Multi-agent traffic incident real-time response system and method

Abstract

The invention discloses a multi-agent traffic event real-time response system and method, and the system comprises an event perception fusion module which is disposed on vehicle-mounted and roadside equipment, fuses the vehicle motion state, environment perception data and meteorological data, and generates an unstructured traffic behavior set in real time through a lightweight recognition model; the edge decision-making module is used for defining decision-making responsibilities by adopting a dynamic role binding mechanism and decomposing tasks based on a hierarchical memory storage architecture to generate response instructions; the central collaborative decision-making module is used for realizing strategy decoupling through role adapter layering and coordinating cross-domain response measures by adopting a three-level conflict resolution mechanism; and the strategy optimization module updates center parameters through federated learning and optimizes a response strategy in combination with a reinforcement learning reward model, so that second-level identification and dynamic response of traffic events can be realized, the congestion dissipation time is shortened, the false report and missing report rate is reduced, the secondary accident rate is reduced, and the disposal efficiency and safety of complex traffic scenes are comprehensively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation systems, and in particular to a multi-agent traffic event real-time response system and method. Background Art

[0002] Traditional traffic incident recognition relies on a single data source, and response measures are mostly static. This leads to data silos, insufficient real-time performance, poor adaptability, and a single response strategy. Current traffic incident recognition and response technologies primarily focus on single onboard sensors, the integration of multiple roadside devices, centralized cloud-based processing, and fixed rule bases, but all of these approaches present significant limitations. For example, patent publication number CN117557570B, based on a vehicle-mounted camera, achieves basic object recognition. However, the single data source is susceptible to interference from lighting, rain, and fog, resulting in a high rate of missed detections. Furthermore, static threshold settings make it difficult to distinguish between speed limits in school zones and actual congestion scenarios, leading to frequent false alarms. Furthermore, large-parameter models like YOLO lack real-time inference capabilities on vehicle terminals, creating a significant conflict between local computing power and model complexity. Response measures are limited to onboard warnings and are not linked to roadside traffic lights or cloud platforms, resulting in inefficient handling.

[0003] To further improve perception reliability, some technologies have turned to roadside radar and vision fusion solutions. For example, the patent disclosed as CN115685185A achieves vehicle trajectory tracking through target-level fusion of millimeter-wave radar and camera, and predicts abnormal behavior based on Kalman filtering. Although such systems have been used in intersection signal control, their data sources are still limited to roadside equipment. They lack refined input from on-board sensors and have difficulty capturing driver intentions or vehicle dynamics. At the same time, trajectory prediction relies on complex models such as cloud-based LSTM, resulting in response delays of more than 5 seconds. In addition, when the radar false alarm rate surges in severe weather, the system cannot correct misjudgments in real time through interactive mechanisms such as driver voice feedback, resulting in insufficient dynamic adaptability. The response level is also limited to single-point control of traffic lights, and has not formed a coordinated strategy with variable lanes, information board information release, etc.

[0004] To meet the needs of multi-source data integration, cloud-based centralized processing platforms attempt to aggregate vehicle-side, roadside, and meteorological data, using multimodal spatiotemporal feature extraction models for event analysis and signal timing optimization. However, a core bottleneck lies in the latency exceeding 10 seconds between data upload, cloud computing, and command issuance, making it impossible to meet the demands of high-real-time scenarios such as emergency braking. Furthermore, the transmission of raw video streams exacerbates bandwidth pressure and privacy risks. Furthermore, centrally trained models struggle to adapt to diverse scenarios such as mountainous roads and urban expressways, lacking the local decision-making capabilities of lightweight edge-based models, limiting their generalizability.

[0005] Another traditional approach relies entirely on manually pre-set rule bases. While these systems can achieve response times within seconds, their rule bases require frequent manual maintenance, making them unable to dynamically respond to multi-event scenarios like "unusual parking combined with the passage of large vehicles." Furthermore, they lack the ability to autonomously optimize strategies based on historical data. Coarse-grained control of fixed parameters also struggles to adapt to real-time traffic flow changes, and in complex road networks, can easily lead to conflicts between local optimization and global congestion. Summary of the Invention

[0006] In order to solve the problem of chain response failure in multi-vehicle rear-end collisions in heavy rain scenarios, the present invention proposes a multi-agent real-time response system and method for traffic incidents. By loading agents with lightweight language models on vehicle-mounted and roadside equipment, a local decision-making closed loop is achieved, bypassing cloud-based delays and improving the system's recognition and response speed.

[0007] A further object of the present invention is to achieve a second-level response to complex traffic events through hierarchical coordinated decision-making at the equipment layer, roadside edge layer, and central collaboration layer.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a multi-agent real-time traffic incident response system, characterized by comprising: The event perception fusion module, deployed on vehicle-mounted and roadside equipment, integrates vehicle motion status, environmental perception data, and meteorological data to generate unstructured traffic behavior sets in real time through a lightweight recognition model; The edge decision module uses a dynamic role binding mechanism to define decision responsibilities and generates response instructions based on the hierarchical memory storage architecture to decompose tasks; The central collaborative decision-making module achieves policy decoupling through layered role adapters and adopts a three-level conflict resolution mechanism to coordinate cross-domain response measures; The strategy optimization module updates the central parameters through federated learning and optimizes the response strategy in combination with the reinforcement learning reward model.

[0009] This technical solution achieves instant recognition and dynamic response to traffic events by building a multi-agent architecture for vehicle-road-cloud collaboration, combining lightweight edge perception modules, a role-customized decision-making mechanism, a dynamic conflict resolution engine, and a federated reinforcement learning optimization strategy. Specifically, multi-source data fusion is used to enhance perception robustness, policy decoupling is achieved through layered injection of domain knowledge via role adapters, a three-level conflict resolution decision-making mechanism is employed to coordinate cross-domain response measures, and a policy library is continuously optimized based on federated learning and reinforcement learning. Ultimately, a closed-loop "perception-decision-making-collaboration-optimization" system is formed, significantly improving the efficiency and safety of handling complex traffic scenarios.

[0010] Preferably, the event perception fusion module includes a collision risk warning unit, which calculates the collision time estimate in real time through an adaptive Kalman filter model. When the collision time estimate is less than the critical safety time corresponding to the current vehicle speed and road conditions, the collision risk probability value is calculated to generate an alarm text for the collision risk warning.

[0011] Preferably, calculating the collision risk probability value includes: obtaining the maximum deceleration braking time based on the road friction coefficient, dynamically adjusting the mean correction term and the variance correction term of the log-normal distribution based on the real-time attention score and the line of sight focus to obtain the driver reaction time, wherein the mean correction term is negatively correlated with the real-time attention score, and the variance correction term is negatively correlated with the line of sight focus; obtaining the collision risk probability value based on the driver reaction time, the maximum deceleration braking time and the system response delay, combined with the collision time estimation.

[0012] Preferably, the event perception fusion module has a voice feedback unit, including: Lightweight large model edge deployment, loading pre-trained large models on the roadside and on-board terminals; Unstructured text fusion interface, which uniformly encodes multi-source warning texts into natural language sequence input; Dynamic prompt word engine, when the event confidence is lower than the set threshold, generates query text for fuzzy semantics; Voice feedback closed-loop correction, real-time update of event judgment results based on the user's voice input on the vehicle side.

[0013] Preferably, the event perception fusion module includes a vehicle slip detection unit, which establishes a dynamic slip rate calculation model through cross-validation of four-wheel speed sensors and visual speed measurement, and obtains the actual slip rate based on the Pacejka magic formula. When the slip rate of any wheel is greater than a first threshold and lasts for more than a second threshold time, an alarm text of a vehicle slip event is generated.

[0014] Preferably, the vehicle skid detection unit comprises: The multi-source verification engine, through cross-verification of wheel speed sensors and visual speed measurement, triggers IMU arbitration when data conflicts reach the fourth threshold; Dynamic weight allocator, dynamically adjusts the multi-sensor data fusion weight according to the real-time road friction coefficient; The slip event decider generates a slip warning text when the arbitrated slip rate continuously exceeds a dynamic threshold for 500ms; the dynamic threshold is negatively correlated with the road friction coefficient.

[0015] Preferably, the edge decision module includes: The role definition layer dynamically binds agent responsibilities through prompt word templates; A hierarchical memory storage layer, including a short-term memory cache unit that uses a sliding window to store real-time event context and a long-term memory retrieval unit that associates historical event processing schemes based on feature similarity; At the task decomposition decision layer, Bi-LSTM extracts spatiotemporal features and improves the TOPSIS algorithm to sort task priorities.

[0016] Preferably, the central collaborative decision-making module is equipped with a role adapter, including: Feature extraction unit, extracts global context features of input events; The role transformation unit compresses features to the bottleneck dimension using a dimensionality reduction matrix and restores them using a dimensionality increase matrix after nonlinear activation; The residual fusion unit performs weighted fusion of the output of the role transformation unit and the original input features to generate role decision features.

[0017] Preferably, the reinforcement learning reward model of the policy optimization module uses the negative expected logarithmic sigmoid difference as the training loss function; the loss function for parameter learning of the intelligent agent model based on human evaluation includes: a reward-driven term, which maximizes the reward score of human-approved responses; a policy anchoring term, which constrains the deviation of the current policy from the basic safety policy by weighting the logarithm of the policy probability ratio; and a dynamic balance mechanism, which adjusts the constraint strength coefficient according to the urgency of the event.

[0018] The present invention also adopts the following technical solution: a multi-agent traffic incident real-time response method, based on the above-mentioned multi-agent traffic incident real-time response system, comprising the following steps: S1, identifies traffic events in real time and generates unstructured event feature sets; S2, dynamically adjusts role responsibilities based on event type and constructs preliminary response instructions based on event characteristics; S3, collaborative decoupling of multi-agent strategies, dynamically generating global optimization instructions through three-level conflict resolution; S4 updates parameters through federated learning and continuously iterates and evolves the strategy library in combination with the reinforcement learning reward model.

[0019] The beneficial effects of the present invention are: 1) Improve recognition accuracy and real-time performance: Enhance event detection robustness through multi-source data fusion from vehicle, road, and cloud, and Kalman filtering for denoising. 2) Dynamic response strategy optimization: Build an intelligent agent based on a lightweight large language model, combine federated learning with reinforcement learning, and dynamically generate multi-measure coordinated response plans; 3) Reduce edge computing load: Deploy lightweight models on vehicles or roadside equipment to reduce computing power requirements through unstructured text interaction; 4) Enhance system adaptability: Introducing user voice feedback and dynamic thresholds, such as the magic formula for slip rate and the TTC collision time compensation coefficient, to improve adaptability in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is an overall system block diagram of a multi-agent traffic incident real-time response system according to embodiment 1 of the present invention.

[0021] Figure 2 This is a system block diagram of the event perception fusion module of Example 1 of the present invention.

[0022] Figure 3 This is a system block diagram of the edge decision module of Example 1 of the present invention.

[0023] Figure 4 This is a system block diagram of the central collaborative decision-making module of Example 1 of the present invention.

[0024] Figure 5 This is a logic diagram of the three-level conflict resolution mechanism of Example 1 of the present invention.

[0025] Figure 6 This is a system block diagram of the strategy optimization module of Example 1 of the present invention.

[0026] Figure 7 This is a training flow chart of the strategy optimization module of Example 1 of the present invention. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] Example 1 This embodiment provides a multi-agent traffic incident real-time response system. Figure 1 As shown, it includes an event perception fusion module, an edge decision-making module, a central collaborative decision-making module and a strategy optimization module, which can achieve second-level identification and dynamic response of traffic events, reduce congestion dissipation time, lower false alarm and missed alarm rates and secondary accident rates, and comprehensively improve the handling efficiency and safety of complex traffic scenarios.

[0029] The event perception fusion module is deployed on vehicle-mounted and roadside equipment, integrating multi-source traffic data and generating structured event features in real time through a lightweight recognition model. Figure 2 .

[0030] The vehicle-side sensors mainly include accelerator, brake, steering wheel direction, camera, radar, inertial unit, GPS, voice, etc., using 9-axis IMU and GPS data as input. When both types of data are available, Kalman filtering is used to fuse and remove noise.

[0031] The surrounding environment is dynamically identified through on-board or roadside visual or radar fusion data. The target recognition model used includes but is not limited to YOLO 11 to determine the surrounding vehicle type, location, speed, road channelization, speed limit, traffic lights, signs, etc.

[0032] By deploying a variety of lightweight recognition models on vehicle-mounted terminals or roadside edge devices, unstructured traffic behavior sets can be identified and generated. The lightweight recognition models include but are not limited to the following sub-modules.

[0033] Abnormal parking monitoring unit: When the vehicle speed is less than 3km / h, an alarm text of abnormal parking event is generated, and the driving speed of surrounding vehicles is supplemented.

[0034] Low-speed detection unit: When the vehicle speed is lower than 20km / h, a low-speed warning text is generated, and the driving speed of surrounding vehicles is supplemented.

[0035] The emergency braking detection unit uses an improved sliding window variance algorithm to analyze triaxial acceleration data in real time, calculating the acceleration mutation characteristics within the window. When the variance in the vehicle's forward direction exceeds a third threshold, an emergency braking event warning is generated. For example, using N=18 and Δt=300ms, in this embodiment, the third threshold is set to 3.5m / s².

[0036] It should be noted that the acceleration variance of emergency braking in heavy rain weather will increase in both collision and non-collision scenarios. Therefore, a fixed threshold will not lead to missed detection, but false alarms may occur. In this embodiment, false alarms can be effectively prevented through multi-event collaborative detection. In some other implementation methods, a dynamic adjustment mechanism of threshold-related weather can also be set. For example, in heavy rain weather, the third threshold is raised according to the empirical coefficient to improve the accuracy of the single detection results of the emergency braking detection unit.

[0037] Abnormal lane change detection unit: takes the original trajectory P as input and uses a recurrent neural network model (such as GRU) to predict the trajectory P' in the next 2 seconds, using the Hausdorff distance d H Quantify the trajectory deviation, when the Hausdorff distance d H When the distance is greater than 1.2m, an alarm text of an abnormal lane change event is generated, and the event label is supplemented according to data such as the turn signal and the target steering wheel angle.

[0038] Vehicle Slip Detection Unit: This unit establishes a dynamic slip calculation model through cross-validation of four-wheel speed sensors and visual speed measurement. It derives the actual slip rate λ based on the Pacejka magic formula. When λ for any wheel exceeds 15% and persists for more than 500ms, a vehicle slip event warning is generated.

[0039] Large vehicle monitoring unit: Identifies target size through radar perception data or video data. If the target speed exceeds a threshold (e.g., 40 km / h on urban roads and 80 km / h on expressways), it generates a warning text indicating a large vehicle driving event.

[0040] Special vehicle monitoring unit: Through traffic control releases, surrounding vehicle type identification, special vehicle voiceprint matching, etc., it detects the passage needs of special vehicles (such as police cars, ambulances, fire trucks, etc.) and generates alarm texts for special vehicle traffic maintenance events.

[0041] Collision risk warning unit: The predicted collision time TTC is calculated in real time through the adaptive Kalman filter model. Specifically, the ratio of relative distance to speed is used as the base number, and the function of acceleration is used as the product correction term to obtain the predicted collision time. The product correction term is negatively correlated with the acceleration. For example, the product correction term can be set to 2 / [1+γsign(a rel )], where a rel represents acceleration, and γ is the dynamic compensation coefficient.

[0042] When the predicted collision time is less than 2.5 seconds, the collision risk is estimated based on the predicted collision time and the driver's reaction time. When the critical safety time corresponding to the current vehicle speed and road conditions is greater than the predicted collision time, the complementary percentage of the ratio of the predicted collision time to the critical safety time corresponding to the current vehicle speed and road conditions is used as the collision risk probability value. When the critical safety time corresponding to the current vehicle speed and road conditions is not greater than the predicted collision time, the collision risk probability value is 0.

[0043] Specifically, the critical safety time corresponding to the current vehicle speed and road conditions is affected by the vehicle braking time, system response delay, and human reaction delay. In this embodiment, the three are accumulated to obtain the critical safety time corresponding to the current vehicle speed and road conditions.

[0044] Among them, the vehicle braking time varies according to the current vehicle speed v and road conditions, and can be specifically expressed as v / μg, where μ is the road friction coefficient, generally between 0.3-0.9. The actual value is dynamically adjusted based on the real-time precipitation intensity data obtained from the meteorological platform in real time, and g is the gravity acceleration at the current position; the system response delay is obtained based on empirical values, and in this embodiment, it can be taken as 0.3s.

[0045] The human reaction delay time, that is, the driver's reaction time, obeys the log-normal distribution. The mean correction term and variance correction term are formed by the driver's real-time attention score and vision focus, respectively. The mean correction term is negatively correlated with the real-time attention score, and the variance correction term is negatively correlated with the vision focus.

[0046] Specifically, the mean correction term can be expressed as α(1-S attention / 100), the variance correction term can be expressed as e -βSfocus , add the mean correction term to the mean to get the corrected mean, and multiply the variance correction term by the variance to get the corrected variance.

[0047] Among them, S attention is a real-time attention score, which can be obtained based on the driver's facial features captured in real time by the on-board camera; Sfocus is the sight focus, which can be obtained based on the matching degree between the eye gaze point and the key area of ​​the road; α and β are learning parameters, and in this embodiment, both take the value of 0.15.

[0048] According to the collision risk probability value, a warning text of the collision risk warning is finally generated.

[0049] Pedestrian intrusion warning unit: Through thermal imaging or video data, an improved YOLOv5-Human detection network is used to generate warning text for pedestrian intrusion warnings.

[0050] Abnormal weather warning unit: connected to the public meteorological platform, it generates warning texts for severe weather such as strong winds, heavy rain, ice and snow, and high temperatures.

[0051] User Voice Feedback Unit: A pre-trained, lightweight, large language model is deployed at the edge, acting as an intelligent agent for edge data fusion and analysis. It uniformly accepts multi-source, multi-type alert text, constructs targets for each topological road segment through unstructured text input, comprehensively analyzes the spatiotemporal correlation of this information, and extracts the location, type, and adverse impact of abnormal events. Prompt words alert the model to key issues where judgments are ambiguous, inquire with the vehicle user, and adjust judgments based on user voice feedback.

[0052] The knowledge base is constructed based on seven dimensions: measure type, execution method, configuration parameters, response speed, event correlation, combination rules, and traffic impact, and complete prompt words for various intelligent agents are written for each response measure.

[0053] The edge decision module connects the event perception fusion module, uses a dynamic role binding mechanism to define decision responsibilities, and generates response instructions based on the hierarchical memory storage architecture to decompose tasks. Figure 3 .

[0054] The edge decision-making module can build a basic single intelligent agent, and adopts a four-layer progressive architecture to automatically build and guide the single intelligent agent to respond to traffic events, realizing the autonomous decision-making ability of the intelligent agent in the edge computing environment.

[0055] The four-layer progressive architecture includes a role definition layer, a hierarchical memory storage layer, a three-level task decomposition decision layer, and an execution feedback layer.

[0056] The role definition layer generates the agent identity feature matrix through pre-trained prompt word templates. Specifically, the prompt word template of the identity feature matrix is ​​"As {agent type}, your core responsibilities include: 1. Real-time monitoring of {data type} in {jurisdiction area}; 2. Assessing the event risk level (1-5) based on {decision-making model}; 3. Triggering the top three priority disposal plans in the {response measure library}; 4. Constraints: response delay <{max_latency}ms, decision confidence >{min_confidence}%; 5. Dynamic correspondence between device ID and responsibility area."

[0057] Realizing dynamic identity binding of a single agent through an identity feature matrix can solve the problem that traditional edge devices have fixed roles and cannot dynamically adjust responsibilities according to event types.

[0058] The hierarchical memory storage layer includes a short-term memory cache unit and a long-term memory retrieval unit. The short-term memory cache unit uses a sliding window to store real-time event context; the long-term memory retrieval unit associates historical event handling plans based on feature similarity.

[0059] Specifically, in this embodiment, short-term memory adopts the context cache mechanism of the Transformer-XL architecture with an adjustable window length (default 128 tokens). Long-term memory is based on the FAISS vector database to implement feature similarity retrieval, and associative memory is triggered when the cosine similarity is greater than 0.85.

[0060] The third-level task decomposition decision layer assists single-agent decision-making by calling segmentation models. For example, it uses Bi-LSTM to extract the spatiotemporal feature vectors of events, applies the improved TOPSIS algorithm (entropy weight method to optimize weights) to prioritize tasks, and decomposes complex instructions into API call sequences based on the measure implementation knowledge base.

[0061] The execution feedback layer presets 12 types of standard operation interfaces (signal control, drone scheduling, etc.).

[0062] In this embodiment, the edge decision module is also provided with an exception handling mechanism, and when the API returns an error code, the backup solution selector is started.

[0063] The central collaborative decision-making module connects multiple edge decision-making modules, realizes policy decoupling through role adapter layering, and adopts a three-level conflict resolution mechanism to coordinate cross-domain response measures. Figure 4 .

[0064] The following is a detailed description of the role adapter layered architecture.

[0065] The roles of intelligent agents are divided into edge event recognition agent, central decision-making agent, edge traffic control agent, edge information interaction agent, edge emergency rescue agent, and edge rapid disposal agent. Based on the pre-trained large model (such as DeepSeek 671B), a pre-trained understanding model LLM is established. pre (·).

[0066] To ensure the independence of central decision-making and the thinking of different single agents, lightweight role adapters are trained for the central agent and edge response agents.

[0067] The specific implementation involves freezing the parameters of the pre-trained large model and injecting role-specific knowledge through a lightweight residual structure. Taking the original event feature vector x (including event type, location, impact range, etc.) as input, the role-customized feature enhancement model generates adaptive features A(x) that incorporate role knowledge. For example, the emergency rescue role focuses on "injury classification" and the signal control role focuses on "timing coordination."

[0068] The process of obtaining the adaptation feature A(x) injected with role knowledge through the role-customized feature enhancement model is as follows.

[0069] First, the self-attention layer of a pre-trained large model (such as DeepSeek 671B) is called to extract the global context features of the input x and identify the correlation between the key elements of the event, such as the spatiotemporal binding of "position K12+300" and "east to west direction".

[0070] Then, through the feedforward network layer of the pre-trained large model, the attention features are nonlinearly transformed and domain knowledge is embedded to activate the specific features of traffic events, such as the decision-making association between "congestion and slow movement" and "signal cycle extension".

[0071] Next, downsampling compression is performed to reduce the feature dimension of the pre-trained model from d to the bottleneck dimension k (for example, 8 in this embodiment), reducing computational complexity. Nonlinear activation is performed to introduce the nonlinear characteristics required for role decision making through the activation function σ(·). Upsampling recovery is performed to restore the features to their original dimensions, preserving role knowledge. These steps generate role-sensitive features, such as the "ambulance dispatch priority" feature, which is enhanced by the edge emergency rescue agent.

[0072] Finally, in the residual feature fusion layer, the weighted sum of the original input x and the role adaptation features is taken to preserve the integrity of the original event information and prevent catastrophic forgetting. Only the role difference features are superimposed, such as the central decision-making agent increases the weight of the "global impact of the road network."

[0073] It should be noted that after downsampling compression, a bias calibration term b can be added first to learn the bias vector to compensate for the feature distribution offset of role adaptation in order to adapt to the decision preferences of different agents, such as the signal control agent's preference for a short-delay strategy.

[0074] In this technical solution, the role adapter injects lightweight role-specific decision features while retaining the general capabilities of the pre-trained model through a three-stage transformation of feature compression-role activation-feature reconstruction. Among them, the feature compression layer reduces high-dimensional features to an 8-dimensional bottleneck space; the nonlinear activation layer gives the role differentiated decision-making characteristics; and the residual fusion mechanism prevents the loss of original event information.

[0075] The normalization layer of the role adapter is then used to eliminate the semantic deviation between the role features and the original model, and the adapter output is distributed normalized to eliminate the feature distribution offset caused by role adaptation. The frozen pre-trained language model is then used to generate response instructions based on the normalized features to execute the final decision generation.

[0076] In this technical solution, the parameters of the pre-trained large model are completely frozen, and decision customization is only achieved through the front-end role adapter, which can significantly reduce the computing load.

[0077] The role fine-tuner continues with three-step hierarchical training on the traffic incident corpus.

[0078] The first step is to continue pre-training basic parameters such as the attention layer and feedforward layer on 100,000 traffic event texts to inject domain knowledge.

[0079] In the second step, the base parameters are frozen and each adapter is trained independently using 50,000 character annotation data.

[0080] The third step is to use KL divergence constraint and feature alignment (L KD ) realizes knowledge distillation, constrains sub-agents, and ensures strategic coordination among agents.

[0081] Among them, the third step is realized through a multi-agent collaborative decision-making unit, which includes a probability distribution constrainer, a feature space aligner and a dynamic weight allocator.

[0082] The probability distribution constraint uses the KL divergence loss function to force the edge agents to output decision probability distributions close to the central agent, eliminating decision conflicts, such as avoiding conflicts between signal control and variable lane instructions; the feature space aligner minimizes the Euclidean distance between the intermediate layer features of the central and edge agents, unifying the semantic understanding of events, such as ensuring consistent global "congestion and slow driving" judgment criteria; the dynamic weight allocator adjusts the feature alignment strength according to the urgency of the event.

[0083] In this technical solution, multi-agent strategy collaboration is achieved through the dual mechanisms of probability distribution constraints and feature space alignment.

[0084] The center-cooperative edge multi-agent architecture is described in detail below.

[0085] Using historical event recognition and response data, either manually entered or recorded by the system, a feature encoding is constructed based on the time, location, event type, event status, and a library of upstream and downstream response measures supported by real-world conditions. A Bi-LSTM is then used to extract spatiotemporal feature vectors. The edge event recognition agent encodes the features of real-time events and sends them to the central agent, which searches for similarities in the feature space between events in the historical memory library and real-time events.

[0086] The top ten most relevant event memories are retrieved. Using prompts, the central agent is guided to analyze the logical relationships between events and the response measure library, the current event and historical events, and the response measures and their effects, ultimately determining the types of response measures that can be selected. This is followed by the establishment of edge agents of this type, which analyze the executable measures within the scope of the event. Prompts are then used to communicate the executable measure types, execution terminals, interactions between existing measures, and related historical events to the edge agents. Each agent generates corresponding response actions and parameters based on the prompt dialogue.

[0087] The following is a detailed description of the conflict detection and resolution methods. Figure 5 .

[0088] A three-level conflict resolution mechanism is established: L1, L2, and L3. L1 involves resource exclusivity, for example, when different agents all require an ambulance but only one is available in the immediate area. L2 involves conflicting strategies, such as when an ambulance must ensure a green light to pass, while detour vehicles require shorter green light durations and variable lanes. L3 involves overlapping spatiotemporal influence domains, such as when congestion occurs at adjacent ground intersections and ramp entrances.

[0089] More than 100 conflict rules have been pre-loaded on the central agent, such as "IF Measure A includes "signal control", AND Measure B includes "variable lanes", AND the impact radius overlap is greater than 60%, THEN the conflict level = L2." The central agent collects the decision results of each agent in real time and determines the conflict level based on the knowledge base and event logic.

[0090] For different conflict situations, the central agent handles different conflicts according to the following process based on the conflict level.

[0091] For L1 level conflicts, the “human in the loop” feedback response mechanism is triggered, and the central intelligent agent generates a visual event and impact analysis report, which is pushed to the management and control personnel for decision-making.

[0092] For L2 conflicts, prompt words are used to guide the central agent to evaluate the risks that may arise from different choices, sort the risk types and priorities, and give priority to resolving life-saving risks, followed by ensuring traffic safety.

[0093] For L3 conflicts, a hybrid parameter method is used to generate compromise measures based on the response measures, confidence levels, and decision benefits of independent decisions made by different agents.

[0094] Specifically, the confidence and expected benefits are first fused to generate the measure reliability weights, then the weight distribution is normalized and the absolute weights are converted into probability distribution form, and finally the compromise measures are generated based on the original response measures.

[0095] By dynamically weighting and fusing the independent decisions of each intelligent agent, a globally optimal compromise measure is generated. Each measure is linearly combined according to weights, rather than simply voting or priority coverage, which can resolve conflicts in which the temporal and spatial influence domains overlap in multi-agent collaborative decision-making.

[0096] At the same time, the execution side sets comprehensive safety constraints on the specific instructions issued by the edge agent, such as the step length of traffic lights, the switchable flow direction of variable lanes, etc. If the instructions issued are abnormal, specific abnormal information will be fed back to prompt the edge agent to make corrections.

[0097] The strategy optimization module updates the parameters of the central collaborative decision module through federated learning and optimizes the response strategy by combining the reinforcement learning reward model. Figure 6 and Figure 7 .

[0098] The supervised fine-tuning of the single-agent adapter is described in detail below.

[0099] Using historical records as a prediction library, we train adapters for different agents, calculating the cross-entropy loss function for the next word prediction at each position in the sequence. The central agent's parameters are updated using federated parameter learning.

[0100] The interactive reinforcement learning of multiple agents is explained in detail below.

[0101] A reinforcement learning reward model is established to calculate scalar rewards for text quality for different inputs and outputs. The reward model uses the negative expected log Sigmoid difference as the training loss function. Specifically, the input is a traffic event description x, a high-quality response recognized by humans, and a set of responses to be evaluated generated by the agent. The difference between the reward score of the positive sample response and the reward score of the negative sample response set is calculated. The score difference is converted into a probability value through the Sigmoid function. The expected value of the negative log probability is used as the optimization target, which drives the reward model to improve the response quality differentiation ability.

[0102] The agent's decision-making strategy is optimized based on human preferences, and model parameters are dynamically adjusted using human evaluation data to improve the rationality of event responses. Specifically, a loss function is defined for parameter learning of the agent model based on human evaluation. The input is a traffic event description x and the agent's generated response y. The scalar reward value given by the reward model is calculated to minimize the expected negative reward of the agent's response strategy, driving the strategy to generate high-yield responses that meet human preferences. The logarithm of the probability ratio of the reinforcement learning strategy to the supervised fine-tuning base strategy is calculated, and its expected value is weighted by a negative coefficient -β as a policy anchor to constrain the policy update amplitude and prevent sudden deviations from the safety benchmark in the decision logic. The logarithmic probability of the response generated by the reinforcement learning strategy is calculated, and its expected value is weighted by a positive coefficient γ as an exploration incentive to maintain policy diversity and avoid local optimality.

[0103] The three-part weighted and driven strategy dynamically generates efficient response plans that are consistent with human evaluations while ensuring safety benchmarks, reducing traffic incident response delays and strategy conflict rates.

[0104] Example 2 This embodiment takes the rapid identification and response of multiple vehicle rear-end collisions on urban expressways in rainy days as an example to further illustrate the multi-agent real-time response system for traffic events provided in Example 1.

[0105] During the morning rush hour in a certain city, a sudden downpour caused the road surface on the expressway curve to become slippery. A car skidded due to emergency braking, and several cars behind it were unable to slow down in time due to obstructed vision, resulting in a five-car chain rear-end collision.

[0106] Traditional traffic management systems rely on a single on-board sensor (such as a camera) and fixed threshold rules (such as a sudden drop in speed triggering an alarm), making it difficult to quickly identify the location of an accident and the scope of impact under the interference of rainy days. This leads to delayed traffic light adjustments, the inability of rescue vehicles to arrive quickly, and continued congestion.

[0107] The application process of a multi-agent traffic incident real-time response system of the present invention is described in detail below.

[0108] First, multi-source data is collected and integrated in real time.

[0109] Real-time precipitation intensity data (15 mm / h) was accessed from the meteorological platform, and the road friction coefficient μ was dynamically corrected to 0.35. The accident truck triggered emergency braking (acceleration variance exceeded the threshold of 3.5 m / s²), the on-board IMU detected sideslip (slip rate s=0.25), and the GPS located the accident at the K12+300 curve on the expressway. The radar-vision fusion equipment detected a sudden drop in speed to 8 km / h in the accident area, and the trajectory of the rear vehicle shifted (Hausdorff distance d=4.2 m). The thermal imaging camera captured the pedestrian getting out of the vehicle to avoid danger.

[0110] The second is rapid analysis and event generation by edge agents.

[0111] The roadside edge computing node completes multi-source data fusion within 2 seconds and generates an unstructured alarm text: "[Event Type] Multi-vehicle rear-end collision, [Location] Expressway K12 (east to west), [Impact] 3 lanes blocked, rear queue length 800 meters, average speed 8 km / h, [Associated Risk] Pedestrian intrusion, secondary collision risk."

[0112] Finally, a lightweight large language model parses the text, extracts key information, and generates preliminary response suggestions: "1) Switch to a variable lane at the upstream intersection (turn left to the through lane and open the emergency lane); 2) Trigger a green wave belt for rescue vehicles to have priority; 3) Link the navigation platform to push detour information; 4) The roadside variable information board displays "Accident ahead, reduce speed to avoid secondary collision."

[0113] The implementation effects of this embodiment are as follows.

[0114] Improved response efficiency: Event identification delay is reduced from 30 seconds in traditional solutions to within 5 seconds, response strategy generation time is ≤ 3 seconds, and congestion dissipation time is reduced by 38% (120 minutes in traditional solutions → 75 minutes with this solution).

[0115] Optimization of false alarm and missed alarm rates: Through cross-validation of multi-source data (such as IMU + radar + meteorological data), the false alarm rate has been reduced from 15% to 2%; with the assistance of thermal imaging, the missed detection rate of pedestrian intrusion detection has been reduced from 10% to 1.5%.

[0116] Improved resource utilization: The accuracy of emergency lane opening decisions has increased to 95%, avoiding resource waste caused by ineffective lane switching; dynamic optimization of signal timing has increased auxiliary road traffic by 22%.

[0117] Enhanced user interaction: Driver voice feedback corrects misjudgments (e.g., changing "vehicles slow down" to "yield to pedestrians"); the attention scoring model is linked to the early warning system, reducing the secondary accident rate by 45%.

Claims

1. A multi-agent real-time traffic incident response system, characterized in that: include: The event perception fusion module, deployed on vehicle-mounted and roadside equipment, integrates vehicle motion status, environmental perception data, and meteorological data to generate unstructured traffic behavior sets in real time through a lightweight recognition model; The edge decision module uses a dynamic role binding mechanism to define decision responsibilities and generates response instructions based on the hierarchical memory storage architecture to decompose tasks; The central collaborative decision-making module achieves policy decoupling through layered role adapters and adopts a three-level conflict resolution mechanism to coordinate cross-domain response measures; The strategy optimization module updates the central parameters through federated learning and optimizes the response strategy in combination with the reinforcement learning reward model.

2. A multi-agent traffic incident real-time response system according to claim 1, characterized in that: The event perception fusion module includes a collision risk warning unit, which calculates the collision time estimate in real time through an adaptive Kalman filter model. When the collision time estimate is less than the critical safety time corresponding to the current vehicle speed and road conditions, the collision risk probability value is calculated and an alarm text of the collision risk warning is generated.

3. A multi-agent traffic incident real-time response system according to claim 2, characterized in that: Calculating the collision risk probability value includes: obtaining the maximum deceleration braking time based on the road friction coefficient, dynamically adjusting the mean correction term and variance correction term of the log-normal distribution based on the real-time attention score and the visual focus to obtain the driver's reaction time, wherein the mean correction term is negatively correlated with the real-time attention score, and the variance correction term is negatively correlated with the visual focus; according to the driver's reaction time, the maximum deceleration braking time and the system response delay, the collision risk probability value is obtained in combination with the collision time estimation.

4. A multi-agent traffic incident real-time response system according to claim 1, characterized in that: The event perception fusion module has a voice feedback unit, including: Lightweight large model edge deployment, loading pre-trained large models on the roadside and on-board terminals; Unstructured text fusion interface, which uniformly encodes multi-source warning texts into natural language sequence input; Dynamic prompt word engine, when the event confidence is lower than the set threshold, generates query text for fuzzy semantics; Voice feedback closed-loop correction, real-time update of event judgment results based on the user's voice input on the vehicle side.

5. A multi-agent traffic incident real-time response system according to claim 1, 2 or 4, characterized in that: The event perception fusion module includes a vehicle slip detection unit. Through cross-validation of four-wheel speed sensors and visual speed measurement, a dynamic slip rate calculation model is established. The actual slip rate is obtained based on the Pacejka magic formula. When the slip rate of any wheel exceeds a first threshold and lasts for more than a second threshold time, a warning text of a vehicle slip event is generated.

6. A multi-agent traffic incident real-time response system according to claim 5, characterized in that: The vehicle skid detection unit comprises: The multi-source verification engine, which cross-validates wheel speed sensors and visual speed measurement, triggers arbitration when data conflicts reach a fourth threshold. Dynamic weight allocator, dynamically adjusts the multi-sensor data fusion weight according to the real-time road friction coefficient; The slip event decider generates a slip warning text when the arbitrated slip rate continuously exceeds a dynamic threshold for 500ms; the dynamic threshold is negatively correlated with the road friction coefficient.

7. A multi-agent traffic incident real-time response system according to claim 1, characterized in that: The edge decision module includes: The role definition layer dynamically binds agent responsibilities through prompt word templates; A hierarchical memory storage layer, including a short-term memory cache unit that uses a sliding window to store real-time event context and a long-term memory retrieval unit that associates historical event processing schemes based on feature similarity; At the task decomposition decision layer, Bi-LSTM extracts spatiotemporal features and improves the TOPSIS algorithm to sort task priorities.

8. The multi-agent traffic incident real-time response system according to claim 1, characterized in that: The central collaborative decision-making module has a role adapter, including: Feature extraction unit, extracts global context features of input events; The role transformation unit compresses features to the bottleneck dimension using a dimensionality reduction matrix and restores them using a dimensionality increase matrix after nonlinear activation; The residual fusion unit performs weighted fusion of the output of the role transformation unit and the original input features to generate role decision features.

9. A multi-agent real-time traffic incident response system according to claim 1, characterized in that: The reinforcement learning reward model of the policy optimization module uses the negative expected log-sigmoid difference as the training loss function; The loss function for parameter learning of the agent model based on human evaluation includes: a reward-driven term that maximizes the reward score of human-approved responses; The policy anchor item constrains the deviation of the current policy from the basic security policy by weighting the logarithm of the policy probability ratio; the dynamic balance mechanism adjusts the constraint strength coefficient according to the urgency of the event.

10. A multi-agent traffic incident real-time response method, based on a multi-agent traffic incident real-time response system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1, identifies traffic events in real time and generates unstructured event feature sets; S2, dynamically adjusts role responsibilities based on event type and constructs preliminary response instructions based on event characteristics; S3, collaborative decoupling of multi-agent strategies, dynamically generating global optimization instructions through three-level conflict resolution; S4 updates parameters through federated learning and continuously iterates and evolves the strategy library in combination with the reinforcement learning reward model.

Images