]

[0031] Among them, α represents the distance attenuation factor, d i,k represents the distance between the sensing vehicle i and block k, R sen Indicates the sensing radius of the vehicle sensor;

[0032] S22: Given the sensor output, the probability formula for the occupancy of a single vehicle perception block at detection block k for perception vehicle i is obtained from the above perception result reliability formula:

[0033]

[0034] Among them, S i,k ∈{S + ,S - , S0} is the detection result of the sensor of vehicle i in block k within the RoI; S + Indicates that there is an obstacle, S -Indicates that there is no obstacle, S0 indicates unknown; whether the block is occupied by an obstacle is determined by G i,k ∈{S + ,S -} is given, when S i,k =G i,k This means that the perception result is correct;

[0035] S23: The occupancy probability formula of the multi-vehicle perception block is introduced as follows:

[0036]

[0037] Among them, a i,k ∈{0,1} is the decision variable assigned to the perception task of the requesting vehicle; when a i,k =1 indicates that the perception of vehicle i is the perception block k, a i,k = 0, indicating that the sensing vehicle i does not sense block k;

[0038] S24: Given the occupancy probability formulas for single-vehicle and multi-vehicle perception blocks, the perceptual quality of block k is defined as:

[0039] Q k =|2o k -1|

[0040] S25: The relationship between the required perceived quality threshold and the decrease in the distance between the requesting vehicle and the block is expressed by the following expression:

[0041]

[0042] Among them, x need Indicates half the length of the requesting vehicle’s region of interest (RoI), d 0,k represents the distance between the requesting vehicle V0 and block k, θ 0,k represents the angle between the line vector from the requesting vehicle V0 to block k and the forward direction of the vehicle V0;

[0043] S26: Prioritize vehicles with higher perception quality for perception. Based on the space-quality adaptive coupled perception control architecture, the block selection algorithm of the perception vehicle is finally combined to determine the task load for each perception vehicle:

[0044]

[0045] Among them, l represents the amount of data generated by the car sensing a block, a i,k Represents the decision variable for assigning the perception task to perception vehicle i.

[0046] Furthermore, the collaborative computing stage is modeled by first establishing a transmission model to define the transmission rate when the perception vehicle unloads tasks to the RSU and the collaborative computing vehicle; then a delay model is constructed, including the local task processing delay of the perception vehicle, the processing delay of unloading tasks to the RSU, and the processing delay of unloading tasks to the collaborative computing vehicle; and then by determining the maximum delay of local processing, RSU processing, and collaborative computing vehicle processing, the completion time of a single perception vehicle task can be obtained, and the overall completion time of all perception vehicle tasks can be further derived.

[0047] The collaborative computing stage modeling specifically includes the following steps:

[0048] S31: Constructing the transmission model:

[0049] S311: The transmission rate when sensing vehicle i offloads part of the task to the RSU is:

[0050]

[0051] where β i,rsu represents the bandwidth coefficient allocated by the system for the transmission task of the sensing vehicle i to the RSU; B represents the total bandwidth of the system; p i,rsu Indicates the transmit power; h i,rsu represents the channel gain; d i,rsu represents the distance between sensing vehicle i and RSU; α represents the path loss coefficient; represents the power of additive white Gaussian noise;

[0052] S312: The transmission rate when sensing vehicle i offloads part of the task to its paired collaborative computing vehicle j is:

[0053]

[0054] Among them, β i,j represents the bandwidth coefficient allocated by the system when the perception vehicle i transmits the task to its paired collaborative computing vehicle j; p i,j Indicates the transmit power; h i,j represents the channel gain; d i,j Represents the distance between perception car i and collaborative computing car j; it is necessary to ensure that the perception car and collaborative computing car must be within the communication range d i,j ≤R com , where R com Represents the communication radius of the vehicle; for each task on the sensing vehicle, two types of bandwidth resources, V2V and V2I, need to be allocated: β i =β i,rsu +β i,j In addition, the transmission of all perception tasks must meet the constraints of the system bandwidth: represents the power of additive white Gaussian noise;

[0055] S32: Build a delay model:

[0056] S321: The local task processing delay of perception vehicle i is:

[0057]

[0058] Among them, ω i,i represents the ratio of tasks assigned by perception vehicle i to the vehicle itself; D i represents the amount of tasks formed on the perception vehicle i; S represents the number of CPU revolutions required to calculate 1 bit of the task, and f i is the calculation frequency on the sensing vehicle i;

[0059] S322: The task processing delay of sensing vehicle i partially offloading tasks to RSU is:

[0060]

[0061] Among them, ω i,rsu represents the ratio of tasks assigned by sensing vehicle i to RSU, λ i is the computing resource allocation coefficient of RSU used to process the partially unloaded tasks from sensing vehicle i; f rsu Indicates the calculation frequency on RSU;

[0062] S323: The task processing delay of partially offloading the task from the perception vehicle i to the collaborative computing vehicle j is:

[0063]

[0064] Among them, ω i,j represents the ratio of tasks assigned by perception vehicle i to collaborative computing vehicle j, and f j is the computation frequency on the collaborative computing vehicle j; the perception task allocation ratio is defined to be in the range of [0,1], and the sum of the task allocation ratios is 1: ω i,i ,ω i,rsu ,ω i,j ∈[0,1] and ω i,i +ω i,rsu +ω i,j =1;ω i,i represents the ratio coefficient of the task assigned by the perception vehicle i to the vehicle itself; ω i,rsu represents the distribution ratio coefficient of the task unloading from sensing vehicle i to RSU; ω i,j represents the task allocation ratio coefficient of the perception vehicle i to the collaborative computing vehicle j; the collaborative computing vehicle has sufficient computing power f j ≥f0; where f0 represents the minimum computing frequency requirement that the collaborative computing vehicle needs to meet;

[0065] S33: Constructing the total delay model:

[0066] For each perception vehicle, the entire task is completed only when the execution is completed locally, at the RSU, and on other collaborative computing vehicles:

[0067] T i =max{T i,i ,T i,rsu ,T i,j}

[0068] Only when all the perception car tasks are completed can the entire assisted perception task be completed:

[0069]

[0070] The optimization problem 1 (perception task allocation problem) in the collaborative perception stage is expressed as:

[0071]

[0072] stC 1:d i,k ≤R sen ,

[0073] C 2:a i,k ∈{0,1}.

[0074] The optimization problem 2 (perception task processing problem) in the collaborative computing phase is expressed as:

[0075]

[0076] C4:ω i,i +ω i,rsu +ω i,j =1,

[0077] C5:d i,j ≤R com ,

[0078] C6:f j ≥f0,

[0079]

[0080] Furthermore, a block selection algorithm for perception vehicles is designed to solve optimization problem 1. For each block in the area of ​​interest of the requesting vehicle, an adaptive collaborative perception method based on perception quality priority sorting is used. The collaborative perception unit is iteratively formed through a successive optimal vehicle selection mechanism until the system's preset perception threshold constraints are met to obtain a perception solution.

[0081] Designing the DDPG-HA algorithm to solve optimization problem 2 includes the following steps:

[0082] The computation offloading optimization problem is transformed into a Markov decision process (MDP), whose state, action, and reward functions are described as follows:

[0083] State Space definition is the state variable observed at time t, including the task offloading decision and task transmission rate. The state of the system can be defined as: s(t) = [N(t), R(t)];

[0084] Action Space definition At time t, the action performed by the agent following the action selection strategy consists of three parts: the allocation of perception tasks, the allocation of bandwidth resources, and the allocation of RSU computing resources. The action characteristics of the system can be expressed as: a1(t) = [δ(t), β(t), λ(t)];

[0085] By using the DDPG (Deep Deterministic Policy Gradient) algorithm to solve the task and resource allocation problem and constructing a bipartite graph of the task processing time between the perception vehicle and the collaborative computing vehicle, the optimal range of the maximum task completion time is determined.

[0086] A binary search is performed between the minimum and maximum task completion times. For each intermediate value mid, the Hungarian Algorithm (HA) is used to determine whether an effective task allocation scheme can be found given the maximum task completion time mid. Finally, the task offloading decision a2(t) = [c(t)] can be found, and all actions are a(t) = [c(t), ω(t), β(t), λ(t)].

[0087] Reward Function For the current state s(t), each time an action a(t) is executed, a reward r(t) is obtained. The inverse of the delay in completing all the sensing tasks is used as the reward: r(t) = 1 / T.

[0088] Introducing the Prioritized Experience Replay (PER) mechanism, which calculates the Temporal-Difference Error (TD) of each sample at each sampling, giving higher sampling priority to samples with larger TD errors.

[0089] After multiple rounds of training, the cumulative reward value converges, and the optimal action and task completion delay can be obtained.

[0090] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for joint optimization of collaborative perception and computation offloading in a vehicle network.

[0091] The present invention has the following beneficial effects:

[0092] 1. Joint Optimization of Collaborative Perception and Computation Offloading: Existing research ignores the inherent connection between collaborative perception and perception task processing, leading to inefficient collaboration among sensing, communication, and computing, and reduced resource allocation efficiency. This paper proposes a method to effectively connect the collaborative perception and collaborative computing stages. In the collaborative perception stage, by rationally selecting vehicle perception areas, perception quality is improved while reducing redundancy. In the collaborative computing stage, by jointly optimizing the offloading location and ratio of perception data, system bandwidth resources, and RSU computing resources, the completion time of all perception vehicle perception tasks is minimized while meeting perception quality constraints.

[0093] 2. Algorithmic Innovation and Performance Improvement: This paper combines DDPG with the Hungarian algorithm and adds prioritized experience replay, effectively enhancing the algorithm's exploration capabilities and optimizing the balance between exploration and exploitation, thereby significantly improving decision-making efficiency. Furthermore, compared to traditional DDPG, the proposed DDPG-HA algorithm demonstrates greater robustness and higher learning efficiency, accelerating convergence while increasing rewards.

[0094] This application combines the improvements in the above two aspects. First, based on the division of the area of ​​interest of the requesting vehicle, the spatial position and perception capability of the cooperative vehicle are comprehensively considered to dynamically optimize the perception vehicle selection strategy, thereby improving the overall perception coverage and reducing redundant data. Subsequently, based on factors such as channel status, communication distance and computing power, an offloading node screening mechanism and a joint allocation scheme for task resources are designed to construct an integrated optimization model for perception, transmission and computing. By introducing reinforcement learning and the Hungarian algorithm for collaborative solution, the redundancy of perception data is effectively reduced while the perception quality meets the requirements, the surrounding idle resources are fully utilized and the task completion delay is significantly shortened. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 This is a scene diagram of the Internet of Vehicles collaborative perception and collaborative computing system of the present invention;

[0096] Figure 2 is a flow chart of the present invention;

[0097] Figure 3 This is a comparison chart of the perception results of the collaborative perception solution proposed in the present invention and other perception solutions;

[0098] Figure 4This is a comparison chart of the amount of perception data of the collaborative perception solution proposed in this invention and other perception solutions;

[0099] Figure 5 This is a performance comparison chart of the DDPG-HA algorithm proposed in this invention and the traditional DDPG algorithm;

[0100] Figure 6 This is a comparison chart of the task completion time of this solution under different bandwidths and other solutions. DETAILED DESCRIPTION

[0101] The present invention will be further described below in conjunction with the accompanying drawings. The embodiments of the present invention are only used to illustrate the present invention and are not intended to limit the present invention. Without departing from the technical concept of the present invention, various substitutions and modifications can be made based on common technical knowledge and customary means in the field, and all should be included in the scope of the present invention.

[0102] The present invention aims to solve the problems of limited perception range, redundant perception data, and lack of a unified and efficient mechanism for collaborative optimization of perception and computing resources in collaborative perception systems in the Internet of Vehicles.

[0103] To solve the above problems, the present invention provides a method for the joint optimization of collaborative perception and computational offloading in the Internet of Vehicles. First, by balancing the perception performance and data volume in each perception area, the selection of perception areas by perception vehicles is optimized to ensure high-value perception information and reduce redundant data transmission. Then, for the real-time processing of perception tasks, the present invention proposes a new vehicle collaborative computing strategy, which combines channel conditions, vehicle computing capabilities and communication distance to optimize the selection of offloading nodes, ensure timely processing of perception tasks, and reduce the burden of network transmission and computational processing. Finally, a corresponding optimization problem is constructed and solved by a corresponding optimization algorithm. Under the premise of ensuring high-quality perception information, the offloading nodes are reasonably selected, the allocation of perception tasks is optimized, and communication and computing resources are effectively configured, so as to minimize the processing completion time of multiple perception tasks, thereby improving the overall task processing efficiency of the system. To achieve the above technical objectives, the present invention provides a two-stage method for the joint optimization of collaborative perception and computational offloading in the Internet of Vehicles. The specific steps are as follows.

[0104] S1: Building a collaborative perception scenario

[0105] A one-way three-lane collaborative perception scenario with RSU coverage is constructed, which includes a requesting vehicle, multiple collaborative perception vehicles, and a collaborative computing vehicle. The perception vehicle has a rectangular area of ​​interest, which is divided into several blocks.

[0106] The system model structure is as follows Figure 1As shown in Figure 2, we consider a one-way three-lane cooperative sensing scenario under the coverage of an RSU, where there are many vehicles equipped with wireless communication equipment and certain computing capabilities on the lanes. It is divided into request vehicles, collaborative perception vehicles and collaborative computing vehicles. The request vehicles and collaborative perception vehicles are collectively referred to as perception vehicles. The perception vehicle and collaborative computing vehicle are defined as in, and gather The first one represents the requesting vehicle, and the remaining ones are collaborative sensing vehicles. They are all equipped with sensors within a fixed circular range of radius R. The requesting vehicle needs to obtain environmental information for its RoI. The RoI is divided into blocks of equal size to minimize redundancy. Λ = {1, 2, .., K} represents the set of all blocks within the requesting vehicle's range of interest. Quadtree compression technology is used to compress the sensing data into independent block messages, each with the same bit size. Each block is assigned to a sensing vehicle for sensing according to certain rules. Finally, the sensing task formed by each sensing vehicle can be offloaded locally, to the RSU via vehicle-to-infrastructure (V2I) communication, or to a collaborative computing vehicle with sufficient computing power via vehicle-to-vehicle (V2V) communication.

[0107] S2: Collaborative Perception Stage Modeling

[0108] It includes single-vehicle perception model, multi-vehicle collaborative perception model, and block selection model of perception vehicle.

[0109] During the collaborative perception phase, because sensor perception reliability is affected by the distance between the sensor and the perceived block, the perception reliability is defined as a function of distance. Based on this function, the occupancy probability of a single-vehicle perception block is derived. Furthermore, the occupancy probability of multiple-vehicle perception blocks is introduced to leverage measurements from different angles to improve perception accuracy and confidence. On this basis, a spatial-quality adaptively coupled perception control architecture is constructed to meet the requesting vehicle's need for higher perception quality of closer areas. Ultimately, through the perception vehicle's block selection algorithm, vehicles with higher perception quality are prioritized to perceive blocks within the requested area. The perception control architecture then determines which vehicles are responsible for perceiving each block, clarifying which blocks within the region of interest each perception vehicle should perceive, and ultimately determining the workload assigned to each perception vehicle.

[0110] The details of the collaborative sensing stage model are as follows:

[0111] S21: Since the sensor is not completely reliable, use λ i,k represents the reliability of the perception block k of the sensor of vehicle i, and defines the reliability formula of the perception result as a function related to the distance:

[0112]

[0113] Among them, α represents the distance attenuation factor, d i,k represents the distance between the sensing vehicle i and block k, R sen Indicates the sensing radius of the vehicle sensor.

[0114] S22: Given the sensor output, the probability formula for the occupancy of a single vehicle perception block at detection block k for perception vehicle i is obtained from the above perception result reliability formula:

[0115]

[0116] Among them, S i,k ∈{S + ,S - , S0} is the detection result of the sensor of vehicle i in block k within the RoI, where S + Indicates an obstacle, S - Indicates no obstacles, S0 indicates unknown. Whether the block is occupied by obstacles is determined by G i,k ∈{S + ,S -} is given, when S i,k =G i,k This indicates that the perception result is correct.

[0117] S23: Since measuring the same region of interest from multiple angles can improve perception accuracy and confidence, the occupancy probability formula for the multi-vehicle perception block is introduced as follows:

[0118]

[0119] Among them, a i,k ∈{0,1} is the decision variable assigned to the perception task of the requesting vehicle; when a i,k =1 indicates that the perception of vehicle i is the perception block k, a i,k =0, it means that the sensing vehicle i does not perceive block k.

[0120] S24: Given the occupancy probability formulas for single-vehicle and multi-vehicle perception blocks, the perceptual quality of block k is defined as:

[0121] Q k =|2o k -1|

[0122] As the occupancy probability of block k approaches 1 / 2, the information about whether there is an obstacle on the block becomes more uncertain, and the perceived quality Q k The lower the value of .

[0123] S25: From the perspective of the requesting vehicle, closer blocks have higher risk and shorter response time windows, which directly impacts the vehicle's immediate safety and decision-making. Therefore, the required sensing quality increases as the distance to the block decreases. This relationship is expressed as the following expression:

[0124]

[0125] Among them, x need Indicates half of the length of RoI, d 0,k represents the distance between the requesting vehicle V0 and block k, θ 0,k represents the angle between the line vector from the requesting vehicle V0 to block k and the forward direction of the vehicle V0;

[0126] S26: For each perception block within the requesting vehicle's region of interest, while meeting perception constraints and minimizing perception data redundancy, vehicles with higher perception quality are prioritized for perception. Based on the spatial-quality adaptively coupled perception control architecture, the specific vehicles responsible for perceiving the block are further determined. This clarifies which blocks within the region of interest each perception vehicle should perceive, thereby determining the workload of each perception vehicle:

[0127]

[0128] Among them, l represents the amount of data generated by the car sensing a block, a i,k Represents the decision variable for assigning the perception task to perception vehicle i.

[0129] S3: Collaborative Computational Phase Modeling

[0130] Includes transmission model, delay model and total delay model.

[0131] During the collaborative computing phase, each perception vehicle's perception tasks can be executed locally, offloaded to the RSU, and offloaded to its associated collaborative computing vehicles. First, a transmission model is established to define the transmission rate when the perception vehicle offloads tasks to the RSU and collaborative computing vehicles. Subsequently, a latency model is constructed, including the perception vehicle's local task processing latency, the processing latency of offloading tasks to the RSU, and the processing latency of offloading tasks to the collaborative computing vehicle. By determining the maximum latency of local processing, RSU processing, and collaborative computing vehicle processing, the completion time of a single perception vehicle task can be derived, and further, the overall completion time of all perception vehicle tasks can be derived.

[0132] Since the processing of the perception task can be performed locally, offloaded to the drive test unit, or offloaded to its associated collaborative computing vehicle, the details of the collaborative computing stage model are as follows:

[0133] S31: Constructing the transmission model:

[0134] The transmission model is that the perception task is first formed on the requesting vehicle and multiple collaborative perception vehicles, and then the perception task is processed by partial offloading: if it is offloaded to other locations, it is transmitted first and then calculated, so there is a transmission model.

[0135] S311: The transmission rate when sensing vehicle i offloads part of the task to the RSU is:

[0136]

[0137] where β i,rsu represents the bandwidth coefficient allocated by the system for the transmission task of the sensing vehicle i to the RSU; B represents the total bandwidth of the system; p i,rsu Indicates the transmit power; h i,rsu represents the channel gain; d i,rsu represents the distance between sensing vehicle i and RSU; α represents the path loss coefficient; represents the power of additive white Gaussian noise.

[0138] S312: The transmission rate when sensing vehicle i offloads part of the task to its paired collaborative computing vehicle j is:

[0139]

[0140] Among them, β i,j represents the bandwidth coefficient allocated by the system when the perception vehicle i transmits the task to its paired collaborative computing vehicle j; p i,j Indicates the transmit power; h i,j represents the channel gain; d i,j Represents the distance between perception car i and collaborative computing car j; it is necessary to ensure that the perception car and collaborative computing car must be within the communication range d i,j ≤R com , where R com Represents the communication radius of the vehicle; for each task on the sensing vehicle, two types of bandwidth resources, V2V and V2I, need to be allocated: β i =β i,rsu +β i,j In addition, the transmission of all perception tasks must meet the constraints of the system bandwidth: represents the power of additive white Gaussian noise.

[0141] S32: Build a delay model:

[0142] S321: The local task processing delay of perception vehicle i is:

[0143]

[0144] Among them, ω i,irepresents the ratio of tasks assigned by perception vehicle i to the vehicle itself, D i represents the amount of tasks formed on the perception vehicle i, S represents the number of CPU revolutions required to calculate 1 bit of the task, and f i is the calculation frequency on the sensing car i.

[0145] S322: The task processing delay of sensing vehicle i partially offloading tasks to RSU is:

[0146]

[0147] Among them, ω i,rsu represents the ratio of tasks assigned by sensing vehicle i to RSU, λ i is the computing resource allocation coefficient of the RSU used to process the partially unloaded tasks from the sensing vehicle i. It should be noted that the amount of sensing information after processing and the delay of sensing information return are relatively small, so they can be ignored. rsu Indicates the calculation frequency on RSU.

[0148] S323: The task processing delay of partially offloading the task from the perception vehicle i to the collaborative computing vehicle j is:

[0149]

[0150] Among them, ω i,j represents the ratio of tasks assigned by perception vehicle i to collaborative computing vehicle j, and f j is the computation frequency on the collaborative computing vehicle j; the perception task allocation ratio is defined to be in the range of [0,1], and the sum of the task allocation ratios is 1: ω i,i ,ω i,rsu ,ω i,j ∈[0,1] and ω i,i +ω i,rsu +ω i,j =1;ω i,i represents the ratio coefficient of the task assigned by the perception vehicle i to the vehicle itself; ω i,rsu represents the distribution ratio coefficient of the task unloading from sensing vehicle i to RSU; ω i,j represents the task allocation ratio coefficient of the perception vehicle i to the collaborative computing vehicle j; the collaborative computing vehicle has sufficient computing power f j ≥f0; where f0 represents the minimum computing frequency requirement that the collaborative computing vehicle needs to meet.

[0151] S33: Constructing the total delay model:

[0152] For each perception vehicle, the entire task is completed only when the execution is completed locally, at the RSU, and on other collaborative computing vehicles:

[0153] T i =max{T i,i ,T i,rsu ,T i,j}

[0154] Only when all the perception car tasks are completed can the entire assisted perception task be completed:

[0155]

[0156] S4: Establishing a system optimization model

[0157] Based on the perception stage model and the calculation stage model, a system optimization model is established to minimize the total task completion delay while satisfying the perception performance constraint.

[0158] Based on the modeling of the above-mentioned collaborative perception and collaborative computing stages, an optimization problem is further defined. Under the constraints of perception block location, perception distance, task division, communication distance, computing power of collaborative computing vehicles, bandwidth resources, and RSU computing resources, the completion time of all perception tasks is minimized by jointly optimizing the selection of the responsible perception vehicle for each perception block, the offloading location and offloading ratio of perception data, the system's bandwidth resources, and the configuration of RSU computing resources.

[0159] The optimization problem is formulated as:

[0160]

[0161] stC1:d i,k ≤R sen ,

[0162] C2:a i,k ∈{0,1},

[0163]

[0164] C4:ω i,i +ω i,rsu +ω i,j =1,

[0165] C5:d i,j ≤R com ,

[0166] C6:f j ≥f0,

[0167]

[0168] Where a is the association matrix between perception vehicles and blocks, c is the pairing matrix between perception vehicles and collaborative computing vehicles, ω is the partial offloading allocation matrix, β is the bandwidth allocation matrix, and λ is the RSU computing resource allocation matrix; T represents the completion time of the perception tasks of all perception vehicles (i.e., the total task completion time); d i,k represents the distance between the sensing vehicle i and block k; R sen Indicates the perception radius of the perception vehicle; a i,k ∈{0,1} is the decision variable for the perception task of the requesting vehicle, indicating whether the perception vehicle i perceives block k; c i,j ∈{0,1} indicates whether the sensing vehicle i unloads the task to the collaborative computing vehicle j; V s represents the set of sensing cars; ω i,i represents the ratio of tasks assigned by perception vehicle i to the vehicle itself, ω i,rsu represents the ratio of tasks assigned by sensing vehicle i to RSU, ω i,j represents the ratio of tasks assigned by perception vehicle i to collaborative computing vehicle j; f j is the computation frequency on collaborative computing vehicle j, λ i is the computational resource allocation coefficient of the RSU used to process the partially unloaded tasks from the sensing vehicle i. i,j represents the distance between perception vehicle i and collaborative computing vehicle j; R com represents the communication radius between vehicles; f0 represents the minimum computing frequency limit that the collaborative computing vehicle must meet; β i It represents the bandwidth resource allocation coefficient allocated by the system to the perception vehicle i.

[0169] C1 ensures that any block perceived by the sensing vehicle is within its sensing range; C2 defines the sensing decision, which indicates whether the sensing vehicle i senses block k; C3 is a binary variable for the selection of collaborative computing vehicles, and ensures that each collaborative computing vehicle provides task offloading services to at most one sensing vehicle; C4 ensures that the sum of the task allocation ratios of each sensing vehicle is 1; C5 ensures that the collaborative computing vehicle is within the communication range of the corresponding sensing vehicle; C6 ensures that the selected collaborative computing vehicle meets the minimum required computing power requirements; C7 constrains the total available bandwidth resources; finally, C8 implements the computing power constraints of the roadside unit. S5: The optimization problem is decoupled into collaborative perception phase optimization problem 1 and collaborative computing phase optimization problem 2.

[0170] Because the optimization problem is a non-convex mixed-integer nonlinear programming (MINLP) problem, direct solution is extremely difficult. To this end, we propose a two-stage decomposition method that decomposes the original problem into two sub-modules with clear physical meaning by decoupling the perception-communication coupling constraints: the selection of perception blocks by each perception vehicle and the processing of perception tasks on each perception vehicle. This method can optimize and adjust the different characteristics and constraints of the problem to better meet the specific requirements of the perception task and effectively address the limitations of communication and computing resources. Specifically, in the first stage, the collaborative perception stage, we first calculate the number of perception tasks formed on each perception vehicle based on the perception quality constraints; in the second stage, the collaborative computation stage, we focus on solving the adaptive transmission and computation problems, aiming to minimize the task execution delay. The details of each stage will be introduced in the subsequent sections.

[0171] The perception task allocation problem in the collaborative perception stage is stated as follows:

[0172] The collaborative perception stage ensures that each perception vehicle can meet the perception threshold requirement with minimal perception redundancy. The optimization problem 1 is expressed as:

[0173]

[0174] stC 1:d i,k ≤R sen ,

[0175] C 2:a i,k ∈{0,1}.

[0176] The optimization problem 2 in the collaborative computing phase is expressed as follows:

[0177] The collaborative computing phase ensures that each perception vehicle minimizes the delay in completing all tasks under the constraints of task allocation, communication distance, computing power of the collaborative computing vehicle, bandwidth allocation, and RSU computing resource allocation. Based on the data from the collaborative perception phase, it first determines the perception area of ​​each perception vehicle and the number of perception tasks it needs to handle. On this basis, the collaborative computing phase aims to optimize the processing of perception tasks by rationally selecting offloading nodes, optimizing the allocation of multi-vehicle perception tasks, and efficiently allocating communication resources and bandwidth resources to minimize the maximum task completion time. Optimization Problem 2 is expressed as:

[0178]

[0179] C4:ω i,i +ω i,rsu +ω i,j =1,

[0180] C5:d i,j ≤R com ,

[0181] C6:f j ≥f0,

[0182]

[0183] S6: Solving optimization problems

[0184] To solve optimization problem 1, a block selection algorithm for sensing vehicles was designed; to solve optimization problem 2, a DDPG-HA algorithm was designed. Finally, a sensing solution and a sensing task processing solution were obtained.

[0185] S61: Solving Optimization Problem 1

[0186] In order to enable each perception vehicle to meet the perception threshold requirements with minimal perception redundancy, for each block in the requesting vehicle's area of ​​interest, the present invention proposes an adaptive collaborative perception method based on perception quality priority sorting, and iteratively establishes collaborative perception units through a successive optimal vehicle selection mechanism until the system's preset perception threshold constraints are met.

[0187] To this end, we designed Algorithm 1 to select blocks within the requesting vehicle's area of ​​interest:

[0188]

[0189]

[0190] To demonstrate the effectiveness of the proposed perception scheme, a comprehensive evaluation of the proposed method was conducted. The effectiveness and advantages of the proposed block selection algorithm for vehicle perception in minimizing the redundancy of perception data while improving perception performance were verified. The following comparison scheme was set up:

[0191] Comparison scheme 1: Independent perception by the own vehicle: The blocks within the region of interest are only perceived by the requesting vehicle;

[0192] Comparison Scenario 2: Perception by the vehicle with the best perception quality: The blocks within the region of interest are only perceived by the vehicle with the best perception quality;

[0193] Comparison plan 3: All vehicle perception: All perception vehicles participate in perception.

[0194] Through comparative simulation with the above solutions, this example gives specific simulation and conclusions.

[0195] like Figure 3A comparison of the perception performance of different perception schemes was presented, revealing that any scheme involving collaborative perception vehicles outperformed independent perception schemes, demonstrating that collaborative perception can effectively improve perception quality. Furthermore, when the vehicle with the best perception quality is selected for perception, and its perception results meet the threshold of the perception quality benchmark, that vehicle is selected for perception; however, if the perception of the vehicle with the best perception quality fails to meet the threshold, a multi-vehicle joint perception strategy is required based on the established scheme.

[0196] Then, by comparing the proposed perception scheme with all vehicle perception schemes, we can find that Figure 3 There is almost no difference in the perception results of the five blocks. Figure 4 Analysis shows that the proposed sensing solution generates significantly less data than all other vehicle sensing solutions. This demonstrates that the proposed sensing solution maintains comparable perception quality to all other vehicle sensing solutions while effectively reducing data usage. While maintaining perception quality, it also optimizes data processing efficiency and resource utilization.

[0197] After the collaborative perception stage is completed, the selection of the perception area by each perception vehicle can be determined, thereby clarifying the number of perception tasks that each perception vehicle needs to handle.

[0198] S62: Solve optimization problem 2:

[0199] Since optimization problem 2 contains binary offloading decision variables, and both task allocation decisions and resource allocation decisions are continuous variables, it has the following two key characteristics: first, task allocation, bandwidth resource allocation, and RSU computing resource allocation are all affected by the offloading decision and change with it; second, when the computing resource and communication resource allocation strategies are unknown, it is difficult to directly evaluate the optimality of the offloading decision.

[0200] Furthermore, considering that this problem is a mixed continuous and discrete problem, and the Deep Deterministic Policy Gradient (DDPG) algorithm is primarily designed to handle continuous actions, while a simple continuous action discretization method can solve both continuous and discrete action problems, this also prevents the DDPG algorithm from fully leveraging its advantages. The Hungarian Algorithm (HA) is widely used in matching problems to minimize latency, minimizing total execution time by finding the optimal matching solution. Therefore, the present invention proposes the DDPG-HA algorithm, which is specifically as follows: based on the DDPG algorithm, tasks and resources are allocated. By constructing a bipartite graph of the task processing time between the perception vehicle and the collaborative computing vehicle, it is used to solve the problem of selecting the perception vehicle to offload tasks to the collaborative computing node, and cleverly transforms this problem into a matching problem. The Hungarian algorithm can efficiently solve the matching results and, due to its superiority in minimizing total execution time, is widely used in such optimization problems.

[0201] However, the core goal of this invention is to minimize the maximum task completion time, that is, to ensure that the maximum value of the completion time of all tasks is minimized. Directly applying the Hungarian algorithm may cause some tasks to complete excessively long tasks, while others complete prematurely, thus failing to achieve a globally optimal solution. Therefore, we introduce a bisection algorithm that gradually adjusts the upper limit of task completion time, uses DDPG to process continuous actions, and combines the Hungarian algorithm with the bisection algorithm to process discrete actions, thereby improving the overall efficiency of perception task processing.

[0202] The details are as follows:

[0203] The computation offloading optimization problem is transformed into a Markov decision process (MDP), whose state, action, and reward functions are described as follows:

[0204] State Space definition are the state variables observed at time t, including task offloading decisions and task transmission rates. The state of the system can be defined as: s(t) = [N(t), R(t)].

[0205] Action Space definition At time t, the action performed by the agent following the action selection strategy consists of three parts: the allocation of perception tasks, the allocation of bandwidth resources, and the allocation of RSU computing resources. The action characteristics of the system can be expressed as: a1(t) = [δ(t), β(t), λ(t)].

[0206] The DDPG algorithm is used to solve the task and resource allocation problem and construct a bipartite graph of the task processing times between the perception vehicle and the collaborative computing vehicle. This determines the range within which the maximum task completion time is optimized. Next, a binary search is performed between the minimum and maximum task completion times. For each intermediate value mid, the Hungarian algorithm is used to determine whether a valid task allocation solution can be found given a maximum task completion time of mid. To achieve this goal, the Hungarian algorithm first constructs a cost matrix. For tasks with a completion time exceeding mid, the cost is set to a very large value, indicating that the task cannot be allocated. The Hungarian algorithm then searches for the optimal task allocation solution within this modified cost matrix. If the completion time of all tasks does not exceed mid, the current mid value is considered feasible. By continuously adjusting the binary search range and combining the Hungarian algorithm with feasibility testing for each mid value, the task offloading decision a2(t) = [c(t)] is ultimately found, with all actions a(t) = [c(t), ω(t), β(t), λ(t)].

[0207] Reward Function For the current state s(t), each time an action a(t) is executed, a reward r(t) is obtained. The inverse of the latency of all perceived tasks is used as the reward: r(t) = 1 / T.

[0208] In addition, DDPG extracts samples uniformly from the experience pool, while the relatively optimal action space in this problem is very narrow, which means that many of the samples extracted are unimportant experiences. Therefore, the Prioritized Experience Replay (PER) mechanism is introduced. By calculating the TD error (Temporal-Difference Error, TD) of the sample at each sampling, samples with larger TD errors are given higher sampling priority, allowing the agent to learn more valuable experience and thus use samples more effectively.

[0209] The DDPG-HA algorithm described above can show better performance on this problem. The algorithm structure diagram is as follows:

[0210]

[0211] Initial parameters and the environment state are input, and the system state and accumulated reward are reset at each training round. At each step in each round, the agent outputs the corresponding perception task allocation, bandwidth resource allocation, and RSU computation allocation based on the current environment state s(t). It also generates a bipartite graph of the task completion time from the perception vehicle to the collaborative computation. The optimal vehicle pairing is then determined using a combination of the Hungarian algorithm and the bipartite algorithm. The combination of continuous and discrete actions is the overall action a(t). When the action is applied to the environment, the environment state shifts to s(t+1), and the corresponding environmental feedback, i.e., reward r(t), is obtained. The experience tuple (s(t), a(t), r(t), s(t+1)) is stored in a prioritized experience replay pool. During each learning step, the agent selects a batch of samples from the prioritized experience replay pool based on their experience values. Using an improved target policy smoothing regularization technique and target action value estimation method, the parameters of the estimation network are updated using gradient descent, and the target network is updated using soft updates. After multiple rounds of training, the accumulated reward converges, resulting in the optimal action and task completion delay.

[0212] The effectiveness and advantages of the DDPG-HA algorithm proposed in this paper in solving the problem of minimizing the delay in completing all tasks are verified, and the following comparison scheme is set up:

[0213] Comparison plan 1: Shortest distance matching plan: The perception vehicle and the collaborative computing vehicle are matched using the shortest distance algorithm.

[0214] Comparison plan 2: Stable matching plan: The perception vehicle and the collaborative computing vehicle are matched using a stable matching algorithm.

[0215] Comparison scheme 3: PER-DDPG algorithm: uses the DDPG algorithm improved by priority experience replay to optimize actions.

[0216] Through comparative simulation with the above solutions, this example gives specific simulation and conclusions.

[0217] Figure 5 The performance of the proposed DDPG-HA algorithm is compared with that of the traditional DDPG algorithm. Both the proposed task processing algorithm and the DDPG algorithm ultimately converge. In particular, the proposed algorithm demonstrates excellent learning efficiency and optimized final performance within the reinforcement learning framework. Compared to the traditional DDPG algorithm, the proposed DDPG-HA algorithm converges significantly faster and maintains a high level of reward throughout the training cycle.

[0218] Figure 6The task completion times of different schemes under different bandwidth conditions are presented. As bandwidth increases, data transmission capacity is significantly enhanced, and the task completion time of all schemes shows a downward trend. Notably, the downward trend gradually slows. This phenomenon can be attributed to the fact that task completion time is composed of two major components: transmission time and computation time. Under bandwidth-limited conditions, due to the low data rate, the impact of transmission time on task completion time is more significant. When the bandwidth reaches a higher level, the impact of reduced transmission time on overall task completion time becomes less significant. Even with further increases in bandwidth, the improvement in overall task completion time becomes relatively limited due to the larger absolute value of computation time compared to transmission time. When the bandwidth is 1.5 MHz, the task completion time of the proposed algorithm is 34.8% lower than that of the stable matching scheme. At 2 MHz, the performance of the proposed algorithm is comparable to that of the PER-DDPG algorithm, and at 2 MHz, it outperforms the stable matching scheme. These results demonstrate that, under the same conditions, the proposed algorithm can effectively utilize bandwidth.

[0219] In summary, the present invention uses perception performance and task completion delay as indicators, verifies the effectiveness of the method in simulation results, and demonstrates superior perception performance and the ability to process perception tasks.

[0220] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for joint optimization of collaborative perception and computation offloading in a vehicle network.

[0221] The present invention proposes a method for joint optimization of collaborative perception and computation offloading. First, by calculating the balance between perception performance and data volume in each perception area, the selection of perception vehicles is optimized to ensure high-value perception information and reduce redundant data transmission. Then, the selection of collaborative computing vehicles is optimized based on the perception task size, channel conditions and vehicle computing power to reduce the burden of network transmission and computation processing, improve the overall task processing efficiency of the system, and thus minimize the processing completion time of multiple perception tasks.

[0222] The present invention specifically includes:

[0223] In response to the problem that existing research ignores the intrinsic connection between collaborative perception and perception task processing, resulting in inefficient collaboration between sensing, communication and computing, and reduced resource allocation efficiency, the present invention proposes a method to effectively connect the collaborative perception and collaborative computing stages. In the collaborative perception stage, by reasonably selecting the vehicle perception area, the perception quality is improved while reducing redundancy, and the task volume of each perception vehicle is determined. In the collaborative computing stage, under the constraints of perception quality, perception block location, perception distance, task division, communication distance, collaborative computing vehicle computing power, bandwidth resources and RSU computing resources, by jointly optimizing the offloading location and proportion of perception data, system bandwidth resources and RSU computing resources, the completion time of all perception tasks of perception vehicles is minimized while meeting the perception quality constraints.

[0224] This paper combines DDPG with the Hungarian algorithm and adds prioritized experience replay, effectively enhancing the algorithm's exploration capabilities and optimizing the balance between exploration and exploitation, thereby significantly improving decision-making efficiency. Furthermore, compared to traditional DDPG, the proposed DDPG-HA algorithm exhibits greater robustness and higher learning efficiency, accelerating convergence while increasing rewards.

[0225] The present invention is applicable to multiple fields such as intelligent connected vehicles, autonomous driving, and vehicle networking, and has broad application prospects and market potential. By integrating the collaborative optimization strategy of perception and computing resources, it not only improves the vehicle's perception quality of the dynamic environment, but also significantly enhances the system's real-time processing capabilities for multiple tasks. This technology is particularly suitable for high-density traffic, complex urban road conditions, and smart highway scenarios, and can effectively improve the safety, reliability, and operational efficiency of the transportation system. At the same time, the present invention has a good engineering implementation foundation, is easy to integrate with existing intelligent transportation infrastructure, and has high industrial promotion value.

[0226] The conventional techniques and schemes not described in detail in the above embodiments are well known in the art and will not be described in detail here. The above embodiments and experimental examples describe the preferred embodiments of the present invention in detail, but the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, the technical scheme of the present invention can be subjected to various simple modifications, and these simple modifications all fall within the scope of protection of the present invention.

Claims

1. A method for joint optimization of collaborative perception and computation offloading in an Internet of Vehicles, characterized in that: The following steps are involved: A one-way three-lane cooperative perception scenario is constructed, covered by a Roadside Unit (RSU). This scenario includes a requesting vehicle, multiple cooperative perception vehicles, and a cooperative computing vehicle. The perception vehicle has a rectangular region of interest, which is divided into several blocks. Collaborative perception stage modeling: including single-vehicle perception model, multi-vehicle collaborative perception model, and perception vehicle block selection model; Collaborative computing stage modeling: including transmission model, delay model and total delay model; Based on the perception stage model and the computation stage model, a system optimization model is established to minimize the total task completion delay while satisfying the perception performance constraints. The optimization problem is expressed as: s.t.C1:d i,k ≤R sen , C2:a i,k ∈{0,1}, C3: C4:oh i,i +oh i,rsu +oh i,j =1, C5:d i,j ≤R com , C6:f j ≥f0, C7: C8: Where a is the association matrix between the perception vehicle and the block, c is the pairing matrix between the perception vehicle and the collaborative computing vehicle, ω is the partial offloading allocation matrix, β is the bandwidth allocation matrix, and λ is the RSU computing resource allocation matrix; T represents the total task completion time; d i,k represents the distance between the sensing vehicle i and block k; R sen Indicates the perception radius of the perception vehicle; a i,k ∈{0,1} indicates whether the sensing vehicle i senses block k; c i,j ∈{0,1} indicates whether the sensing vehicle i unloads the task to the collaborative computing vehicle j; V s represents the set of sensing cars; ω i,i represents the ratio of tasks assigned by perception vehicle i to the vehicle itself, ω i,rsu represents the ratio of tasks assigned by sensing vehicle i to RSU, ω i,j represents the ratio of tasks assigned by perception vehicle i to collaborative computing vehicle j; d i,j represents the distance between perception vehicle i and collaborative computing vehicle j; R com represents the communication radius between vehicles; f j is the computation frequency on collaborative computing vehicle j; f0 represents the minimum computation frequency limit that the collaborative computing vehicle must meet; β i represents the bandwidth resource allocation coefficient allocated by the system to the perception vehicle i; i is the computing resource allocation coefficient of the RSU used to process the partially unloaded tasks from the sensing vehicle i; Decouple the optimization problem into the collaborative perception phase optimization problem 1 and the collaborative calculation phase optimization problem 2; A block selection algorithm for perception vehicles is designed to solve optimization problem 1, and a DDPG-HA algorithm is designed to solve optimization problem 2. Finally, the perception solution and perception task processing solution are obtained.

2. The method for joint optimization of collaborative perception and computation offloading in the Internet of Vehicles according to claim 1, characterized in that: The collaborative perception stage modeling first defines the reliability of the perception result as a function of distance, then derives the occupancy probability of a single-vehicle perception block based on this function, and further introduces the occupancy probability of multiple-vehicle perception blocks, thereby constructing a spatial-quality adaptively coupled perception control architecture. Finally, combined with the block selection algorithm of the perception vehicle, it is clear which blocks in the area of ​​interest each perception vehicle should perceive, and then the task volume formed by each perception vehicle is determined.

3. The method for joint optimization of collaborative perception and computation offloading in the Internet of Vehicles according to claim 2, wherein the collaborative perception stage modeling specifically comprises the following steps: S21: Use λ i,k represents the reliability of the perception block k of the sensor of vehicle i, and defines the reliability formula of the perception result as a function related to the distance: Among them, α represents the distance attenuation factor, d i,k represents the distance between the sensing vehicle i and block k, R sen Indicates the sensing radius of the vehicle sensor; S22: Given the sensor output, the probability formula for the occupancy of a single vehicle perception block at detection block k for perception vehicle i is obtained from the above perception result reliability formula: Among them, S i,k ∈{S + ,S - , S0} is the detection result of the sensor of vehicle i in block k within the RoI; S + Indicates that there is an obstacle, S - Indicates that there is no obstacle, S0 indicates unknown; whether the block is occupied by an obstacle is determined by G i,k ∈{S + ,S - } is given, when S i,k =G i,k This means that the perception result is correct; S23: The occupancy probability formula of the multi-vehicle perception block is introduced as follows: Among them, a i,k ∈{0,1} is the decision variable assigned to the perception task of the requesting vehicle; when a i,k =1 indicates that the perception of vehicle i is the perception block k, a i,k = 0, indicating that the sensing vehicle i does not sense block k; S24: Given the occupancy probability formulas for single-vehicle and multi-vehicle perception blocks, the perceptual quality of block k is defined as: Q k =|2o k -1| S25: The relationship between the required perceived quality threshold and the decrease in the distance between the requesting vehicle and the block is expressed by the following expression: Among them, x need Indicates half the length of the requesting vehicle’s region of interest (RoI), d 0,k represents the distance between the requesting vehicle V0 and block k, θ 0,k represents the angle between the line vector from the requesting vehicle V0 to block k and the forward direction of the vehicle V0; S26: Prioritize vehicles with higher perception quality for perception. Based on the space-quality adaptive coupled perception control architecture, the block selection algorithm of the perception vehicle is finally combined to determine the task load for each perception vehicle: Among them, l represents the amount of data generated by the car sensing a block, a i,k Represents the decision variable for assigning the perception task to perception vehicle i.

4. The method for joint optimization of collaborative perception and computation offloading in the Internet of Vehicles according to claim 1, characterized in that: The collaborative computing stage modeling first establishes a transmission model to define the transmission rate when the perception vehicle unloads tasks to the RSU and the collaborative computing vehicle; A delay model is then constructed, including the local task processing delay of the perception vehicle, the processing delay of offloading tasks to the RSU, and the processing delay of offloading tasks to the collaborative computing vehicle. By determining the maximum delay of local processing, RSU processing, and collaborative computing vehicle processing, the completion time of a single perception vehicle task can be obtained, and the overall completion time of all perception vehicle tasks can be further derived.

5. The method for joint optimization of collaborative perception and computation offloading in the Internet of Vehicles according to claim 4, characterized in that: The collaborative computing stage modeling specifically includes the following steps: S31: Constructing a transmission model: S311: The transmission rate when sensing vehicle i offloads part of the task to the RSU is: where β i,rsu represents the bandwidth coefficient allocated by the system for the transmission task of the sensing vehicle i to the RSU; B represents the total bandwidth of the system; p i,rsu Indicates the transmit power; h i,rsu represents the channel gain; d i,rsu represents the distance between sensing vehicle i and RSU; α represents the path loss coefficient; represents the power of additive white Gaussian noise; S312: The transmission rate when sensing vehicle i offloads part of the task to its paired collaborative computing vehicle j is: Among them, β i,j represents the bandwidth coefficient allocated by the system when the perception vehicle i transmits the task to its paired collaborative computing vehicle j; p i,j Indicates the transmit power; h i,j represents the channel gain; d i,j Represents the distance between perception car i and collaborative computing car j; it is necessary to ensure that the perception car and collaborative computing car must be within the communication range d i,j ≤R com , where R com Represents the communication radius of the vehicle; for each task on the sensing vehicle, two types of bandwidth resources need to be allocated for it: Vehicle-to-Vehicle (V2V) communication and Vehicle-to-Infrastructure (V2I) communication: β i =β i,rsu +β i,j In addition, the transmission of all perception tasks must meet the constraints of the system bandwidth: represents the power of additive white Gaussian noise; S32: Build a delay model: S321: The local task processing delay of perception vehicle i is: Among them, ω i,i represents the ratio of tasks assigned by perception vehicle i to the vehicle itself; D i represents the amount of tasks formed on the perception vehicle i; S represents the number of CPU revolutions required to calculate 1 bit of the task, and f i is the calculation frequency on the sensing vehicle i; S322: The task processing delay of sensing vehicle i partially offloading tasks to RSU is: Among them, ω i,rsu represents the ratio of tasks assigned by sensing vehicle i to RSU, λ i is the computing resource allocation coefficient of RSU used to process the partially unloaded tasks from sensing vehicle i; f rsu Indicates the calculation frequency on RSU; S323: The task processing delay of partially offloading the task from the perception vehicle i to the collaborative computing vehicle j is: Among them, ω i,j represents the ratio of tasks assigned by perception vehicle i to collaborative computing vehicle j, and f j is the computation frequency on the collaborative computing vehicle j; the perception task allocation ratio is defined to be in the range of [0,1], and the sum of the task allocation ratios is 1: ω i,i ,ω i,rsu ,ω i,j ∈[0,1] and ω i,i +ω i,rsu +ω i,j =1;ω i,i represents the ratio coefficient of the task assigned by the perception vehicle i to the vehicle itself; ω i,rsu represents the distribution ratio coefficient of the task unloading from sensing vehicle i to RSU; ω i,j represents the task allocation ratio coefficient of the perception vehicle i to the collaborative computing vehicle j; the collaborative computing vehicle has sufficient computing power f j ≥f0; where f0 represents the minimum computing frequency requirement that the collaborative computing vehicle needs to meet; S33: Constructing the total delay model: For each perception vehicle, the entire task is completed only when the execution is completed locally, at the RSU, and on other collaborative computing vehicles: T i =max{T i,i ,T i,rsu ,T i,j } Only when all the perception car tasks are completed can the entire assisted perception task be completed:

6. The method for joint optimization of collaborative perception and computation offloading in the Internet of Vehicles according to claim 1, characterized in that: The optimization problem 1 (perception task allocation problem) in the collaborative perception stage is expressed as: s.t.C 1:d i,k ≤R sen , C 2:a i,k ∈{0,1}. The optimization problem 2 (perception task processing problem) in the collaborative computing phase is expressed as: stC3: C4:oh i,i +oh i,rsu +oh i,j =1, C5:d i,j ≤R com , C6:f j ≥f0, C7: C8:

7. The method for joint optimization of collaborative perception and computation offloading in the Internet of Vehicles according to claim 6, characterized in that: A block selection algorithm for perception vehicles is designed to solve optimization problem 1. For each block in the area of ​​interest of the requesting vehicle, an adaptive collaborative perception method based on perception quality priority sorting is used. The collaborative perception unit is iteratively formed through a successive optimal vehicle selection mechanism until the system's preset perception threshold constraints are met to obtain a perception solution.

8. The method for joint optimization of collaborative perception and computation offloading in the Internet of Vehicles according to claim 6, characterized in that: Designing the DDPG-HA algorithm to solve optimization problem 2 includes the following steps: The computation offloading optimization problem is transformed into a Markov decision process (MDP), whose state, action, and reward functions are described as follows: State Space definition is the state variable observed at time t, including the task offloading decision and task transmission rate. The state of the system can be defined as: s(t) = [N(t), R(t)]; Action Space definition At time t, the action performed by the agent following the action selection strategy consists of three parts: the allocation of perception tasks, the allocation of bandwidth resources, and the allocation of RSU computing resources. The action characteristics of the system can be expressed as: a1(t) = [δ(t), β(t), λ(t)]; By using the DDPG (Deep Deterministic Policy Gradient) algorithm to solve the task and resource allocation problem and constructing a bipartite graph of the task processing time between the perception vehicle and the collaborative computing vehicle, the optimal range of the maximum task completion time is determined. A binary search is performed between the minimum and maximum task completion times. For each intermediate value mid, the Hungarian Algorithm (HA) is used to determine whether an effective task allocation scheme can be found given the maximum task completion time mid. Finally, the task offloading decision a2(t) = [c(t)] can be found, and all actions are a(t) = [c(t), ω(t), β(t), λ(t)]. Reward Function For the current state s(t), each time an action a(t) is executed, a reward r(t) is obtained. The inverse of the delay in completing all the sensing tasks is used as the reward: r(t) = 1 / T. Introducing the Prioritized Experience Replay (PER) mechanism, which calculates the Temporal-Difference Error (TD) of each sample at each sampling, giving higher sampling priority to samples with larger TD errors. After multiple rounds of training, the cumulative reward value converges, and the optimal action and task completion delay can be obtained.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for joint optimization of collaborative perception and computation offloading in a connected vehicle network as described in any one of claims 1 to 8 are implemented.

Images