An AI-driven intelligent road condition recognition system

Abstract

This invention relates to the field of intelligent road condition recognition technology, specifically to an artificial intelligence-driven intelligent highway road condition recognition system, comprising a roadside sensing unit, a mobile sensing unit, an emergency status acquisition unit, and an intelligent analysis platform. This system constructs a dynamic causal graph with traffic congestion events as the root node based on multimodal real-time data and emergency response status data through the intelligent analysis platform. It uses a stage recognition model to infer the emergency handling stages and their expected durations, and generates lane opening / closing schemes and guidance strategies through a lane-level traffic management decision module. An information dissemination unit then distributes these schemes to vehicles. Through causal graph evolution prediction and lane-level dynamic traffic management, the system updates the congestion situation in congested areas in real time and provides precise lane-level traffic guidance.

Description

Technical Field

[0001] This invention relates to the field of intelligent road condition recognition technology, and more specifically, to an artificial intelligence-driven intelligent road condition recognition system. Background Technology

[0002] When sudden events such as traffic accidents, temporary construction, or road closures occur on highways or urban expressways, drivers in queues behind are often in a complete information blind spot because they cannot directly observe the actual situation of the congestion ahead. Specifically, drivers can only see the brake lights of vehicles in their lane and adjacent lanes illuminate, their speed suddenly decreases, or they come to a complete stop, but they cannot know the root cause of the congestion. The cause of the congestion may be a multi-vehicle rear-end collision, a single vehicle breakdown, cargo spillage, road collapse, temporary maintenance work, or traffic control closures, and the corresponding handling procedures and durations vary greatly depending on the type of emergency.

[0003] For drivers stuck in traffic jams, the information blind spots include at least the following aspects: First, the specific cause of the congestion is unclear, and drivers cannot determine whether it is a minor scrape that can be cleared on its own within minutes, or a serious accident involving heavy clearing and the transfer of the injured; Second, the specific lanes blocked are unclear, and drivers cannot determine which lanes are directly occupied by accident vehicles or construction areas, or which lanes are still passable or about to be reopened; Third, the real-time progress of emergency response is unclear, and drivers do not know whether traffic police have arrived at the scene, whether the investigation and evidence collection have been completed, or whether the clearing vehicles have been dispatched or are in operation; Fourth, the estimated time for clearing the congestion is unclear, and drivers cannot make reasonable plans based on this information, such as whether to wait in place or take an alternative route in advance.

[0004] The lack of such information can trigger a series of negative consequences. From a driver's psychological perspective, an unknown waiting time is far more likely to induce anxiety and anger than a known long wait. Driven by anxiety, some drivers will frequently honk their horns and attempt to change lanes to check the situation ahead. From a traffic behavior perspective, when some drivers decide to break out of the current queue and forcibly change lanes, the changing vehicles will collide with vehicles traveling normally in the target lane, forcing those vehicles to brake suddenly. This disturbance will spread along the traffic flow, causing the congestion to spread from a single lane to adjacent lanes, and in severe cases, even to the opposite lane. More seriously, disorderly lane changes and cutting in significantly increase the risk of lateral and rear-end collisions, easily triggering secondary accidents behind existing congested areas, further exacerbating traffic paralysis and posing a risk of injury or death.

[0005] In existing technologies, traffic guidance methods for highways and urban expressways mainly include the following schemes.

[0006] The first approach involves displaying text-based warnings via roadside variable message signs. These signs, typically mounted on road gantry or roadside posts, display brief messages such as "Accident ahead, slow down" or "Road closed due to construction ahead." However, the spacing between these signs is usually several kilometers or more, and the displayed content is limited by the number of characters and fixed templates. They cannot accurately describe whether the accident occupies the left, middle, or right lane, nor can they provide real-time updates on the arrival, investigation, and clearing of obstacles by traffic police. Drivers behind cannot make informed lane-changing decisions based on this vague information.

[0007] The second approach involves using navigation software to display congested road segment markers and estimated travel times. This software relies on spatiotemporal aggregation analysis of floating car trajectory data to render congested sections in red or dark red on electronic maps and calculate estimated travel times based on historical data or a simple empirical formula of queue length versus dissipation time. However, this approach has significant drawbacks: historical statistical models cannot adapt to dynamically changing on-site situations; for example, the early arrival of tow trucks or sudden secondary accidents can cause significant deviations in the preset dissipation times. Furthermore, the navigation software can only provide a travel time prediction for the entire road segment, failing to provide detailed suggestions for specific lanes or inform drivers of lane-level information such as whether the first lane is passable while the second and third lanes are occupied. Additionally, the limited sampling frequency and positioning accuracy of the navigation software, based on in-vehicle GPS, makes it difficult to accurately distinguish the specific lane a vehicle is in during dense queues, thus lacking the data foundation for generating lane-level guidance strategies.

[0008] The third approach involves integrating drone aerial imagery into traffic monitoring. Drones can fly over accident sites and transmit high-resolution overhead images or video streams, allowing monitoring center personnel to intuitively understand the location of vehicles involved in the accident, the number of lanes occupied, and queue lengths. However, in current applications, drones are often used purely as image acquisition tools. Operators control the drones remotely or via preset flight paths, and then rely on manual observation and subjective judgment. This approach lacks deep integration of multimodal perception data and emergency response status data. Emergency response status data refers to structured information from traffic management platforms, emergency call platforms, or construction reporting platforms, including alarm records, police arrival times, investigation start times, clearing vehicle dispatch status, and estimated time of traffic restoration. Existing solutions fail to semantically align the visual features in real-time drone images with the aforementioned emergency status data and to model causal relationships. Therefore, they cannot automatically infer the current stage of the emergency response, predict the remaining duration based on the processing stage, or automatically convert the inference results into differentiated traffic management plans for each lane.

[0009] In summary, existing technologies are insufficient in three aspects: real-time acquisition of detailed status of traffic congestion areas ahead, accurate prediction of congestion dissipation time, and generation of lane-level orderly traffic management strategies. There is an urgent need for an intelligent system that can deeply integrate multimodal on-site perception with emergency response progress and perform causal reasoning and lane-level decision-making. Summary of the Invention

[0010] The purpose of this invention is to provide an artificial intelligence-driven intelligent road condition recognition system, thereby solving the problems mentioned in the background art.

[0011] To achieve the above objectives, the AI-driven intelligent road condition recognition system includes a roadside sensing unit, a mobile sensing unit, an emergency status acquisition unit, and an intelligent analysis platform. The roadside sensing unit is used to collect multimodal real-time data from a fixed perspective in the traffic jam area; The mobile sensing unit is deployed on a drone that can fly over the traffic jam area and is used to supplement the collection of multimodal real-time data of the blind spots covered by the roadside sensing unit under the control of the received dispatch instructions. The emergency status acquisition unit is used to acquire emergency response status data from at least one of the traffic management platform, the emergency call platform, and the construction reporting platform. The intelligent analysis platform includes an event cause-effect graph construction module, a stage identification and time prediction module, a lane-level traffic management decision module, and an information release unit. The event causal graph construction module is used to extract event elements and construct a dynamic causal graph with the current traffic jam event as the root node based on the multimodal real-time data and the emergency response status data. The nodes of the dynamic causal graph represent entities or states that have a causal relationship with the traffic jam event, and the edges represent the causal associations and temporal effects between nodes. The stage identification and time prediction module is used to perform state reasoning on the dynamic causal graph using a trained stage identification model, and output emergency response stage labels and their expected durations. The lane-level traffic management decision module is used to input the emergency handling stage label and expected duration, and the lane-level traffic flow status into the multi-objective optimizer, and solve the lane opening / closing scheme and lane-level guidance strategy in a rolling time-domain optimization manner within a micro traffic simulation environment. The information dissemination unit is used to send the lane opening / closing scheme and lane-level guidance strategy to vehicles affected by traffic congestion.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This AI-driven intelligent highway traffic condition recognition system constructs a dynamic causal graph with traffic congestion events as the root node based on multimodal real-time data and emergency response status data through an intelligent analysis platform. It uses a stage recognition model to infer the emergency handling stages and their expected durations, and a lane-level traffic management decision module generates lane opening / closing plans and guidance strategies. An information dissemination unit then distributes these plans to vehicles. Through causal graph evolution prediction and lane-level dynamic traffic management, the system updates the congestion situation in congested areas in real time and provides precise lane-level traffic guidance. Attached Figure Description

[0013] Figure 1 This is a block diagram of the overall system structure of the present invention; Figure 2 This is a system interface diagram for real-time acquisition of highway status according to the present invention; Figure 3 This is a diagram of the interface of the drone scheduling and real-time video display system of the present invention; Figure 4 This is a diagram of the emergency status display interface of the present invention; Figure 5 This is an event causal interface diagram of the present invention; Figure 6 This is a diagram of the stage identification and time prediction interface of the present invention; Figure 7 This is a diagram of the lane-level traffic management decision interface of the present invention. Detailed Implementation

[0014] The technical solutions in 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.

[0015] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0016] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0017] Please see Figure 1 As shown, an artificial intelligence-driven intelligent road condition recognition system is provided, including a roadside sensing unit, a mobile sensing unit, an emergency status acquisition unit, and an intelligent analysis platform; Roadside sensing units are used to collect multimodal real-time data from a fixed perspective in traffic congestion areas; Mobile sensing units are deployed on drones that can fly over congested areas to supplement the collection of multimodal real-time data in blind spots covered by roadside sensing units under the control of received dispatch instructions. The emergency status acquisition unit is used to acquire emergency response status data from at least one of the traffic management platform, emergency call platform, and construction reporting platform. The intelligent analysis platform includes an event cause-effect graph construction module, a stage identification and time prediction module, a lane-level traffic management decision-making module, and an information dissemination unit; Among them, the event cause-effect graph construction module is used to extract event elements and construct a dynamic cause-effect graph with the current traffic jam event as the root node based on multimodal real-time data and emergency response status data. The nodes of the dynamic cause-effect graph represent entities or states that have a causal relationship with the traffic jam event, and the edges represent the causal associations and time sequence effects between nodes. The phase identification and time prediction module is used to perform state reasoning on the dynamic causal graph using a trained phase identification model, and output emergency response phase labels and their expected durations. The lane-level traffic management decision module is used to input the emergency response stage label and expected duration, lane-level traffic flow status into the multi-objective optimizer, and solve the lane opening / closing scheme and lane-level guidance strategy in a rolling time-domain optimization manner within a micro traffic simulation environment. The information dissemination unit is used to send lane opening / closing plans and lane-level guidance strategies to vehicles affected by traffic congestion.

[0018] The specific implementation method is as follows: First, in order to obtain the current traffic congestion status (traffic congestion factors, including traffic congestion due to accidents, traffic congestion due to lane borrowing, and traffic congestion due to road closures), this solution collects multimodal real-time data of the traffic congestion area from a fixed perspective through roadside sensing units, and cooperates with drones deployed by mobile sensing units that can fly over the traffic congestion area to supplement the collection of multimodal real-time data of the blind spots covered by the roadside sensing units under the control of received dispatch instructions. The roadside sensing unit consists of fixed sensing devices deployed along the highway, used to collect multimodal real-time data on congested areas from a fixed perspective. This unit includes surveillance cameras and millimeter-wave radar deployed at intervals along both sides of the road or on gantries. The surveillance cameras output visible light image data as a continuous video stream, while the millimeter-wave radar simultaneously outputs point cloud data containing the position, velocity, and echo intensity of moving targets. The two types of data are time-stamped and then jointly calibrated spatially to form a multimodal sensing stream covering the assigned road segment. The fixed installation location of the roadside sensing unit allows it to continuously monitor the traffic conditions of a preset area; however, blind spots inevitably exist in areas with poor sensing coverage, such as curves, hilltops, large vehicle obstructions, or large gaps between devices. Figure 2 The diagram shows the system interface for the roadside sensing unit to acquire real-time road conditions. The diagram displays the monitoring content of cameras (roadside camera A1) and millimeter-wave radars (millimeter-wave radars B1 and C1) at different locations. At the same time, this system also maintains data connection with the corresponding traffic lights and weather sensors to acquire the real-time status of traffic lights and weather conditions in the corresponding monitoring areas.

[0019] Meanwhile, the drones deployed by the mobile sensing units are equipped with gimbal cameras and edge computing modules, and are normally parked in automated drone nests along the roadside. When the intelligent analysis platform detects that the perception confidence of the roadside sensing units in a certain area is lower than the threshold or that there is a completely uncovered blank area, the platform performs a quality assessment on the multimodal data transmitted back in real time by the roadside sensing units to obtain the average confidence of target detection within the coverage area of ​​each roadside sensing device. When the average confidence level of a certain sub-region is lower than a preset threshold or no moving target is detected for three consecutive acquisition cycles, the sub-region is determined to be a perception blind zone. The platform extracts the polygon boundary vertex coordinate sequence, center point coordinates, and required supplementary acquisition spatial resolution of the blind zone. The intelligent analysis platform encapsulates the perception blind zone information into a perception coverage requirement message and sends it to the mobile sensing unit through the internal communication link. The corresponding perception coverage requirement message contains the polygon vertex coordinate sequence of the blind zone, the expected minimum ground sampling distance, and the task priority identifier.

[0020] After receiving the sensing coverage request message, the mobile sensing unit reads the status information of all available drones in the current drone cluster from the drone status database, including each drone's unique identifier, current geographic coordinates, current remaining battery power, and the lens focal length and pixel size of its onboard camera. Simultaneously, the mobile sensing unit constructs a joint optimization model for drone task assignment and waypoint parameters, aiming to minimize the weighted sum of the lack of sensing blind spot coverage and the total energy consumption cost of the drone cluster. The specific construction method is as follows: A polygonal area of ​​perception blind spot to be covered is provided. Its area The currently available drone set is as follows: The current position coordinates of drone i are The remaining power is The minimum required ground sampling distance is ; The corresponding decision variables include: assignment variables ,in This indicates that drone i has been assigned to perform a blind spot detection task; the hovering position of drone i. Flight altitude Gimbal tilt angle ; The final objective function is: ; in This indicates the area within the blind zone that is still not covered by any camera's field of view after being covered by a drone swarm. This represents the energy consumed by drone i to perform the blind spot detection task. This represents the full charge level of drone i, used as a normalization benchmark. These are pre-defined weighting coefficients used to balance the importance of coverage quality and energy consumption. The corresponding blind spot area Calculate using the following formula: ; in Indicates that drone i is in the hovering position Flight altitude and gimbal tilt angle Under these conditions, the field of view projection area of ​​its mounted camera on the ground, It is represented as the union of the field-of-view projection areas of all assigned drones. This represents a function for calculating the area of ​​a polygon. This represents the intersection of the blind zone polygon and the union of the covered area.

[0021] The energy cost of the corresponding drone i performing the task The calculation formula is as follows: ; in Let i be the drone starting from its current position. Fly to target hover position The three-dimensional path distance This indicates the preset hovering duration for this blind spot detection task. The energy consumption coefficient, representing the energy consumption per unit flight distance, is determined by the characteristics of the drone's battery and its flight efficiency. The energy consumption coefficient representing the unit hovering time; The corresponding conditions include: Remaining battery power constraints ensure that the assigned drone has sufficient energy reserves to complete the mission. ; Ground sampling distance constraints ensure that the spatial resolution of the acquired images meets the requirements for blind zone analysis. ; in, Let be the ground sampling distance of UAV i under the current parameters, and its calculation formula is: ; in, This indicates the pixel size of the camera mounted on the drone. This indicates the focal length of the camera lens.

[0022] At the same time, airspace restrictions need to be imposed on drones to ensure that the hovering position is within the geographical range where flight is permitted: ; in This represents the pre-defined horizontal projection area of ​​the flyable airspace. and These are the minimum and maximum permissible flight altitudes, respectively. and These represent the working range of the gimbal's pitch angle.

[0023] Subsequently, the mobile sensing unit solves the above optimization model. Based on the geometric center of the blind zone, the current position of each drone, and its remaining battery power, it uses a greedy algorithm to divide the blind zone into several sub-regions and assigns them to the corresponding drones, thus determining... The initial value is then determined. Then, for each assigned UAV, targeting its covered sub-region, a gradient-based local search method is used to iteratively adjust the value while satisfying ground sampling distance constraints and spatial constraints. , and The goal is to minimize the sum of the uncovered area and energy cost within the sub-region. The iteration terminates when the change in the objective function is less than a set threshold, yielding the optimal hovering position, flight altitude, and gimbal pitch angle for each UAV.

[0024] Finally, the mobile sensing unit generates scheduling instructions based on the solution results. Each scheduling instruction includes the UAV identifier, target waypoint longitude, target waypoint latitude, flight altitude, gimbal pitch angle, and hovering duration. The scheduling instructions are then sent to the corresponding UAVs via wireless communication links to perform blind spot filling flight tasks.

[0025] like Figure 3The diagram shows the system interface for drone scheduling and real-time video display. Based on the results, scheduling commands are generated to dispatch the corresponding drones (e.g., Drone Alpha-1 in the diagram is the drone executing this scheduling). The diagram displays the drone's status (flight status and standby status). Simultaneously, based on the drone's flight status, its battery level, distance, and speed are displayed. The drone is controlled to reach its designated coordinates (latitude and longitude) according to the corresponding scheduling commands. After reaching the designated monitoring location, the drone uses its onboard camera to monitor the current traffic congestion in real time, feeding back the video feed to the system interface for live display. This allows for the acquisition of real-time images of currently uncovered areas, enabling full coverage monitoring of congested areas in conjunction with roadside cameras.

[0026] Furthermore, in order to respond to the current traffic congestion status in the congested area, this solution utilizes the emergency status acquisition unit to obtain emergency response status data from at least one of the traffic management platform, emergency call platform, and construction reporting platform.

[0027] The emergency status acquisition unit establishes a secure data link with the aforementioned external platform through a network interface, and receives structured emergency response status data in a preset polling cycle or an event-driven subscription push method.

[0028] The specific content of emergency response status data includes: traffic accident type, precise accident location, lane number occupied, traffic police arrival time, investigation start time, clearing vehicle dispatch status, and expected time of traffic restoration obtained from the traffic management platform; accident nature described by the caller, call time, number of vehicles involved and license plate information, and police force status dispatched after receiving the call obtained from the emergency call platform; and construction permit number, start and end chainage of the construction section, lane occupied by construction, construction start time, planned end time, and whether the current construction activity is suspended or in progress, etc., obtained from the construction reporting platform.

[0029] Simultaneously, the emergency status acquisition unit performs field mapping and format normalization on the received heterogeneous data, uniformly encapsulating it into emergency response status records with a preset data structure. These records are then continuously pushed to the intelligent analysis platform via an internal communication link, allowing the event cause-effect graph construction module to call upon them to construct and update event status nodes and processing stage attributes in the dynamic cause-effect graph. For example... Figure 4 The image shows the system's emergency status display interface. The image displays the location of the current traffic jam area and the type of traffic jam (such as traffic jams caused by vehicle breakdowns or road maintenance work). When a new traffic jam occurs, it is added to the emergency status display interface according to the corresponding traffic jam type. For example, 12 traffic accidents and 5 construction items were added today.

[0030] Furthermore, the intelligent analysis platform, as the core processing node, communicates with the roadside sensing unit, mobile sensing unit, and emergency status acquisition unit. Internally, the intelligent analysis platform includes an event cause-effect graph construction module, a stage identification and time prediction module, and a lane-level traffic management decision-making module.

[0031] The event cause-effect graph construction module extracts event elements based on multimodal real-time data and emergency response status data, and constructs a dynamic cause-effect graph with the current traffic congestion event as the root node. The nodes of the dynamic cause-effect graph represent entities or states that have a causal relationship with the traffic congestion event, and the edges represent the causal relationships and temporal effects between nodes. The specific construction method is as follows: First, entity and state extraction is performed to identify vehicles, obstacles and closed areas occupying lanes from the perception fusion results, and the occupancy length and throughput rate of each lane are calculated. For lane occupancy detection, based on the lane boundary lines or visual lane line detection results provided by the high-precision map, the current road segment is first divided into continuous cross-sections along the longitudinal direction, and a lane number sequence from the left edge to the right edge of the road is constructed for each cross-section. Each target in the perception fusion result is projected onto the road plane coordinate system, and the set of lane numbers crossed by the target is determined according to the spatial overlap relationship between the target's bounding box and the boundaries of each lane. For targets classified as motor vehicles, non-motor vehicles, or pedestrians, if their instantaneous speed amplitude is lower than a preset static determination threshold and they remain stationary for more than a set minimum number of continuous frames, the target is marked as a vehicle or obstacle occupying a lane. For continuous closed polygons composed of cones, construction barriers, and guardrails, the entire polygon is regarded as a closed area, and the set of lanes occupied by the closed area is determined by calculating the spatial overlap relationship between the closed area and each lane.

[0032] The occupancy length of each lane is calculated as follows: For a specific lane, extract all vehicles, obstacles, and closed areas marked as occupying that lane. Project the bounding rectangles of these occupants onto the longitudinal direction of the lane's centerline to obtain the projection interval of each occupant on the lane's centerline. Union all projection intervals to obtain the total occupancy interval set for that lane. Let the occupancy interval set of lane k be... ,in and Let be the start and end coordinates of the j-th occupied section along the longitudinal direction of the lane, and be the corresponding occupied length of lane k. The calculation formula is as follows: ; in This indicates the number of occupied sections on lane k after merging. The first term is the sum of the lengths of each section, and the second term is the sum of the lengths of the overlapping parts between adjacent sections. By subtracting the overlapping parts, we ensure that the total occupied length after merging accurately reflects the longitudinal span of continuous occupancy.

[0033] In calculating the throughput rate for each lane, a virtual flow detection line is set up at a cross-section downstream of the occupied area where traffic flow is not yet affected by the occupancy. Using millimeter-wave radar or vision-based vehicle detection and tracking results, vehicles crossing this virtual detection line are counted within a preset statistical time window. Let the length of the statistical time window be... The number of vehicles crossing the virtual detection line in lane k within this window is Then the throughput of lane k for: ; If lane k is completely occupied or closed, resulting in no vehicles arriving at the detection line, then Through flow rate =0. For partially occupied lanes, the detection line only covers the unoccupied passable width, and the number of vehicles crossing the detection line reflects the actual throughput capacity of the lane when it is occupied.

[0034] The calculation results of the above-mentioned occupancy length and through flow rate are used as part of the lane-level traffic flow state and are passed to the event cause-effect graph construction module to update the attributes of the regional nodes in the dynamic cause-effect graph and the input parameters of the lane-level traffic management decision module.

[0035] The event cause-effect graph construction module then reads the current processing stage code and expected clearance conditions from the emergency status record. Emergency response status data is pushed in JSON format, with typical fields including: event ID, type (rear-end collision, breakdown, construction, road closure, etc.), alarm time, police arrival time, and current processing stage code (0 - alarm not confirmed, 1 - police en route, 2 - arrived, 3 - investigation underway, 4 - awaiting clearance, 5 - clearance work in progress, 6 - restoration imminent). Figure 5 The diagram shown is the event causality interface of this system. It generates the event causality graph of the current traffic jam event in real time, as well as the expected clearing time. Subsequently, a causal ontology library is built, generating the following node types: Root node: Traffic jam event ( ).

[0036] Cause of the event: such as "rear-end collision" ( "Lane construction" ).

[0037] Vehicle node: Vehicles involved in accidents or malfunctions ( , ...), and severely obstructed key vehicles behind.

[0038] Area node: A closed or obstructed lane area ( , , (These correspond to the first, second, and third lanes, respectively).

[0039] State node: Current processing stage (S).

[0040] Environmental nodes: weather, lighting, etc. (optional).

[0041] Directed edges are created based on the node type and the causal ontology library. For example, "rear-end collision" - [cause] -> "traffic jam"; "rear-end collision" - [occupy] -> "lane area". "; "lane area "-[Cause]->"Traffic Congestion Incident"; "Current Processing Stage" -[Decision]->"Lane Area" The expected release time is set. An influence weight is assigned to the edge, initially 1.0, and adjusted according to the actual evolution.

[0042] Then, every preset time interval (e.g., 10 seconds) or when new emergency data arrives, the attributes of the corresponding nodes in the dynamic cause-effect graph are updated; For example, when the processing stage label changes, the attribute value of the state node S changes; If a tow truck arrives, a new "Tow Truck Node" is added, and an edge is created between it and the "Lane Area" to indicate that clearing is in progress.

[0043] Through the incremental updates described above, the dynamic cause-effect graph fully describes the causal chain of the entire process of a traffic jam event, from its occurrence and handling to its dissipation.

[0044] In the specific identification and prediction process, the stage identification and time prediction module in the intelligent analysis platform uses a temporal graph neural network to reason about the dynamic causal graph, outputting emergency response stage labels and their expected durations. The specific method is as follows: First, let the image snapshot sequence be... ,in ,in This represents the set of nodes at time t. Let represent the set of directed edges at time t. This indicates the preset observation time window length, which is the number of consecutive snapshots involved in a single inference.

[0045] Correspondingly, each node i has an initial feature vector before being input into the network. Initial feature vector The construction method is as follows: For a vehicle node, the initial features include the vehicle type code, the speed normalization value, the one-hot code of the lane number, and a binary identifier indicating whether it is in an occupied state. For a region node, the initial features include the normalized value of the occupancy length, the normalized value of the throughput rate, and the region type encoding; For event state nodes, the initial features include one-hot encoding of the processing phase and normalized values ​​of the duration. For the root node, the initial feature is a fusion of the normalized value of the current queue length and the normalized value of the number of affected lanes.

[0046] All numerical features are processed by min-max normalization, and categorical features are concatenated with numerical features after one-hot encoding to form an initial feature vector of uniform dimension.

[0047] For each time-series graph snapshot, the stage identification and time prediction module first employs a multi-layer graph attention network to spatially aggregate and update the features of each node. The graph attention layer is calculated as follows: ; in This represents the feature vector of node i in the l-th layer. The initial feature vector constructed above, This represents the feature vector of node i after being updated by the l-th layer of the graph attention network. Let represent the set of neighboring nodes of node i. When the graph is directed, it includes all source nodes that have directed edges pointing to node i. Let represent the learnable weight matrix of the l-th layer, used to perform a linear transformation on the features of neighboring nodes. Let be the attention coefficient of neighbor node j to node i, representing the weight of the contribution of node j's features to updating node i's features. The activation function is non-linear, employing an exponential linear unit function, with corresponding attention coefficients. The calculation formula is as follows: ; in This represents the learnable attention vector of the l-th layer. This represents a vector concatenation operation. Represented as a modified linear unit activation function with leakage, it has a small positive slope on the negative half-axis, and the denominator is for all neighboring nodes. Summing the exponential terms normalizes the attention coefficient, making... .

[0048] The graph attention network consists of L stacked layers, each with an independent weight matrix. and attention vector After L layers of graph attention calculation, each node i obtains the final spatial aggregated feature vector. .

[0049] Subsequently, the stage identification and time prediction module uses a readout function to aggregate the feature vectors of all nodes in each timescale snapshot into a single graph-level representation vector. The readout function uses a concatenation of average pooling and max pooling. ; in The graph-level representation vector representing time t. This represents the element-wise average of the final feature vectors of all nodes. It is represented as the element-wise maximum value of the final feature vector of all nodes.

[0050] Perform the above spatial feature extraction and readout operations on T consecutive image snapshots within the observation time window to obtain a sequence of graph-level representation vectors. .

[0051] A sequence of graph-level representation vectors is input into a temporally gated convolutional network (TGN) to capture the evolutionary features of dynamic causal graphs along the time axis. The TGN consists of multiple stacked dilated convolutional layers. Each dilated convolutional layer contains two parallel convolutional branches: one branch uses a hyperbolic tangent function for activation to extract temporal features, and the other branch uses an sigmoid growth curve function for activation to generate a gating signal. The outputs of the two branches are multiplied element-wise to obtain the output of the layer. An increasing dilation factor is applied between adjacent dilated convolutional layers to expand the receptive field without increasing the number of layers. Let the total number of layers in the TGN be K, and the final output of the network be a high-level temporal feature vector. The corresponding expression is as follows: ; in This represents the forward computation function of a temporally gated convolutional network.

[0052] Subsequently, high-level temporal feature vectors They are simultaneously fed into two parallel downstream task networks: The first downstream task is a stage classifier, which outputs the label of the current emergency response stage. The stage classifier consists of a fully connected layer and a flexible maximum normalized exponential function layer. The fully connected layer processes the high-level temporal feature vectors... A linear mapping is used to represent a score vector with dimension equal to the total number of categories in the processing stage. A flexible maximum normalized exponential function transforms this score vector into a predicted probability distribution for each stage. The stage category with the highest probability value is selected as the output emergency processing stage label. Let the set of emergency processing stage categories be... The predicted probabilities for each category are: Then the output stage label is: ; The second downstream task is a duration regressor, which outputs the expected duration. The duration regressor consists of another fully connected layer that integrates the high-level temporal feature vectors. The linear mapping is a scalar value, which is then passed through a modified linear unit activation function to ensure that the output is non-negative. The final output is the expected duration from the current moment, in minutes.

[0053] The training process of the stage identification model employs a multi-task joint training strategy. The training dataset is constructed by collecting dynamic causal graph sequences from historical traffic congestion events and their corresponding actual processing stage labels and actual duration annotations. The loss function consists of a weighted sum of the cross-entropy loss for stage classification and the mean squared error loss for duration regression, as shown in the following expression: ; in The cross-entropy loss for stage classification is calculated as follows: ,in Represented as the one-hot encoded truth value of the stage label; correspond The mean squared error loss representing the duration is calculated as follows: , Indicates the predicted expected duration. Indicates the actual duration. It represents the hyperparameter that balances the weights of classification loss and regression loss.

[0054] Finally, during training, an adaptive moment estimation optimizer is used to update the parameters of the graph attention network, the temporally gated convolutional network, and the fully connected layer through end-to-end gradient backpropagation until the loss function converges, thus obtaining the trained stage recognition model.

[0055] like Figure 6 The diagram shows the current system's stage identification and time prediction interface. Based on the prediction results and analysis, the processing stage and processing time of the current traffic congestion area are obtained in real time. For example, the first stage is the accident handling stage, which is expected to take 25 minutes; the second stage is the vehicle evacuation stage, which is predicted to take 15 minutes; and the third stage is the road restoration stage, which is expected to take 10 minutes.

[0056] After completing the prediction of the expected duration, the lane-level traffic management decision module formulates a lane control plan based on the current emergency response stage and the expected duration.

[0057] The lane-level traffic management decision-making module receives two types of input data: the first is the current emergency response stage label and the corresponding expected duration output by the stage identification and time prediction module; the second is the lane-level traffic flow status generated by the roadside sensing unit and the mobile sensing unit after fusion processing, including the real-time number of vehicles, average speed and occupancy rate of each lane.

[0058] The lane-level traffic management decision module contains three sub-components: a decision variable encoder, a simulation inferrer, and a multi-objective optimization solver. The overall module adopts a rolling time-domain optimization framework for online decision-making.

[0059] The basic process of rolling time-domain optimization is as follows: A fixed decision cycle duration is set. At the beginning of each decision cycle, the latest emergency response stage label, expected duration, and lane-level traffic flow status are obtained. Based on the current state, multi-step predictions of future traffic evolution are made in a microscopic traffic simulation environment. The optimal lane opening / closing scheme and lane-level guidance strategy are solved with the goal of minimizing multiple performance indicators weighted averages. This scheme is then implemented in the next decision cycle. The above process is repeated when the decision cycle ends.

[0060] The corresponding decision variable encoder is used to map the lane control scheme into continuous or discrete variable forms that the optimization algorithm can handle. For a road segment containing M lanes, the decision variables include the open / closed status and speed limit of each lane. The specific encoding method is as follows: For the m-th lane, define a binary variable. Indicates an open state. This indicates that the lane is open to traffic. This indicates that the lane is closed; at the same time, a continuous variable is defined. This indicates the suggested speed limit for that lane, and its value ranges from the preset minimum speed limit. up to the legally mandated maximum speed limit for this section of road Between. Combine the decision variables of all lanes into a vector. In multi-objective particle swarm optimization algorithms, the decision vector is... As the position vector of a particle, the velocity vector of a particle has the same properties as... The same dimension corresponds to the adjustment range of each decision variable. Meanwhile, the simulation simulator is the core of the lane-level traffic management decision-making module for scheme evaluation. The simulator embeds a microscopic traffic simulation environment that perfectly corresponds to the current road segment's geometry, number of lanes, and entrance / exit ramp locations. This simulation environment uses an intelligent driver model to describe vehicle following behavior and a minimized overall braking-induced lane-changing model to describe vehicle lane-changing behavior. In the intelligent driver model, the vehicle... acceleration The formula is given below, and the specific calculation formula is as follows: ; in This indicates the vehicle's maximum acceleration. Indicates vehicle The current speed, This represents the vehicle's desired speed; when there is no speed limit, the free-flow speed is used. This represents the acceleration exponent, which is set to 4 in this design. Indicates vehicle The current distance between the vehicle and the vehicle in front in the same lane. This indicates the speed difference between this vehicle and the vehicle in front. This represents the desired minimum vehicle spacing, and ,in Indicates the safe distance at rest. Indicates the safe headway. Reduce speed for comfort.

[0061] Meanwhile, the lane-change model in this scheme adopts a minimization of the overall braking-induced lane-change model, where the vehicle evaluates whether changing lanes to the adjacent lane can achieve a higher driving speed or a lower local density. The lane-change decision is made through a dual judgment based on an excitation criterion based on acceleration gain and a safety criterion that considers the braking effect of vehicles behind in the adjacent lane.

[0062] The simulation inference system's inputs include: the initial position, speed, and target lane of each vehicle in each lane at the current moment, provided by the lane-level traffic flow state; and the lane open / closed status and speed limit values ​​given by the decision variable encoder. These control measures are reflected in the simulation environment as setting no-entry zones or limiting the maximum speed of vehicles in the corresponding lanes. The simulation inference system updates the state of all vehicles and records key performance indicators at each simulation step. The preset simulation duration is H simulation steps, corresponding to the prediction range in the rolling time domain, typically taken as the upper limit of the expected duration.

[0063] At the end of each simulation run, the simulator outputs the following three performance metrics: The total travel time delay D refers to the sum of the differences between the actual travel time of all affected vehicles during the simulation period and the time required to travel the same distance under free-flow conditions. Assume that N vehicles are recorded in the simulation. For vehicle i, its actual travel time is the time from entering the simulated road segment to leaving or the end of the simulation. The free-flow time is the road segment length divided by the free-flow speed, denoted as . The total travel time delay is: ; The weighted travel time E for emergency vehicles refers to the sum of the actual travel times of emergency vehicles such as breakdown vehicles, ambulances, and police cars within the simulation environment, multiplied by their corresponding priority weights. The set of emergency vehicles is defined as follows: For emergency vehicle j, its actual travel time is recorded as follows: The priority weight is ,but: ; The total number of lane changes, L, refers to the sum of all lane changes completed by all vehicles during the simulation period. Each lane change is triggered by the lane change model, and is counted as one when a vehicle successfully moves completely from one lane into an adjacent lane. This metric reflects the orderliness of traffic flow.

[0064] The multi-objective optimization solver then searches for the optimal lane control scheme in the decision variable space, aiming to minimize the weighted sum of the three performance metrics mentioned above. The solver employs a multi-objective particle swarm optimization algorithm, but in its implementation, to meet real-time requirements, the three objectives are weighted and combined into a single fitness function for scalarization. The fitness function F is constructed as follows: ; in, , as well as The preset weighting coefficients reflect the relative importance of delays, emergency vehicle passage efficiency, and traffic flow stability, respectively, to meet the following requirements. , This represents the maximum total travel time delay among all possible routes for all particles in the current particle swarm, used for normalization. The weighted passage time of the largest emergency vehicle in the particle swarm. This represents the maximum total number of lane changes in the particle swarm.

[0065] Finally, the particle swarm optimization algorithm is used for the search process, which is as follows: At the start of the decision cycle, an initial particle swarm is generated based on the optimal solution of the previous cycle and random perturbations. Each particle's position vector corresponds to a set of decision variable codes. For each particle, a simulation inference tool is invoked to perform microscopic traffic simulation, obtaining three performance indicators and calculating the fitness function value. The particle's velocity and position are updated based on its individual historical best position and global best position. The velocity update formula is as follows: ; ; in This represents the velocity vector of particle i at the k-th iteration. This represents the position vector of particle i at the k-th iteration. This represents the inertia weight, used to balance global exploration and local development capabilities. and These are individual learning factors and social learning factors, respectively. and These are random numbers uniformly distributed between 0 and 1. This represents the historical best position of particle i. This represents the global optimal position of the entire particle swarm.

[0066] A fitness threshold is set, and the iterative process continues until the preset maximum number of iterations is reached or the fitness value improvement is less than the fitness threshold for several consecutive generations. Finally, the decision vector corresponding to the globally optimal particle is selected as the optimal solution for this decision cycle. The optimal solution is decoded into lane opening / closing schemes, i.e., determining the opening state and speed limit of each lane.

[0067] Finally, the lane-level guidance strategy is generated based on the optimized lane opening / closing scheme and the vehicle trajectory distribution output by the simulation inference tool. Specifically, for each vehicle about to enter the controlled section, a suggested lane, suggested lane-changing timing, and target lane are provided based on its current lane and target exit. The lane-changing timing is calculated as follows: along the vehicle's direction of travel, the distance from the starting section of the lane closure ahead is calculated, and then, based on the current speed and downstream lane traffic density, the suggested longitudinal position for initiating the lane change is set to ensure the vehicle has sufficient safe lane-changing space and time. Simultaneously, based on the speed limit and expected travel time for each lane, the expected travel time for each lane is provided for the driver's reference, such as... Figure 7 The diagram shown is a lane-level traffic management decision-making interface, used to display the lane-level guidance strategy for the current congested area. For example, a highway has four lanes, namely... , , as well as ,in The lane where the accident occurred is now closed, and the speed limit is 0. Due to the accident lane Adjacent to each other, the corresponding guiding decision is traffic restriction, with a speed limit of 20km / h. as well as This is a regular lane, the lane is open, and the speed limit is the standard speed of 40km / h.

[0068] By repeating the above solution steps in each decision cycle, dynamic response to traffic congestion incident handling and rolling optimization control of lane-level traffic flow are achieved.

[0069] Finally, the information dissemination unit receives lane opening / closing plans and lane-level guidance strategies output by the lane-level traffic management decision module. The lane-level guidance strategy includes the suggested driving status for each lane, the suggested lane-changing timing and target lane, and the lane-level expected travel time. The information dissemination unit internally includes a message encoding module and a roadside communication unit using vehicle-to-everything (V2X) wireless communication technology.

[0070] The message encoding module encodes lane opening / closing schemes and lane-level guidance strategies into signal phase and lane guidance messages in a preset format. The data structure of this message includes a message identifier, a message generation timestamp, the coordinates of the start and end points of the controlled road segment, the total number of lanes within the segment, and a list of guidance information for each lane. Each item in the lane guidance information list corresponds to a lane and specifically includes fields for lane number, lane status, suggested speed limit, lane change prompt type, suggested lane change starting distance, and expected travel time. The lane status field has three possible values: open, closed, and suggested departure. When the lane status is closed, the lane change prompt type field indicates the lane number of the target lane, and the suggested lane change start distance field gives the longitudinal distance between the current vehicle and the starting point of the lane closure section. When the lane status is suggested departure, the lane change prompt type field indicates the reference point for the recommended lane change timing. When the lane status is open, the lane change prompt type field is empty, the suggested speed limit field gives the maximum speed allowed in this lane, and the expected travel time field gives the estimated time required for the vehicle to pass through the controlled section under the suggested speed limit conditions.

[0071] The roadside communication unit (Roadside Communication Unit) of the vehicle wireless communication technology wirelessly transmits the encoded signal phase and lane guidance message to all vehicle terminals within its coverage area via a direct communication interface at a preset broadcast period. The broadcast period is a fixed value between 100 milliseconds and 500 milliseconds. The vehicle terminals receive the signal phase and lane guidance message through the corresponding direct communication interface. The message decoding module parses the message according to the reverse process of the message encoding module to recover all fields of the lane opening / closing scheme and lane-level guidance strategy.

[0072] The vehicle terminal determines its current lane based on its own positioning information and indexes guidance information items corresponding to the current lane and adjacent lanes from the parsed results. The vehicle terminal transmits the indexed guidance information items to the human-machine interface control module, which drives the navigation display screen or head-up display device to draw lane-level guidance graphics. The drawing method is as follows: a top-down view of the lane model of the controlled road section ahead is embedded in the map interface of the navigation display screen. Each lane is distinguished by color: open lanes are rendered with green fill, lanes recommended to be exited are rendered with yellow fill, and closed lanes are rendered with red fill. For lanes that require lane changes, a lane change arrow graphic with directional indicators is superimposed above the lane model, with the arrow pointing to the target lane of the suggested lane change. At the same time, the suggested speed limit and expected travel time are marked with text below the lane model. The head-up display device uses a similar color coding scheme, projecting simplified lane status bars and lane change arrow symbols on the windshield, and updating the remaining distance from the lane change starting section in digital form in real time.

[0073] In one specific embodiment, a two-vehicle rear-end collision occurs on a three-lane highway, blocking the second middle lane. Roadside cameras capture images of the accident, and radar detects a sudden drop in vehicle speed. The emergency status acquisition unit receives an alarm message and a "police officers are on the scene" status. After the dynamic causal graph is initially generated, the stage identification model outputs a "Investigation and Evidence Collection in Progress" label, with an expected duration of 14 minutes. The traffic management decision module tests various solutions in a micro-simulation: closing only the second lane and limiting the speed of the first and third lanes; closing the second lane and opening the hard shoulder as a temporary third lane; adjusting the speed limits of each lane, etc. Through a multi-objective particle swarm optimization algorithm, the final solution is to close the second lane, limit the speed of the first lane to 80 km / h, the third lane to 60 km / h, and guide vehicles in the second lane to change lanes to the third lane in advance. The information dissemination unit immediately updates the content displayed on the roadside display screen and the vehicle terminal, and drivers of vehicles behind change lanes in an orderly manner according to the lane-level guidance. Throughout the event handling process, the system continuously updates and optimizes the causal graph based on the emergency status update (e.g., "Clearance work in progress"), dynamically adjusting the traffic management plan until full traffic is restored.

[0074] This invention utilizes an intelligent analysis platform to construct a dynamic causal graph with traffic congestion events as the root node based on multimodal real-time data and emergency response status data. It employs a stage identification model to infer emergency response stages and their expected durations, and generates lane opening / closing plans and guidance strategies through a lane-level traffic management decision-making module. An information dissemination unit then distributes these plans to vehicles. Through causal graph evolution prediction and lane-level dynamic traffic management, the congestion situation in congested areas is updated in real time, providing precise lane-level traffic guidance.

[0075] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-driven intelligent road condition recognition system, characterized in that, It includes roadside sensing units, mobile sensing units, emergency status acquisition units, and intelligent analysis platforms; The roadside sensing unit is used to collect multimodal real-time data from a fixed perspective in the traffic jam area; The mobile sensing unit is deployed on a drone that can fly over the traffic jam area and is used to supplement the collection of multimodal real-time data of the blind spots covered by the roadside sensing unit under the control of the received dispatch instructions. The emergency status acquisition unit is used to acquire emergency response status data from at least one of the traffic management platform, the emergency call platform, and the construction reporting platform. The intelligent analysis platform includes an event cause-effect graph construction module, a stage identification and time prediction module, a lane-level traffic management decision module, and an information release unit. The event causal graph construction module is used to extract event elements and construct a dynamic causal graph with the current traffic jam event as the root node based on the multimodal real-time data and the emergency response status data. The nodes of the dynamic causal graph represent entities or states that have a causal relationship with the traffic jam event, and the edges represent the causal associations and time sequence effects between nodes. The stage identification and time prediction module is used to perform state reasoning on the dynamic causal graph using a trained stage identification model, and output emergency response stage labels and their expected durations. The lane-level traffic management decision module is used to input the emergency handling stage label and expected duration, and the lane-level traffic flow status into the multi-objective optimizer, and solve the lane opening / closing scheme and lane-level guidance strategy in a rolling time-domain optimization manner within a micro traffic simulation environment. The information dissemination unit is used to send the lane opening / closing scheme and lane-level guidance strategy to vehicles affected by traffic congestion.

2. The AI-driven intelligent road condition recognition system according to claim 1, characterized in that, The roadside sensing unit includes surveillance cameras and millimeter-wave radars deployed at intervals on both sides of the road or on gantry frames; The surveillance camera outputs visible light image data in the form of a continuous video stream; The millimeter-wave radar synchronously outputs point cloud data containing the position, velocity, and echo intensity of the moving target.

3. The AI-driven intelligent road condition recognition system according to claim 1, characterized in that, The method for supplementing the collection of multimodal real-time data of the blind spots covered by the roadside sensing unit in the mobile sensing unit includes the following steps: S1. The quality of the multimodal data transmitted in real time by the roadside sensing unit is evaluated through the intelligent analysis platform to obtain the average confidence level of target detection within the coverage area of ​​each roadside sensing device; S2. When the average confidence level of a certain sub-region is lower than the preset threshold or no moving target is detected for three consecutive collection cycles, the sub-region is determined to be a perception blind zone. S3. Extract the coordinate sequence of the polygon boundary vertices of the perception blind zone, the coordinates of the center point, and the required supplementary acquisition spatial resolution. S4. The intelligent analysis platform encapsulates the blind spot information into a sensing coverage requirement message and sends it to the mobile sensing unit through the internal communication link. S5. After receiving the sensing coverage request message, the mobile sensing unit reads the status information of all available drones in the current drone nest from the drone status database, including the unique identifier of each drone, current geographical coordinates, current remaining battery power, and lens focal length and pixel size of the camera. S6. The mobile sensing unit aims to minimize the weighted sum of the lack of coverage in the sensing blind spot and the total energy consumption cost of the UAV swarm, and constructs a joint optimization model for UAV task assignment and waypoint parameters. S7. Solve the joint optimization model of UAV mission assignment and waypoint parameters; S8. The mobile sensing unit generates scheduling instructions based on the solution results. Each scheduling instruction includes the UAV identifier, target waypoint longitude, target waypoint latitude, flight altitude, gimbal pitch angle, and hovering duration. S9. The scheduling instructions are sent to the corresponding UAVs via wireless communication links to perform blind spot flight tasks.

4. The AI-driven intelligent road condition recognition system according to claim 1, characterized in that, The emergency response status data in the emergency status acquisition unit includes, but is not limited to: The traffic management platform provides information on the type of traffic accident, its precise location, the lane number occupied, the time the traffic police arrived, the start time of the investigation, the status of the clearing vehicles, and the expected time of traffic restoration. The emergency call platform provides information on the nature of the accident described by the caller, the time of the call, the number of vehicles involved and their license plate information, and the status of the police force dispatched after receiving the call. The construction reporting platform provides information on the construction permit number, the start and end chainage of the construction section, the lanes occupied by the construction, the start time of the construction, the planned end time, and whether the current construction activity is suspended or in progress.

5. The AI-driven intelligent road condition recognition system according to claim 1, characterized in that, The method for constructing a dynamic cause-effect graph with the current traffic congestion event as the root node in the event cause-effect graph construction module includes the following steps: S20. Identify vehicles, obstacles, and closed areas occupying lanes from the perception fusion results, and calculate the occupancy length and throughput rate of each lane; S21. The calculation results of the occupancy length and through flow rate are used as part of the lane-level traffic flow state and are passed to the event cause-effect graph construction module to update the attributes of the regional nodes in the dynamic cause-effect graph and the input parameters of the lane-level traffic management decision module. S22. The event cause-effect graph construction module reads the current processing stage code and expected clearing conditions from the emergency status record; S23. Construct a causal ontology library and generate the corresponding node types; S24. Based on the type of node constructed, establish directed edges according to the causal ontology library, and assign influence weights to the directed edges. S25. Update the attributes of the corresponding nodes in the dynamic cause-effect graph according to the preset time or when new emergency state data arrives, and generate the final dynamic cause-effect graph.

6. The AI-driven intelligent road condition recognition system according to claim 5, characterized in that, The node types in S23 include root nodes, cause event nodes, vehicle nodes, region nodes, status nodes, and environment nodes.

7. The AI-driven intelligent road condition recognition system according to claim 6, characterized in that, The method for outputting emergency response stage labels and their expected durations in the stage identification and time prediction module includes the following steps: S200, the construction graph snapshot sequence is as follows ,in ,in This represents the set of nodes at time t. Let represent the set of directed edges at time t. Indicates the preset observation time window length; S201. Construct corresponding initial feature vectors based on different node types. ; S202. For each moment of the graph snapshot, the stage identification and time prediction module uses a multi-layer graph attention network to aggregate and update the features of each node in the spatial dimension, and outputs the feature vector of the node after being updated by different numbers of layers of graph attention network. S203, the stage identification and time prediction module uses a readout function to aggregate the feature vectors of all nodes in each time-map snapshot into a single graph-level representation vector. The readout function uses a combination of average pooling and max pooling. ; S204. Perform the above spatial feature extraction and readout operations on the T consecutive snapshots within the observation time window to obtain the graph-level representation vector sequence. ; S205. Input the graph-level representation vector sequence into a temporally gated convolutional network to capture the evolutionary features of the dynamic causal graph along the time axis. The final output of the output network is a high-level temporal feature vector. ; S206, Transfer the high-level temporal feature vector Simultaneously fed into two parallel downstream task networks; The first downstream task is a stage classifier, which outputs the label of the current emergency response stage. ; The second downstream task is a duration regressor, which is used to output the expected duration.

8. The AI-driven intelligent road condition recognition system according to claim 1, characterized in that, The lane-level traffic management decision module includes three sub-components: a decision variable encoder, a simulation inferrer, and a multi-objective optimization solver. The decision variable encoder is used to map lane control schemes into continuous or discrete variable forms that can be processed by optimization algorithms. The simulation inference engine uses an intelligent driver model to describe vehicle following behavior and a minimization of the overall braking-induced lane-changing model to describe vehicle lane-changing behavior, and outputs the following three performance indicators: Total travel time delay D refers to the sum of the differences between the actual travel time of all affected vehicles during the simulation period and the time required to travel the same distance under free-flow conditions; The weighted travel time E for emergency vehicles refers to the sum of the actual travel time of emergency vehicles such as breakdown vehicles, ambulances, and police cars in the simulation environment, multiplied by their corresponding priority weights. The total number of lane changes L refers to the sum of all lane changes completed by all vehicles during the simulation period; The multi-objective optimization solver aims to minimize the weighted sum of the three performance indicators mentioned above and searches for the optimal lane control scheme in the decision variable space.

9. The AI-driven intelligent road condition recognition system according to claim 8, characterized in that, The method for searching the optimal lane control scheme in the multi-objective optimization solver includes the following steps: S30. The three objectives are weighted and combined into a single fitness function for scalarization. The corresponding fitness function F is constructed as follows: ； in, , as well as The preset weighting coefficients reflect the relative importance of delays, emergency vehicle passage efficiency, and traffic flow stability, respectively, to meet the following requirements. , This represents the maximum total travel time delay among all possible routes for all particles in the current particle swarm, used for normalization. The weighted passage time of the largest emergency vehicle in the particle swarm. This represents the maximum total number of lane changes in the particle swarm; S31. Search processing is performed using the particle swarm optimization algorithm; S32. At the beginning of the decision cycle, an initial particle swarm is generated based on the optimal solution of the previous cycle and the random perturbation. The position vector of each particle corresponds to a set of decision variable codes. S33. For each particle, call the simulation inference tool to perform microscopic traffic simulation, obtain three performance indicators and calculate the fitness function value, and update the particle's velocity and position based on the particle's individual historical best position and global best position. S34. Set a fitness threshold and continue the iteration process until the preset maximum number of iterations is reached or the fitness value improvement is less than the fitness threshold for several consecutive generations. S35. Select the decision vector corresponding to the globally optimal particle as the optimal solution for this decision cycle, and decode the optimal solution into a lane opening / closing scheme.

10. The AI-driven intelligent road condition recognition system according to claim 1, characterized in that, The information publishing unit is equipped with a message encoding module and a roadside communication unit for vehicle wireless communication technology. The message encoding module encodes the lane opening / closing scheme and lane-level guidance strategy into a preset format of signal phase and lane guidance message; The roadside communication unit of the vehicle wireless communication technology wirelessly transmits the encoded signal phase and lane guidance message to all vehicle terminals within the coverage area of ​​the roadside communication unit through a direct communication interface at a preset broadcast period.

Images