A method and system for returning control of key perception results for a short-time safety window of a blind area of a city viaduct ramp merging

Abstract

This invention provides a method and system for timely return of key perception results within a short-term safety window for merging blind spots on urban elevated ramps. The method includes: acquiring local observation information of a target vehicle; mapping the local observation information to an information set belonging to the target vehicle and selecting joint actions on the information set; calculating the collaborative perception performance benefit, total energy consumption, and total latency of the target vehicle based on the joint actions of multiple vehicles; calculating the counterfactual value corresponding to the target vehicle selecting other candidate actions on the current information set based on comprehensive real-time utility, obtaining the real-time regret for each candidate action; updating the action selection probability on the information set, and outputting the average strategy for each information set after reaching a preset stopping condition; and online invoking joint actions in subsequent time slots based on the average strategy to control the timely return rate of key collaborative perception results within the safety decision window. This invention improves the timely return rate of key perception results within the safety decision window.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and vehicle-road-cloud collaborative perception technology, specifically involving a control method and system for timely return of key perception results for short-term safety windows when merging into blind spots on urban elevated ramps. Background Technology

[0002] In urban elevated highway ramp merging scenarios, vehicles on the main road and those on the ramp typically intersect within a short distance. This intersection is easily obstructed by bridge piers, guardrails, curves, slopes, and large vehicles, making it difficult for a single vehicle to obtain complete information about vehicles approaching from the side and rear in a timely manner. For vehicles about to merge, whether the blind spot cooperative perception results can be returned within a limited safety decision window directly affects whether they can make a smooth, timely, and safe merging decision. Therefore, timely return of key perception results for blind spots during elevated highway ramp merging has become a technical problem with significant scenario constraints in vehicle-road-cloud cooperative systems.

[0003] Existing technologies typically mitigate this problem through the following approaches: first, utilizing onboard perception processing units to directly perform local detection and decision-making; second, leveraging roadside units (RSUs) to provide auxiliary perception or relay capabilities; third, offloading perception tasks or intermediate results to edge nodes for processing; and fourth, jointly optimizing communication resources, computing resources, or task scheduling order to improve latency, energy consumption, or perception quality indicators. While these solutions cover different technical approaches such as local processing, roadside collaboration, edge offloading, and resource scheduling, most are geared towards general vehicle-to-everything (V2X) collaborative perception or general edge task offloading scenarios.

[0004] However, in the specific scenario of merging into blind spots on urban elevated highway ramps, multiple vehicles often simultaneously compete for uplink and edge processing resources within a very short period. Each vehicle can only monitor its own local channel, queue, and battery status in real time, making it difficult to obtain the complete status and immediate intentions of other vehicles. Under these conditions, simply increasing transmission power, enhancing local computing power, or expanding the edge offloading ratio may temporarily reduce individual processing latency, but it easily leads to link congestion, amplified edge queuing, and increased terminal power consumption, causing critical sensing results to miss the safety decision-making window. Therefore, this scenario essentially faces the scenario-specific impossible triangle of "timely return rate—terminal power consumption—cooperative sensing benefits," which cannot be simultaneously and optimally satisfied. Existing technologies generally lack a technical solution specifically designed for timely return of critical sensing results to the control target and for joint decision-making on local computing power levels, uplink transmission power levels, and task offloading modes under locally observable conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a control method for timely return of key perception results in the short-term safety window of urban elevated ramp merging blind spots. Under the condition that multiple vehicles concurrently request collaborative perception and only have local observable information, the method coordinates the selection of local computing power level, uplink transmission power level and task offloading mode to improve the timely return rate of key perception results within the safety decision window, while taking into account terminal energy consumption constraints and collaborative perception benefits.

[0006] To achieve the above objectives, the present invention provides the following solution: A control method for timely return of key perception results to the short-term safety window for merging blind spots on urban elevated ramps includes: When the target vehicle enters the blind spot associated area on the elevated ramp, the local observation information of the target vehicle in the current discrete decision time slot is obtained; The local observation information is mapped to the information set to which the target vehicle belongs, and a joint action is selected on the information set. The joint action includes local computing power level, uplink transmission power level, and task offloading mode. Under the condition of multiple vehicles concurrently requesting collaborative perception, the collaborative perception performance gain, total energy consumption and total latency of the target vehicle are calculated based on the joint actions of multiple vehicles, wherein the total latency represents the timeliness of the return of the perception result relative to the safety decision window. The comprehensive real-time utility is calculated based on the performance benefits of collaborative sensing, total energy consumption, and total latency to characterize the overall advantages and disadvantages of the current joint action under the scenario-based impossible triangle of "on-time return rate - terminal energy consumption - collaborative sensing benefits". While keeping other vehicle strategies unchanged, based on the comprehensive instant utility, the counterfactual value corresponding to the target vehicle choosing other candidate actions in the current information set is calculated sequentially, and compared with the information set value under the current strategy to obtain the instant regret of each candidate action; The immediate regrets are accumulated into counterfactual regrets according to the candidate actions, and the action selection probability on the information set is updated only when the accumulated regrets are positive. After the preset stopping condition is reached, the average strategy on each information set is output. Based on the aforementioned average strategy, joint actions in subsequent time slots are invoked online to control the on-time return rate of key collaborative perception results within the security decision window.

[0007] Preferably, the local observation information includes local channel state information, task queue state information, remaining battery power state information, and task load state information.

[0008] Preferably, the method for calculating the comprehensive immediate utility includes: The collaborative perception performance gain is calculated by subtracting the congestion penalty term after weighted summation of local processing gains and edge collaboration gains. The local computing power level of the target vehicle is used to calculate the local computing power consumption, the uplink communication power level and the effective transmission duration of the target vehicle are used to calculate the uplink communication power consumption, and the local computing power consumption and the uplink communication power consumption are added together to obtain the total power consumption. The local computing latency is calculated based on the current perception task computation volume of the target vehicle and the local computing power level. The uplink communication latency is calculated based on the amount of data to be uploaded by the target vehicle and the uplink data transmission rate. The task queuing latency, the local computing latency and the uplink communication latency are added together to calculate the total latency. The comprehensive real-time utility is calculated based on the collaborative perception performance gains, the total energy consumption, the total latency, and the preset weighting coefficients.

[0009] Preferably, the formula for calculating immediate remorse is: in Indicates the first During round iteration, the target vehicle is in the information set Forced selection of candidate actions The counterfactual value of time; This represents the value of the information set under the current hybrid strategy.

[0010] Preferably, the expression for counterfactual remorse is as follows: The expression for updating the action selection probability on the information set when the accumulated regret is positive is as follows: in , Represents vehicle intelligent agent In information set The set of actions to choose from.

[0011] Preferably, the expression for the averaging strategy is as follows: Represents vehicle intelligent agent In information set upper action The cumulative policy weights are used to calculate the average policy for the action at the end.

[0012] This invention also provides a control system for timely return of key perception results to the short-term safety window for merging blind spots of urban elevated ramps, used to implement the method, including: The data acquisition module is used to acquire local observation information of the target vehicle in the current discrete decision time slot when the target vehicle enters the blind zone associated area of ​​the elevated ramp. The information set mapping module is used to map the local observation information to the information set to which the target vehicle belongs, and to select a joint action on the information set. The joint action includes local computing power level, uplink transmission power level and task offloading mode. The calculation module is used to calculate the collaborative perception performance gain, total energy consumption and total latency of the target vehicle based on the joint actions of multiple vehicles under the condition of multiple vehicles concurrently requesting collaborative perception, wherein the total latency represents the timeliness of the return of the perception result relative to the safety decision window. The comprehensive evaluation module is used to calculate the comprehensive real-time utility based on the performance benefits of collaborative sensing, total energy consumption, and total latency, so as to characterize the comprehensive advantages and disadvantages of the current joint action under the scenario-based impossible triangle of "on-time return rate - terminal energy consumption - collaborative sensing benefits". The instant regret acquisition module is used to calculate the counterfactual value of the target vehicle choosing other candidate actions in the current information set based on the comprehensive instant utility, while keeping the strategies of other vehicles unchanged, and compare it with the information set value under the current strategy to obtain the instant regret of each candidate action. The average strategy acquisition module is used to accumulate the immediate regret into counterfactual regret according to the candidate actions, and update the action selection probability on the information set only when the accumulated regret is positive, and output the average strategy on each information set after reaching the preset stopping condition. The on-time return rate control module is used to make online calls to joint actions in subsequent time slots according to the average strategy, so as to control the on-time return rate of key collaborative perception results within the security decision window.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention targets the specific scenario of a short-term safety window when urban elevated ramps merge into blind spots. It focuses on the technical goal of timely return of key collaborative perception results and jointly controls local computing power, uplink transmission power and task offloading method. Therefore, it has clear scenario targeting and problem orientation. (2) Without relying on complete global state information, the present invention achieves distributed decision updates through a counterfactual regret minimization mechanism based on information sets, thereby improving the engineering feasibility under local observable conditions and the collaborative adaptability under multi-vehicle concurrent scenarios. (3) The present invention can jointly weigh the benefits of collaborative sensing, terminal energy consumption and result return latency, and improve the on-time return rate of key sensing results within the security decision window, the stability of policy invocation and the long-term operation effect by averaging policy output. Attached Figure Description

[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a schematic diagram of the system architecture of the present invention; Figure 3 This is a diagram illustrating the impossible trilemma and CFR update mechanism of the present invention; Figure 4 This is a scene construction diagram for the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] Example 1 In this implementation, the application targets are multiple intelligent connected vehicles, at least one roadside unit, and one edge collaborative processing node in the merging blind zone scenario of urban elevated ramps. The "elevated ramp merging blind zone" refers to a scenario where, during the short-distance merging process between main road vehicles and ramp vehicles, the perception field of a single vehicle is incomplete due to obstructions from bridge piers, guardrails, curves, slopes, and large vehicles. Therefore, the merging decision must be made with the assistance of vehicle-road-cloud collaborative perception results. This scenario has the following practical characteristics: First, the risk window is short; if the perception results cannot be returned within seconds before merging, they easily lose their decision-making value. Second, multiple vehicles will simultaneously request uplink and edge computing power within a similar timeframe, forming short-term bursts of resource competition. Third, if a high computing power and high power strategy is adopted for a long period, it will significantly increase terminal energy consumption and thermal load.

[0019] The specific, real, and niche pain point technical problem addressed by this implementation can be further summarized as follows: Within a short decision window of a few seconds before the entrance of an urban elevated highway ramp, vehicles waiting to merge on the ramp and adjacent vehicles on the main road will almost simultaneously initiate blind spot collaborative perception requests. However, each vehicle can only grasp local information such as its own link, queue, and battery level, and cannot obtain the complete status and intentions of other vehicles in real time. If high-power uploading and high-computing-power execution are uniformly adopted, although individual latency may be shortened locally, it will cause uplink congestion, amplified edge queuing, and increased terminal power consumption overall, causing key perception results to miss the merging decision window. Therefore, this invention does not generalize to pursue maximum throughput or minimum average latency, but specifically solves the more specific engineering problem of "returning blind spot perception results on time within a short merging window with controllable costs." In other words, this scenario essentially faces a typical "impossible triangle": timely return of perception results, controllable terminal power consumption, and maximization of collaborative perception benefits are difficult to achieve simultaneously without cost. The extreme optimization of any one of them may come at the expense of sacrificing at least one of the other two.

[0020] To address the aforementioned pain points, this implementation unifies the vehicle's local state, action selection, benefit evaluation, and counterfactual regret update in each discrete time slot into a distributed decision-making framework.

[0021] For ease of explanation, the vehicle set is denoted as: in The number of vehicles participating in the collaborative sensing for ramp merging. The system in each discrete time slot... Make a decision on the vehicle Only local information can be observed And based on this, determine the information set. Then select an action from the information set.

[0022] like Figure 1 As shown, a control method for timely return of key perception results for short-term safety windows in urban elevated highway merging blind spots includes: S1: When the target vehicle enters the blind zone associated area on the elevated ramp, acquire the local observation information of the target vehicle within the current discrete decision time slot; a further implementation method is that the local observation information includes local channel state information, task queue state information, remaining battery power state information, and task load state information. Specifically: Local channel state information, such as instantaneous channel quality from vehicle to roadside unit, uplink gain, available bandwidth status, or access quality level; Task queue status information, such as the queue length of currently pending sensing tasks, the amount of data to be uploaded, and the task arrival intensity; Remaining battery status information, such as the current remaining battery power, thermal load status, or allowable power consumption level of the vehicle terminal; Task load status information, such as the number of computation cycles required for the current perception task, the complexity level of the target detection model, and whether it belongs to a high-priority inbound security task.

[0023] Regarding scene randomization, the initial local channel gain can be set in the range of Within this range, different location-based obstruction and channel fading conditions are simulated; the task data volume is set within a certain range. Within this range, different sensory frame or feature compression package sizes are simulated; the task computation load is set within a certain range. Internally, it can simulate the complexity of different detection models or fusion tasks; the vehicle scale can also be adjusted when necessary. The percentage of vehicles unloaded simultaneously is used to reflect changes in the intensity of resource competition.

[0024] Vehicles in time slots The local observation can be represented as This step plays the role of "transforming specific scenario pain points into algorithm input" in the seven steps. In other words, the three real-world constraints of blind spots on elevated ramps, short-term merging windows, and multi-vehicle concurrent competition are not merely left in the background description, but are compressed into local observation variables, serving as the starting point for subsequent information set partitioning, action selection, and regret update.

[0025] S2: Map local observation information to the information set to which the target vehicle belongs, and select joint actions on the information set. Joint actions include local computing power level, uplink transmission power level, and task offloading mode.

[0026] Specifically, after acquiring local observation information, an information set belonging to the target vehicle is constructed according to preset mapping rules. The purpose of this information set is to classify locally observable conditions using a finite state bucket approach when the vehicle cannot obtain the global system state, and to maintain corresponding action strategies for each category. In one specific embodiment, the local channel state can be divided into three levels: low, medium, and high, and the queue pressure and remaining battery power can be jointly divided into three levels: low voltage, medium voltage, and high voltage, thus forming a total of nine information sets. The information set satisfies… Representing intelligent agents The set of all reachable information.

[0027] This step addresses the problem of "how to form an executable decision-making basis when the global state is unavailable." By mapping continuously changing local states to a finite set of information, the complexity of policy maintenance can be reduced, and vehicles can reuse historical learning experience under similar local conditions. This is a prerequisite for the subsequent application of counterfactual regret updates.

[0028] After determining the information set, the system selects the current action from a preset discrete action space. The action is defined as follows: in This is the local computing power tier. This is the uplink transmit power setting. This is for task unloading mode. An action space setting that can be implemented is as follows: local computing power tier. Transmit power level Uninstall mode , Indicates local execution. This indicates edge unloading. In this embodiment, the current action is not directly given by a fixed threshold, but is selected by the target vehicle from the action probability distribution corresponding to the current information set. Different information sets correspond to different candidate action preferences. When the system learns that a certain type of local state is more suitable for a certain set of computing power-power-unloading combinations, it will be more inclined to select the corresponding action in the same or similar local states. The system architecture diagram of this invention is as follows. Figure 2 As shown.

[0029] Therefore, step S2 actually embodies the first layer of meaning of the algorithm design idea of ​​this invention, namely: first, mapping the continuous, local, and incomplete environmental state into a finite information set, and then performing discrete action selection on the information set so as to minimize counterfactual regret in subsequent applications. Without this step, subsequent regret accumulation, complexity control, and average policy output would lack an implementable carrier.

[0030] S3: Under the condition of multiple vehicles concurrently requesting collaborative perception, calculate the collaborative perception performance gain, total energy consumption and total latency of the target vehicle based on the joint actions of multiple vehicles, where the total latency represents the timeliness of the perception result return relative to the safety decision window.

[0031] The impossible trinity and CFR update mechanism diagram of the present invention, as shown below. Figure 3 As shown.

[0032] A further implementation method includes a method for calculating comprehensive immediate utility, comprising: The performance gain of collaborative awareness is calculated by subtracting the congestion penalty term after weighted summation of local processing gains and edge collaboration gains; specifically, the performance gain can be expressed in the following form: in This indicates the computational load of the task, reflecting the intensity of the current task's demand for computing resources; It represents the equivalent channel gain, reflecting the uplink transmission quality between the current vehicle and the roadside unit; This represents the congestion penalty caused by multiple vehicles unloading simultaneously, reflecting the reduction of collaborative benefits due to edge queuing and link contention. , and These are the weight parameters in the performance model, used to adjust the impact of local processing benefits, edge collaboration benefits, and congestion penalties on overall performance, respectively. and The normalization coefficient is used to eliminate the influence of different dimensional parameters on the revenue calculation scale. This formula reflects that increasing local computing power improves local processing quality, increasing offloading success rate and link quality improves collaborative revenue, while multi-vehicle competition leads to performance degradation. The logarithmic form shows that revenue growth exhibits diminishing marginal returns, consistent with the engineering principle that revenue growth slows down after improvements in computing power and links reach a certain level.

[0033] The local computing power consumption is calculated based on the target vehicle's local computing power level, and the uplink communication power consumption is calculated based on the target vehicle's uplink transmit power level and effective transmission duration. The local computing power consumption and uplink communication power consumption are then added together to obtain the total power consumption. Specifically, the power consumption can be calculated as follows: in in The computational energy consumption coefficient related to terminal hardware is used to characterize the energy consumption sensitivity of different vehicle chip platforms to frequency changes. This is the power consumption index, used to reflect the fact that power consumption increases non-linearly as the frequency increases; The effective transmission duration is used to measure the duration of the current uplink process. The above energy consumption model satisfies the basic physical law that "the higher the local computing power and the greater the transmission power, the higher the energy consumption is usually," and is therefore reasonable. In this embodiment, the energy consumption coefficient is taken as... The dynamic power consumption index is used to reflect the sensitivity of the terminal chip platform in converting computing power frequency into power consumption; the dynamic power consumption index is taken as... This is used to describe the degree of nonlinear increase in energy consumption when computing power increases.

[0034] The local computation latency is calculated based on the current perception task computation volume of the target vehicle and the local computing power level. The uplink communication latency is calculated based on the amount of data to be uploaded by the target vehicle and the uplink data transmission rate. The task queuing latency, local computation latency, and uplink communication latency are then added together to obtain the total latency. Specifically, the latency can be calculated as follows: in in To delay queuing, For local calculation latency, For communication / unloading latency, The amount of data to be uploaded. For link speed, To represent the number of CPU cycles required for a task, and For system bandwidth, The noise power is taken as the expression for the communication rate. The above expression uses a Shannon-type logarithmic rate model, which accurately reflects the relationship between increased transmission rate with increasing link gain and transmit power, and decreased transmission rate with increasing noise. Therefore, it can be used as a feasible approximation model in this embodiment. If a 6G-specific coding modulation and resource block allocation mechanism is used in actual deployment, an equivalent rate model can be used instead without changing the core idea of ​​this invention. In this embodiment, the system bandwidth is taken as... Used to characterize the scale of available wireless transmission resources; noise power is taken as... It is used to characterize the background noise at the receiver.

[0035] In one implementation scenario, when multiple vehicles simultaneously approach the merging point and select the edge unloading mode, the system calculates a congestion penalty and queuing delay based on the number of vehicles unloading simultaneously, the edge node capacity, and the current link congestion level. This reflects the short-term, sudden competition characteristics in merging blind zone scenarios. The scenario construction diagram of this invention is as follows: Figure 4 As shown.

[0036] S4: Calculate the comprehensive real-time utility based on the performance benefits of collaborative sensing, total energy consumption, and total latency to characterize the comprehensive advantages and disadvantages of the current joint action under the scenario-based impossible triangle of "on-time return rate - terminal energy consumption - collaborative sensing benefits".

[0037] Specifically, after obtaining the performance gains, total energy consumption, and total latency, the system further calculates the vehicle's overall real-time utility: in , and These are weighting coefficients for performance, energy consumption, and latency, respectively. In one embodiment, they can be taken as... It is used to balance the benefits of collaborative perception, vehicle energy consumption, and latency penalties.

[0038] Correspondingly, from the perspective of long-term operation across multiple time slots, the overall optimization objective of this invention in the short-time window scenario of ramp blind spots can be further expressed as follows: in This represents the joint strategy of each vehicle across different information sets. This represents the total number of time slots for strategy evaluation or learning. This formula indicates that the present invention does not merely pursue the instantaneous optimal action within a single time slot, but rather aims to maximize the long-term cumulative expectation of comprehensive immediate utility throughout the entire ramp merging risk window and multiple rounds of interaction.

[0039] However, the aforementioned optimization objective is difficult to solve directly for at least the following reasons: First, vehicles can only observe local channels, local queues, and remaining battery power, and cannot obtain a complete global state, making it difficult to directly construct the precise input for a centralized global optimization problem. Second, the actions of multiple vehicles are coupled within the same time slot. The unloading, power, and computing power choices of one vehicle can affect the link congestion, edge queuing, and overall benefits of other vehicles, resulting in a significant multi-agent coupling characteristic in the objective function. Third, the system is in a high-speed, time-varying environment. If the global state is re-collected and the joint optimal strategy is solved in each time slot, it will lead to excessive computational and communication overhead, making it difficult to meet the real-time requirements in short-window scenarios with ramp blind spots. Fourth, the action space is a discrete combination space composed of computing power levels, transmission power levels, and unloading modes. As the number of vehicles, information sets, and actions increases, the complexity of direct exhaustive search or centralized optimization will rise rapidly. Therefore, the present invention adopts the counterfactual regret minimization solution method based on information set described in subsequent steps S5 to S7, which approximates the better joint strategy in the long run by updating the information set round by round under local observable conditions.

[0040] The purpose of this step is to transform the multi-objective conflict problem into a unified, comparable scalar, allowing different candidate actions to be selected under the same evaluation scale. For the scenario of merging into a blind spot on a ramp, if the perception result exceeds the safety decision window, the latency penalty weight should be increased to reduce the utility of such actions; if the system is in a low-battery or long-term operating state, the energy consumption penalty weight can be increased to suppress the frequent occurrence of high-power actions. In actual deployment, the weight coefficients can be fixed or configured and switched according to the current road level, traffic risk level, weather conditions, or remaining vehicle battery power. In other words, this step solves the problem of "how to unify the three objectives of timely return, controllable energy consumption, and effective perception into the same decision-making scale," that is, explicitly transforming the "impossible triangle" in this scenario into a computable, weighted, and iteratively optimized comprehensive utility function.

[0041] Therefore, step S4 plays the role of "multi-objective modeling and solving objective definition" in the seven steps. It not only reflects the objective function basis of the algorithm design, but also provides a unified utility reference for the subsequent average regret convergence analysis. This is because only by first defining the comparable benefits of each round on the information set can the counterfactual value and regret have calculable objects.

[0042] To make the model engineering feasible, the following constraints are further introduced: in , , and These represent the vehicle terminal's maximum computing power, maximum transmission power, single-slot energy budget, and maximum tolerable latency, respectively.

[0043] Therefore, the ultimate problem considered in this embodiment can be described as follows: Under the framework of incomplete information extended game theory, by minimizing the cumulative counterfactual regret of each agent on each information set through CFR, a distributed multi-agent resource coordination strategy that balances performance gains, energy consumption constraints, and latency constraints can be obtained in the long-term average sense. This problem statement also corresponds to the core positioning of this invention: "introducing CFR into the 6G incomplete information multi-agent resource coordination scenario relatively early".

[0044] S5: While keeping the strategies of other vehicles unchanged, based on the comprehensive immediate utility, calculate the counterfactual value of the target vehicle choosing other candidate actions in the current information set in turn, and compare it with the information set value under the current strategy to obtain the immediate regret of each candidate action.

[0045] Specifically, after obtaining the immediate utility of the current action, the system further evaluates whether a better result could be achieved by selecting other candidate actions within the current information set. To this end, it is necessary to calculate the counterfactual value for each candidate action in the information set.

[0046] Specifically, while keeping the current strategies of other vehicles unchanged, we assume in turn that the target vehicle selects other candidate actions on the current information set, recalculate the possible utility values, and thus obtain the counterfactual values ​​of the candidate actions.

[0047] The entire 6G multi-slot interaction process is modeled as an extended game. For any intelligent agent Its information set upper action The counterfactual value is defined as: , That is, according to the strategy profile in other information sets Action, and in the information set Forced selection action At that time, intelligent agent Expected returns.

[0048] Accordingly, information set The counterfactual value of this is: Then in the During round iteration, actions Immediate counterfactual remorse is defined as: in Indicates the first During round iteration, the target vehicle is in the information set Forced selection of candidate actions The counterfactual value of time; This represents the value of the information set under the current hybrid strategy.

[0049] If immediate regret is positive, it means that choosing this action more often in the past would have yielded a higher reward; if immediate regret is negative, it means that this action has no advantage over the current hybrid strategy. By repeating this process on each round of information, the system can gradually distinguish which actions are more suitable for the current local state.

[0050] This invention employs a counterfactual regret minimization mechanism because it directly addresses the core challenge in this scenario: vehicles cannot know the immediate unloading choices, power choices, and edge queuing states of other vehicles, but can perform counterfactual comparisons of different computing power-power-unloading combinations under the same information set based on historical results. This mechanism does not require a complete global model or static environmental assumptions, and can continuously refine its strategy using actual interaction results. Therefore, it exhibits better adaptability compared to fixed-rule methods and better real-time feasibility compared to centralized optimization methods that rely on global observations.

[0051] Furthermore, step S5 forms the computational basis for the average regret convergence analysis. In other words, whether the average regret decreases is not a theoretical appendix discussed separately from the algorithm steps, but is directly based on the round-by-round calculation of the counterfactual value of each information set and each candidate action in this step.

[0052] S6: Accumulate immediate regrets into counterfactual regrets based on candidate actions, and update the action selection probability on the information set only when the accumulated regret is positive. Output the average policy on each information set after reaching the preset stopping condition; after obtaining immediate regrets, the system accumulates them into counterfactual regrets: The counterfactual regret minimization (CFR) method is used to update the action probabilities on the information set. Specifically, the expression for updating the action selection probabilities on the information set when the cumulative regret is positive is as follows: in , Represents vehicle intelligent agent In information set The following is a set of selectable actions. The impossible triangle and CFR update mechanism diagram of this invention are shown below. Figure 3 As shown.

[0053] A further implementation method is that the expression for the averaging strategy is as follows: Represents vehicle intelligent agent In information set upper action The cumulative policy weights are used to calculate the average policy for the final action. Represents vehicle intelligent agent In information set Above, for another candidate action The cumulative strategy weight.

[0054] The core idea of ​​this step is to update action probabilities only using positive regret, that is, prioritizing the increase of weights for actions that have historically been more worthwhile, while naturally reducing the selection probability of actions that have performed poorly over the long term. If the positive regret of all candidate actions in a certain information set is zero, it indicates that the system has not yet formed a clear preference, and in this case, a uniform distribution is used to re-explore. In this implementation, the target vehicle independently maintains a set of cumulative regret and action probabilities for each information set, so the same vehicle may exhibit different decision tendencies in different local states.

[0055] From a technical perspective, steps S5 and S6 together constitute the core improvement of this invention compared to general threshold-based offloading methods: they do not directly specify a threshold corresponding to a specific action, but rather allow the system to automatically adjust the action probability based on long-term historical performance on different information sets, thereby transforming "local experience" into a "reusable distributed collaborative strategy." This process constitutes an important technical feature of this invention in 6G incomplete information multi-agent resource collaboration scenarios.

[0056] From a complexity perspective, in a single update, each vehicle needs to traverse candidate actions across its reachable information set and perform counterfactual value calculations and positive regret normalization updates. Therefore, the main computational load in a single round can be described as... Same order; in total number of iterations At that time, the overall computational workload is correspondingly the same as Same-order growth. Therefore, it can be seen that the complexity analysis of this invention can also find its direct point in step S6, rather than being detached from the seven steps. In this embodiment, the number of iterations in the main experiment is taken as... This is used to ensure that the policy learning process has a sufficient convergence observation window.

[0057] At the same time, step S6 is also the realization step of the conclusion of "average regret convergence", because the reason why average regret can decrease is that this step continuously increases the probability of historically better actions and reduces the probability of long-term poor performance, thereby causing the cumulative counterfactual regret to gradually converge.

[0058] Specifically, regarding complexity analysis: Let the number of vehicles be... The average number of information sets per vehicle is The average number of actions per information set is The main overheads of a single round of CFR updates include: The calculation of the combined action benefit has a complexity of approximately or It depends on the implementation of wireless resource scheduling and edge queuing models: the counterfactual value and regret update of each information set, with a complexity of approximately The complexity of policy normalization and average policy accumulation is also approximately [missing information]. .

[0059] Therefore, in carrying out K After rounds of iteration, the overall complexity of the algorithm can be expressed as: .

[0060] For scenarios with a large action space, further compression of complexity can be achieved by using action pruning, sampled CFR, or function approximation CFR.

[0061] S7: Based on the average strategy, online invocation of joint actions in subsequent time slots is performed to control the on-time return rate of key collaborative perception results within the security decision window. For example... Figure 3 As shown.

[0062] Specifically, when the preset maximum number of iterations is reached, or when the probability change of actions in several consecutive rounds falls below a preset threshold, the system stops iterating and outputs the average strategy across all information sets as the deployment strategy. The reason for using an average strategy instead of a single-round instantaneous strategy is that the average strategy can smooth out short-term fluctuations, reduce drastic action swings caused by individual abnormal states, and is more suitable for continuous execution in real-world road environments. The average strategy can be expressed as... The task of merging into the ramp is completed on time, and the threshold is determined. This means that the blind spot perception result must be returned before this threshold is considered to still have effective value for the inflow decision.

[0063] In actual deployment, the onboard collaborative decision-making module can cache the average policy table in the local controller. Whenever a vehicle re-enters the blind zone on an elevated ramp and collects new local observation information, it can first determine the information set it belongs to, and then directly call the corresponding average policy to give the current action suggestion. The action suggestion may include: whether to prioritize local processing of the current blind zone perception task, whether to offload the task to the edge node, what transmit power level to use for uplink transmission, what local computing power level to use to run the perception model, and whether to switch to a low-power conservative execution mode in low battery or high congestion conditions.

[0064] The adoption of an average strategy instead of a single-round instantaneous strategy also reflects the technical consideration of "stability of the average strategy": since the average strategy can smooth out local fluctuations in a single round of updates, it is more suitable for continuous deployment of real vehicle-road-cloud systems and can also avoid drastic fluctuations in action between adjacent time slots.

[0065] Regarding the experimental closed loop, step S7 already encompasses the simulation scenario, experimental metrics, experimental design, and implementation plan: the scenario corresponds to multiple vehicles competing for link and edge resources within a short window of the ramp blind zone; its metrics include at least average utility, average performance, average energy consumption, average latency, average regret, and on-time task completion rate; its design includes a comparison of CFR with threshold strategies, random strategies, and locally greedy strategies; and its results, in turn, verify the effectiveness of the algorithmic closed loop formed by steps S1 to S6. Therefore, the experiment is not separate from the seven steps but rather a direct extension of the deployment and verification phase of step S7.

[0066] In summary, the above seven steps can be further refined as follows: Step S1 addresses the question of "what to see"; Step S2 addresses the question of "how to form an executable decision-making basis when the overall picture is incomplete"; Step S3 addresses the question of "what are the actual benefits, costs, and latency of the current joint action"; Step S4 addresses the question of "how to uniformly evaluate multi-objective conflicts"; Step S5 addresses the question of "how to compare the merits of alternative actions under local information conditions"; Step S6 addresses the question of "how to transform historical comparison results into a better choice tendency for the next time"; and Step S7 addresses the question of "how to stably output the learning results as an actual deployment strategy". Through the above closed loop, this invention can address the core pain point of "whether the key collaborative perception results can be returned in a timely manner within the safety decision window," and form an implementable technical solution applicable to the scenario of merging into blind spots on elevated ramps. Compared with methods that only pursue global throughput or single-item performance optimization, this invention emphasizes the collaborative control of task timeliness, terminal energy consumption, and local implementability in specific high-risk, niche scenarios.

[0067] Example 2 like Figure 2 , Figure 4 As shown, the present invention also provides a control system for timely return of key perception results to short-term safety windows for merging blind spots on urban elevated ramps, for implementing a method comprising: The data acquisition module is used to obtain local observation information of the target vehicle in the current discrete decision time slot when the target vehicle enters the blind zone associated area on the elevated ramp, based on the scenario-based impossible triangle formed by "time return rate - terminal energy consumption - collaborative perception benefits". The information set mapping module is used to map local observation information to the information set to which the target vehicle belongs, and select joint actions on the information set. The joint actions include local computing power level, uplink transmission power level, and task offloading mode. The calculation module is used to calculate the collaborative perception performance gain, total energy consumption and total latency of the target vehicle based on the joint actions of multiple vehicles under the condition of multiple vehicles concurrently requesting collaborative perception. The total latency represents the timeliness of the return of the perception result relative to the safety decision window. The comprehensive evaluation module is used to calculate the comprehensive real-time utility based on the performance benefits of collaborative sensing, total energy consumption, and total latency, so as to characterize the comprehensive advantages and disadvantages of the current joint action under the scenario-based impossible triangle of "on-time return rate - terminal energy consumption - collaborative sensing benefits". The instant regret acquisition module is used to calculate the counterfactual value of the target vehicle choosing other candidate actions in the current information set based on comprehensive instant utility, while keeping the strategies of other vehicles unchanged, and compare it with the information set value under the current strategy to obtain the instant regret of each candidate action. The average policy acquisition module is used to accumulate immediate regret into counterfactual regret according to candidate actions, and only update the action selection probability on the information set when the accumulated regret is positive. After reaching the preset stopping condition, the average policy on each information set is output. The on-time return rate control module is used to make online calls to joint actions in subsequent time slots based on the average strategy, so as to control the on-time return rate of key collaborative perception results within the security decision window.

[0068] In one feasible example, a specific implementation device is provided, wherein the vehicle acquires uplink measurement values ​​through a local communication module, obtains the current task complexity through a perception computing module, and obtains remaining battery power information through a power management module, and sends the above data to the vehicle-mounted collaborative decision-making module. If historical time slot data exists, the scheduling results of the previous time slot, the power level used, and the edge return delay statistics can also be input as extended observations. In a specific implementation, the data acquisition module is implemented by calling the local communication module to obtain the local channel state, calling the perception computing module to obtain the task load state, and calling the power management module to obtain the remaining battery power state. In addition to the data acquisition module, the information set mapping module, the calculation module, the comprehensive evaluation module, the instant regret acquisition module, the average strategy acquisition module, and the on-time return rate control module are all integrated or run in the vehicle-mounted collaborative decision-making module.

[0069] Furthermore, the system also includes a roadside cooperative control module deployed on the roadside unit and an edge resource processing module deployed on the edge node. Specifically: the roadside cooperative control module aggregates uplink requests from multiple vehicles and broadcasts cooperative status information, including edge node load status and channel congestion level; the data acquisition module receives this cooperative status information via a local communication module as part of the local observation information; the edge resource processing module receives the unloading perception task when a vehicle selects the edge unloading mode, allocates edge computing resources for processing, and returns the perception results to the vehicle for the vehicle to execute an inbound decision.

[0070] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for controlling the return of key perception results in time for a short safety window in a blind area for a city viaduct ramp merge, characterized in that, include: When the target vehicle enters the blind spot associated area on the elevated ramp, the local observation information of the target vehicle in the current discrete decision time slot is obtained; The local observation information is mapped to the information set to which the target vehicle belongs, and a joint action is selected on the information set. The joint action includes local computing power level, uplink transmission power level, and task offloading mode. Under the condition of multiple vehicles concurrently requesting collaborative perception, the collaborative perception performance gain, total energy consumption and total latency of the target vehicle are calculated based on the joint actions of multiple vehicles, wherein the total latency represents the timeliness of the return of the perception result relative to the safety decision window. The comprehensive real-time utility is calculated based on the performance benefits of collaborative sensing, total energy consumption, and total latency to characterize the overall advantages and disadvantages of the current joint action under the scenario-based impossible triangle of "on-time return rate - terminal energy consumption - collaborative sensing benefits". While keeping other vehicle strategies unchanged, based on the comprehensive instant utility, the counterfactual value corresponding to the target vehicle choosing other candidate actions in the current information set is calculated sequentially, and compared with the information set value under the current strategy to obtain the instant regret of each candidate action; The immediate regrets are accumulated into counterfactual regrets according to the candidate actions, and the action selection probability on the information set is updated only when the accumulated regrets are positive. After the preset stopping condition is reached, the average strategy on each information set is output. Based on the aforementioned average strategy, joint actions in subsequent time slots are invoked online to control the on-time return rate of key collaborative perception results within the security decision window.

2. The method according to claim 1, characterized in that, The local observation information includes local channel status information, task queue status information, remaining power status information, and task load status information.

3. The method according to claim 1, characterized in that, The method for calculating the overall immediate utility includes: The performance gain of collaborative perception is calculated by subtracting the congestion penalty term after weighted summation of local processing gains and edge collaboration gains. The local computing power level of the target vehicle is used to calculate the local computing power consumption, the uplink communication power level and the effective transmission duration of the target vehicle are used to calculate the uplink communication power consumption, and the local computing power consumption and the uplink communication power consumption are added together to obtain the total power consumption. The local computing latency is calculated based on the current perception task computation volume of the target vehicle and the local computing power level. The uplink communication latency is calculated based on the amount of data to be uploaded by the target vehicle and the uplink data transmission rate. The task queuing latency, the local computing latency and the uplink communication latency are added together to calculate the total latency. The comprehensive real-time utility is calculated based on the collaborative perception performance gains, the total energy consumption, the total latency, and the preset weighting coefficients.

4. The method according to claim 1, characterized in that, The formula for calculating immediate remorse is: in Indicates the first During round iteration, the target vehicle is in the information set Forced selection of candidate actions The counterfactual value of time; This represents the value of the information set under the current hybrid strategy.

5. The method according to claim 4, characterized in that, The expression for counterfactual remorse is as follows: The expression for updating the action selection probability on the information set when the accumulated regret is positive is as follows: in , Represents vehicle intelligent agent In information set The set of actions to choose from.

6. The method according to claim 1, characterized in that, The expression for the averaging strategy is as follows: Represents vehicle intelligent agent In information set upper action The cumulative policy weights are used to calculate the average policy for the action at the end.

7. A control system for timely return of key perception results to a short-term safety window for merging blind spots of urban elevated ramps, used to implement the method described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire local observation information of the target vehicle in the current discrete decision time slot when the target vehicle enters the blind zone associated area of ​​the elevated ramp. The information set mapping module is used to map the local observation information to the information set to which the target vehicle belongs, and to select a joint action on the information set. The joint action includes local computing power level, uplink transmission power level and task offloading mode. The calculation module is used to calculate the collaborative perception performance gain, total energy consumption and total latency of the target vehicle based on the joint actions of multiple vehicles under the condition of multiple vehicles concurrently requesting collaborative perception, wherein the total latency represents the timeliness of the return of the perception result relative to the safety decision window. The comprehensive evaluation module is used to calculate the comprehensive real-time utility based on the collaborative sensing performance benefits, total energy consumption, and total latency, so as to characterize the comprehensive advantages and disadvantages of the current joint action under the scenario-based impossible triangle of "on-time return rate - terminal energy consumption - collaborative sensing benefits". The instant regret acquisition module is used to calculate the counterfactual value of the target vehicle choosing other candidate actions in the current information set based on the comprehensive instant utility, while keeping the strategies of other vehicles unchanged, and compare it with the information set value under the current strategy to obtain the instant regret of each candidate action. The average strategy acquisition module is used to accumulate the immediate regret into counterfactual regret according to the candidate actions, and update the action selection probability on the information set only when the accumulated regret is positive, and output the average strategy on each information set after reaching the preset stopping condition. The on-time return rate control module is used to make online calls to joint actions in subsequent time slots according to the average strategy, so as to control the on-time return rate of key collaborative perception results within the security decision window.

Images