Traffic intention recognition control method and system based on traffic scene understanding

Abstract

The application provides a traffic intention recognition control method and system based on traffic scene understanding, and relates to the technical field of traffic control. The method comprises the following steps: acquiring motion trajectory and traffic scene data, performing semantic segmentation on the scene data to extract topological graph nodes and label attributes, establishing a directed connection edge, and generating a scene semantic graph through a graph neural network; anchoring the trajectory data to the graph node sequence, extracting path orientation features through a graph convolutional neural network, calculating geometric changes for the trajectory and interactive state time slices, and judging the game state to extract interactive game features; combining the two features to generate an intention candidate set, applying Bayesian inference to obtain a traffic intention recognition result; determining a conflict area according to the recognition result, constructing a multi-agent game state space and an action set, constructing a cumulative reward function, and solving an optimal game strategy through strategy iteration to generate a traffic control strategy, and executing traffic control based on the strategy.

Description

Technical Field

[0001] This invention relates to the field of traffic control technology, and in particular to a method and system for recognizing and controlling traffic intentions based on traffic scene understanding. Background Technology

[0002] Traffic intention recognition and control methods typically rely on the historical movement trajectories and instantaneous motion states of traffic participants for prediction. Common methods utilize motion models such as Kalman filtering or particle filtering to make short-term estimates of the future positions of vehicles or pedestrians, and use Hidden Markov Models or Recurrent Neural Networks to capture temporal change patterns to infer the intention category, such as going straight, turning, or stopping. Alternatively, machine learning classifiers such as decision trees or support vector machines are used, combined with features such as vehicle speed, acceleration, and turn signal signals to determine the intention. At the control level, rule-based methods are often employed, such as setting safe distance thresholds or conflict zone boundaries to trigger deceleration or avoidance actions. Some solutions also incorporate reinforcement learning frameworks to train strategies through simulated environments, but intention recognition and control are usually decoupled into independent modules.

[0003] The above methods have significant drawbacks. They lack a structural understanding of the semantic hierarchy of traffic scenarios, such as road topology connections, traffic sign meanings, and lane function divisions. Relying solely on motion trajectories and low-level features makes it difficult to accurately identify intentions at complex intersections or open areas, because the same motion pattern may correspond to different intentions in different scenarios; for example, slowing down when approaching an intersection may indicate preparation to stop or turn. Most methods ignore the interactive game relationships between multiple traffic participants, treating other vehicles or pedestrians as static obstacles or independent moving objects. In real-world scenarios, each party dynamically adjusts its intentions based on the other's behavior. This game process is not incorporated into a unified framework for intention recognition and control, resulting in traffic control strategies that are either too conservative and inefficient, or too aggressive and pose safety hazards. Summary of the Invention

[0004] The present invention provides a traffic intention recognition and control method and system based on traffic scene understanding, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a traffic intention recognition and control method based on traffic scene understanding, comprising: Acquire movement trajectory data and traffic scene data of the target traffic participants; The traffic scene data is semantically segmented to extract topological graph nodes, and the semantic attributes of the nodes are labeled. Directed connecting edges are established and the semantic attributes of the edges are labeled. A scene semantic graph is generated through a graph neural network. The motion trajectory data is anchored to the node sequence of the scene semantic graph. The spatial propagation features of the node sequence are extracted as path guidance features using a graph convolutional neural network. The motion trajectory data and interaction state attributes are temporally segmented to calculate the change in interaction geometric relationship, determine the interaction game state, and extract interaction game features. The path guidance features and interactive game features are used to generate an intent candidate set. The posterior probability distribution of the intent of the target traffic participants is solved by a Bayesian inference network to obtain the traffic intent recognition result. Based on the passage intention recognition result, identify the conflict area, construct a multi-agent game state space and define the action set, construct a cumulative reward function and use strategy iteration to solve the optimal game strategy, and generate a passage control strategy. Traffic control is implemented on the target traffic participants based on the traffic control strategy.

[0006] In one optional embodiment, the traffic scene data is semantically segmented to extract topological graph nodes, and the semantic attributes of the nodes are labeled. Directed connecting edges are established and their semantic attributes are labeled. A scene semantic graph is generated through a graph neural network, including: Semantic segmentation is performed on traffic scene data to identify road areas, lane areas, and intersection areas. The boundary contour lines of road areas and the center lines of lane areas are extracted. The intersection of the boundary contour lines and the center lines is used as topology graph nodes. Conflict area identifiers are marked according to whether the topology graph node is located in an intersection area. The number of associated lanes, turning type attributes, and conflict area identifiers are marked for each topology graph node to generate node semantic attributes. Detect the current position of other traffic participants in the dynamic layer, determine the topology graph node occupied by each other traffic participant, establish directed connection edges between the topology graph node occupied by other traffic participants and the adjacent topology graph node in the direction of movement, label the directed connection edges with the instantaneous speed and direction of movement angle of other traffic participants, and generate edge semantic attributes. Using node semantic attributes as the initial node feature matrix and edge semantic attributes as the adjacency relation matrix, node features are propagated through a graph neural network, enabling each topological graph node to aggregate the feature information of its neighboring nodes and generate a scene semantic graph.

[0007] In one optional embodiment, anchoring the motion trajectory data to a node sequence of the scene semantic graph, and extracting the spatial propagation features of the node sequence as path guidance features using a graph convolutional neural network includes: Spatial coordinate analysis is performed on the trajectory points in the motion trajectory data. The spatial distance between each trajectory point and each node in the scene semantic graph is calculated. The node with the smallest distance is selected as the anchor node, and a mapping table from trajectory points to anchor nodes is established. Based on the mapping relationship, all anchor nodes passed by the target traffic participants are extracted, and the anchor nodes are arranged in time sequence according to the timestamp to generate an ordered node sequence. The spatial connectivity attribute of each node in the ordered node sequence and the topological edge weight between adjacent nodes are extracted from the scene semantic graph. A multi-layer graph convolutional neural network is constructed, which takes an ordered sequence of nodes and their topological edge weights as the graph structure input. In the graph convolutional layer, the spatial connectivity attributes of the current node and the spatial connectivity attributes of the neighboring nodes are aggregated by message passing. The aggregated information is weighted and fused using the topological edge weights to generate the node hidden layer features of the fused neighborhood structure. Through layer-by-layer propagation of multi-layer graph convolutional layers, each node in the ordered node sequence is sequentially fused with the topological structure information of its one-hop to multi-hop neighborhood, and the spatial propagation feature vector of each node is output. The spatial propagation feature vectors of each node in the ordered node sequence are concatenated according to the temporal order to form path-guided features.

[0008] In one optional embodiment, the motion trajectory data and interaction state attributes are time-series segmented to calculate the change in interaction geometric relationship, determine the interaction game state, and extract interaction game features, including: The motion trajectory data is divided into time windows to generate multiple time segments. Interaction state attributes corresponding to each time segment are extracted from the scene semantic graph, and a time-series mapping table between time segments and interaction state attributes is established. For each time segment, extract the position coordinates and velocity vectors of the motion trajectory data, extract the position coordinates and velocity vectors of other traffic participants for the corresponding interaction state attributes, calculate the relative displacement vector, relative velocity vector and motion direction angle between the target traffic participant and each other traffic participant, and construct the interaction geometric relationship feature group; The difference of the interactive geometric relationship feature groups of adjacent time segments is calculated to obtain the sequence of changes in interactive geometric relationship. The change in relative displacement vector is decomposed in direction to extract the radial approach rate and tangential circumduction rate. Construct an interactive game state discriminator, calculate the ratio of radial approach rate to tangential detour rate, determine the game tendency coefficient, when the game tendency coefficient is greater than the preset tendency threshold, mark the competitive game state and extract the change in relative velocity vector as the competitive feature component, otherwise mark the avoidance game state and extract the change in the angle between the motion directions as the avoidance feature component. The competitive and avoidance feature components of all time segments are concatenated in time sequence to generate interactive game features.

[0009] In one optional embodiment, the path guidance features and interactive game features are used to generate an intent candidate set. A Bayesian inference network is then used to solve for the posterior probability distribution of the target traffic participants' intentions, resulting in the following traffic intention recognition results: By concatenating path-oriented features and interactive game features, spatial propagation feature vectors from path-oriented features and competition and avoidance feature components from interactive game features are extracted to construct a joint feature representation. Construct an intent candidate space, generate a candidate path set based on the topological connection relationship of each node in the spatial propagation feature vector, generate an interaction strategy set based on the numerical distribution of the competition feature component and the avoidance feature component, and combine the candidate path set and the interaction strategy set to form an intent candidate set. Construct a Bayesian inference network, and calculate the path conditional probability and policy conditional probability for each intent candidate in the intent candidate set based on the joint feature representation; A probability fusion layer is constructed, and the path conditional probability and the policy conditional probability are input into the probability fusion layer for joint probability inference. The joint conditional probability is generated through the coupling propagation of path space constraints and interactive game constraints. The joint conditional probability of each intention candidate in the intention candidate set is normalized to generate the posterior probability distribution of the intention of the target traffic participant. The candidate intent with the highest probability value is selected from the posterior probability distribution of intent as the result of the passage intent recognition.

[0010] In one optional embodiment, conflict areas are identified based on the passage intention recognition results, a multi-agent game state space is constructed and an action set is defined, a cumulative reward function is constructed and the optimal game strategy is solved using policy iteration, and a passage control strategy is generated, including: Based on the traffic intention recognition results, the expected travel path of the target traffic participant is extracted, and spatial overlap detection is performed with the expected travel paths of other traffic participants to identify conflict areas. A multi-agent game state space is constructed. The position, speed, and heading of the target traffic participant in the conflict area are used as the state of the first agent, and the position, speed, and heading of other traffic participants are used as the state of the second agent. The spatial boundary constraints of the conflict area are extracted to construct the game state space. Define sets of actions for acceleration, deceleration, and maintenance for the first agent state and the second agent state respectively, and establish a transition mapping relationship between state and action; A multi-agent reinforcement learning reward mechanism is constructed, which generates a safety reward component based on the relative position vector and relative speed vector of the target traffic participant and other traffic participants, an efficiency reward component based on the travel time, and a collision penalty component based on the predicted trajectory intersection, and constructs a cumulative reward function. The optimal game strategy is solved by iteratively performing action selection and state transition in the game state space, updating the decision strategy based on the cumulative reward function, and using the strategy iteration to perform action selection and state transition in the game state space. The action sequence is extracted based on the optimal game strategy, converted into passage time adjustment instructions and spatial trajectory adjustment instructions, and a passage control strategy is generated.

[0011] In one optional embodiment, a multi-agent reinforcement learning reward mechanism is constructed, generating a safety reward component based on the relative position vector and relative velocity vector of the target traffic participant and other traffic participants, an efficiency reward component based on the travel time, and a collision penalty component based on the predicted trajectory intersection. The cumulative reward function includes: Extract the relative position vector and relative velocity vector between the target traffic participant and other traffic participants. Calculate the braking distance based on the relative velocity vector. If the magnitude of the relative position vector is greater than the braking distance, it is determined that the safe distance condition is met, and a safety reward component is generated. Record the time when the target traffic participant enters the conflict area and the time when the target traffic participant leaves the conflict area, calculate the time difference as the passage time, and generate an efficiency reward component when the passage time is less than a preset benchmark time. Calculate the predicted trajectories of the target traffic participant and other traffic participants within a future time window, determine whether the predicted trajectories intersect, and if so, generate a collision penalty component based on the time interval at the intersection. The safety reward component, the efficiency reward component, and the collision penalty component are weighted and summed to construct a cumulative reward function.

[0012] A second aspect of the present invention provides a traffic intention recognition and control system based on traffic scene understanding, comprising: The data acquisition unit is used to acquire the movement trajectory data and traffic scene data of the target traffic participants; The scene semantic unit is used to perform semantic segmentation on the traffic scene data to extract topological graph nodes, label the semantic attributes of the nodes, establish directed connection edges and label the semantic attributes of the edges, and generate a scene semantic graph through a graph neural network. The trajectory feature unit is used to anchor the motion trajectory data to the node sequence of the scene semantic graph, extract the spatial propagation features of the node sequence as path guidance features using a graph convolutional neural network, perform temporal segmentation of the motion trajectory data and interaction state attributes to calculate the change in interaction geometric relationship, determine the interaction game state, and extract interaction game features. The intent recognition unit is used to generate an intent candidate set from path guidance features and interactive game features, and solve the posterior probability distribution of the intent of the target traffic participants through a Bayesian inference network to obtain the traffic intent recognition result. The game strategy unit is used to identify conflict areas based on the passage intention recognition result, construct a multi-agent game state space and define an action set, construct a cumulative reward function and use strategy iteration to solve the optimal game strategy, and generate a passage control strategy. A traffic control unit is used to perform traffic control on the target traffic participants based on the traffic control strategy.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] In this embodiment of the invention, by semantically segmenting traffic scene data and constructing a scene semantic graph with directed edges, and combining it with a graph neural network to extract path guidance features from node sequences, the system can accurately capture the road topology and the spatial propagation patterns of participants, effectively avoiding misjudgments caused by neglecting scene semantic relationships in traditional methods. In complex dynamic scenarios, the calculation of temporal segmentation interaction geometric relationship changes and the discrimination of interaction game states profoundly reveal the game evolution essence between the target traffic participants and their surrounding environment. The extracted interaction game features provide a highly discriminative basis for subsequent inference. By using a Bayesian inference network to fuse path guidance features and interaction game features, the posterior probability distribution of the target intent is solved, significantly reducing the uncertainty caused by a single feature, thus making the travel intent more accurate. The identification results possess probabilistic interpretability and statistical confidence, making them particularly suitable for multimodal behavior prediction scenarios. Based on accurate intent recognition results, conflict areas are identified and a multi-agent game state space is constructed. Combined with a cumulative reward function and iterative strategy solving, the generated optimal game strategy achieves proactive avoidance of conflict risks in the spatiotemporal dimensions, balancing traffic efficiency and safety. A complete closed loop is formed from scene understanding to behavioral decision-making, adapting to complex traffic environments such as urban intersections and highway ramps without the need for manually preset rules. The traffic control strategy at the execution level has real-time adaptive capabilities, effectively responding to sudden situations such as pedestrians crossing the road suddenly or vehicles forcibly changing lanes, significantly reducing the rate of sudden braking or collisions caused by intent misjudgment, while simultaneously improving the overall road network traffic efficiency to the global optimal level. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a traffic intent recognition and control method based on traffic scenario understanding. Figure 2 Flowchart of the logic for feature extraction in interactive games; Figure 3 The temporal evolution and state discrimination diagram of interactive game features; Figure 4 Decomposition diagram of the cumulative reward function for multi-agent reinforcement learning. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0019] Figure 1 This is a flowchart illustrating the traffic intent recognition and control method based on traffic scene understanding, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire movement trajectory data and traffic scene data of the target traffic participants; The traffic scene data is semantically segmented to extract topological graph nodes, and the semantic attributes of the nodes are labeled. Directed connecting edges are established and the semantic attributes of the edges are labeled. A scene semantic graph is generated through a graph neural network. The motion trajectory data is anchored to the node sequence of the scene semantic graph. The spatial propagation features of the node sequence are extracted as path guidance features using a graph convolutional neural network. The motion trajectory data and interaction state attributes are temporally segmented to calculate the change in interaction geometric relationship, determine the interaction game state, and extract interaction game features. The path guidance features and interactive game features are used to generate an intent candidate set. The posterior probability distribution of the intent of the target traffic participants is solved by a Bayesian inference network to obtain the traffic intent recognition result. Based on the passage intention recognition result, identify the conflict area, construct a multi-agent game state space and define the action set, construct a cumulative reward function and use strategy iteration to solve the optimal game strategy, and generate a passage control strategy. Traffic control is implemented on the target traffic participants based on the traffic control strategy.

[0020] In one optional embodiment, the traffic scene data is semantically segmented to extract topological graph nodes, and the semantic attributes of the nodes are labeled. Directed connecting edges are established and their semantic attributes are labeled. A scene semantic graph is generated through a graph neural network, including: Semantic segmentation is performed on traffic scene data to identify road areas, lane areas, and intersection areas. The boundary contour lines of road areas and the center lines of lane areas are extracted. The intersection of the boundary contour lines and the center lines is used as topology graph nodes. Conflict area identifiers are marked according to whether the topology graph node is located in an intersection area. The number of associated lanes, turning type attributes, and conflict area identifiers are marked for each topology graph node to generate node semantic attributes. Detect the current position of other traffic participants in the dynamic layer, determine the topology graph node occupied by each other traffic participant, establish directed connection edges between the topology graph node occupied by other traffic participants and the adjacent topology graph node in the direction of movement, label the directed connection edges with the instantaneous speed and direction of movement angle of other traffic participants, and generate edge semantic attributes. Using node semantic attributes as the initial node feature matrix and edge semantic attributes as the adjacency relation matrix, node features are propagated through a graph neural network, enabling each topological graph node to aggregate the feature information of its neighboring nodes and generate a scene semantic graph.

[0021] In one specific implementation, when performing semantic segmentation on traffic scene data, the input raw data sources include image sequences captured by vehicle-mounted cameras, LiDAR point cloud projection maps, and high-precision map raster data. The above multi-source data is fused and fed into a semantic segmentation network, assigning a semantic category label to each pixel or point cloud voxel. Three types of regions are identified: road regions (including drivable road surfaces), lane regions (including single-lane driving areas), and intersection regions (including intersection merging areas). The semantic segmentation network employs an encoder-decoder structure. The encoder compresses and extracts multi-scale spatial features layer by layer, while the decoder upsamples layer by layer to restore pixel-level semantic labels, ultimately outputting a semantically labeled map with the same resolution as the input.

[0022] After obtaining the semantic annotation map, boundary contours are extracted for road regions, and centerlines are extracted for lane regions. The boundary contour extraction method involves performing Canny edge detection on the semantic mask of the road region, then fitting continuous contour segments using Hough line transform to obtain several polyline segments representing the road boundary. The lane centerline extraction method involves skeletonizing the semantic mask of each lane region, i.e., performing morphological thinning on the mask, and connecting the central axis pixels of the thinned result into a polyline, which is the lane centerline. The geometric intersections of all boundary contours and lane centerlines are selected as a candidate set of topology map nodes. The selection rule is: if the Euclidean distance between two line segments is less than a preset merging threshold, they are merged into the same node to avoid duplicate nodes due to pixel-level errors. Finally, several discrete topology map nodes are obtained, each node corresponding to a spatial location in the scene with road structural significance.

[0023] The semantic attributes of nodes are labeled in three dimensions. The first dimension is the number of associated lanes. For each node in the topology graph, the number of center lines of the lanes adjacent to it within its surrounding space is counted. This number directly reflects the road capacity at the node's location. The second dimension is the turning type attribute. Based on the angle between the direction vectors of the center lines of the lanes connected to the node, the driving direction at the node is determined, and the turning type is divided into four types: straight, left turn, right turn, and U-turn. Specifically, the determination method is to take the angle between the upstream and downstream direction vectors of the node. ,when The time mark is straight, when The time marker indicates a right turn, when The time marker indicates a left turn, when The time marker is a U-turn. The third dimension is the conflict area identifier, which is determined by whether the pixel coordinates of the topology graph node fall within the semantic mask range of the intersection area: if the node coordinates are within the intersection area, the conflict area identifier is assigned a value of 1; otherwise, it is assigned a value of 0. The combination of the attributes of the three dimensions constitutes the node semantic attribute vector, which serves as the initial node features for subsequent graph neural network processing.

[0024] The construction of the dynamic layer relies on the real-time detection of other traffic participants in the scene. An object detection network processes the current frame image or point cloud, outputting the center coordinates, category label (pedestrian, non-motorized vehicle, motorized vehicle, etc.), and bounding box for each other traffic participant. For each detected other traffic participant, the Euclidean distance from its center coordinates to all topology graph nodes is calculated. The node closest to it and less than the occupancy threshold is marked as the topology graph node currently occupied by that traffic participant. If the center coordinates of a traffic participant are beyond the occupancy threshold for all nodes, it is projected onto the nearest lane centerline, and the nearest node on the lane centerline where the projection point is located is taken as its occupied node, thus ensuring that dynamic information can be effectively mapped onto the static topology structure.

[0025] The directed connection edges are established as follows: For each other traffic participant, calculate its instantaneous velocity vector based on the difference in center coordinates between the current frame and several previous frames. The magnitude of the instantaneous velocity is denoted as... The direction and angle of motion are denoted as , Defined as the clockwise angle between the velocity vector and the due east direction of the coordinate system. A directed edge is established between the currently occupied topological graph node and its nearest neighboring topological graph node in the direction of movement, with the direction consistent with the direction of movement. The "nearest neighboring topological graph node in the direction of movement" refers to the topological graph node directly connected to the current node, whose direction vector is... The one with the smallest angle. If the time interval between the current frame and the historical frame is... The magnitude of the displacement vector is Then the instantaneous velocity magnitude .Will and These are collectively used as the semantic attributes of the directed connection edge and written into the corresponding positions in the adjacency matrix. For edges in a static road structure that do not have dynamic traffic participants, their... Assign a value of 0, Retain the default road direction to ensure the integrity of the adjacency matrix.

[0026] After constructing the node semantic attributes and edge semantic attributes, the node semantic attributes are organized into an initial node feature matrix, and the edge semantic attributes are organized into a weighted adjacency relation matrix. These are then fed into the graph neural network to perform node feature propagation. Each layer of the graph neural network follows a message-passing mechanism: for each target node, the source node features and edge attributes carried by all directed edges ending at that node are collected, weighted, aggregated, and fused with the target node's own features. The hidden layer representation of the target node is then updated using a non-linear activation function. The weighting coefficients are derived from the edge semantic attributes... After normalization, it is determined that dynamic edges with higher speeds have a higher influence weight on the target node, thus reflecting the strong influence of dynamic traffic participants on the local scene. After propagation through a multi-layer graph neural network, the representation vector of each topological graph node simultaneously contains the node's own road structure semantic information (number of associated lanes, turning type, conflict area identifier) ​​and dynamic traffic state information from neighboring nodes (speed and direction of movement of other traffic participants). Combining the final representation vectors of all nodes yields a complete representation of the scene semantic graph. This graph retains the geometric attributes of the road topology and integrates dynamic interaction states, providing a structured scene representation foundation for subsequent path guidance feature extraction and interactive game feature calculation.

[0027] In practical engineering implementations, to ensure the stability of graph neural network propagation, it is necessary to normalize the values ​​in the initial node feature matrix and adjacency relation matrix. The number of associated lanes in the node features is normalized by dividing by the maximum number of lanes in the scene. Steering type is represented by one-hot encoding, and conflict zone identifiers are directly used. Binary. In the semantic attributes of edges. Normalized by dividing by the scene speed limit value, Convert to values ​​in radians and normalize to The above normalization process enables the graph neural network to maintain stable feature propagation performance in scenarios with different traffic densities and speed ranges, ultimately generating a scene semantic graph with high semantic consistency.

[0028] In one optional embodiment, anchoring the motion trajectory data to a node sequence of the scene semantic graph, and extracting the spatial propagation features of the node sequence as path guidance features using a graph convolutional neural network includes: Spatial coordinate analysis is performed on the trajectory points in the motion trajectory data. The spatial distance between each trajectory point and each node in the scene semantic graph is calculated. The node with the smallest distance is selected as the anchor node, and a mapping table from trajectory points to anchor nodes is established. Based on the mapping relationship, all anchor nodes passed by the target traffic participants are extracted, and the anchor nodes are arranged in time sequence according to the timestamp to generate an ordered node sequence. The spatial connectivity attribute of each node in the ordered node sequence and the topological edge weight between adjacent nodes are extracted from the scene semantic graph. A multi-layer graph convolutional neural network is constructed, which takes an ordered sequence of nodes and their topological edge weights as the graph structure input. In the graph convolutional layer, the spatial connectivity attributes of the current node and the spatial connectivity attributes of the neighboring nodes are aggregated by message passing. The aggregated information is weighted and fused using the topological edge weights to generate the node hidden layer features of the fused neighborhood structure. Through layer-by-layer propagation of multi-layer graph convolutional layers, each node in the ordered node sequence is sequentially fused with the topological structure information of its one-hop to multi-hop neighborhood, and the spatial propagation feature vector of each node is output. The spatial propagation feature vectors of each node in the ordered node sequence are concatenated according to the temporal order to form path-guided features.

[0029] In one specific implementation, after acquiring the motion trajectory data of the target traffic participants and constructing the scene semantic graph, it is necessary to spatially associate the trajectory data with the semantic graph so that discrete trajectory points can be represented on a structured graph topology. Spatial coordinate analysis is performed on each trajectory point in the motion trajectory data to extract its two-dimensional planar coordinates in a unified geographic coordinate system. Simultaneously, obtain the center coordinates of all nodes in the scene semantic graph. Calculate the trajectory point and the first point in the graph. Euclidean distance between nodes ,Right now ,in and The first The x and y coordinates of each node. Traverse all nodes in the graph and select... The node corresponding to the minimum value is used as the anchor node for that trajectory point, and a mapping table is established from the trajectory point index to the anchor node identifier. This mapping process discretizes the continuous physical trajectory into a sequence of node visits on the semantic graph, providing a foundation for subsequent graph structure feature extraction.

[0030] Based on the above mapping relationship, all trajectory points of the target traffic participant are traversed sequentially, and their corresponding anchor node sets are extracted. These sets are then sorted in ascending order according to the collection timestamp of each trajectory point to generate an ordered node sequence. ,in The length of the ordered node sequence. Indicates the first position in the sequence Each node in the ordered node sequence is deduplicated and compressed to ensure accurate reflection of the actual travel path of the participants on the road network topology. This is because, considering that target traffic participants may repeatedly anchor to the same node within a short period, the sequence is deduplicated and compressed. Specifically, when two adjacent anchored nodes are identical, the node that appears first is retained. Subsequently, the spatial connectivity attributes of each node in the ordered node sequence are extracted from the scene semantic graph. These attributes include node type labels (e.g., intersection, road segment midpoint, before the stop line), the number of passable directions, and the road level to which the current node belongs. Simultaneously, the topological edge weights between adjacent nodes in the ordered node sequence are extracted. This weight comprehensively encodes factors such as the segment length between adjacent nodes, road traffic priority, and connection direction confidence, providing a quantitative basis for the subsequent weighted aggregation of graph convolution.

[0031] A multi-layer graph convolutional neural network is constructed, using an ordered sequence of nodes and their topological edge weights as input to the graph structure. The core operation of the graph convolutional layer is a message-passing aggregation mechanism: for the current node in the sequence... First, its own spatial connectivity attribute vector The attribute vectors of all its directly adjacent nodes (i.e., one-hop neighbors) in the graph To aggregate, where superscript This represents the layer index of the current graph's convolutional layer. During aggregation, topological edge weights are used. The neighborhood aggregation vector is obtained by weighting the attribute vectors of neighboring nodes and then performing a normalized weighted summation. The calculation method is as follows ,in Represents a node The set of direct neighbor nodes in the scene semantic graph. The attribute vector of the current node itself. Aggregate vector with neighborhood After splicing, a learnable linear transformation matrix is ​​applied. After processing with a nonlinear activation function, the current node is obtained at the [node name missing]. Hidden layer features of the layer ,Right now ,in It is a non-linear activation function. This represents a vector concatenation operation. Through this design, each node can fuse the topological structure information of its immediate neighborhood after a single-layer graph convolution operation, generating hidden layer features of the node that fuse the neighborhood structure, which has a stronger structural awareness capability compared to the original node attributes.

[0032] Through the stacking and propagation of multiple graph convolutional layers, each node in an ordered node sequence can sequentially receive topological information from its one-hop, two-hop, or even multi-hop neighborhood. In the... Each node in the layer has integrated its... The receptive field of a node expands with increasing layers, capturing a wider range of road network topological dependencies. For example, when a target traffic participant is about to reach an intersection, the corresponding node, after passing through multiple graph convolutions, not only contains the connectivity information of the intersection itself but also incorporates the topological characteristics of downstream branch roads, thus providing richer spatial semantic support for path direction determination. In engineering implementation, the number of graph convolutional layers is generally set to 2 to 4. Too many layers can introduce oversmoothing problems, causing the features of different nodes to become consistent and lose their distinctiveness; setting the number of layers to 2 can effectively cover the neighborhood range at the intersection level, achieving a good balance between accuracy and computational cost. Finally, after propagation through all graph convolutional layers, the first node in the ordered node sequence... The spatial propagation feature vector output by each node is denoted as . Its dimension is determined by the output dimension of the last layer of linear transformation matrix.

[0033] ordered node sequence Spatial propagation feature vectors of each node Vectors are concatenated sequentially according to their temporal order to obtain the path guidance feature matrix. This feature matrix fully preserves the temporal travel information of target traffic participants on the road network topology and the spatial connectivity semantics at each location, effectively representing the participants' path selection preferences and spatial orientation characteristics. The path guidance features not only reflect the structured representation of the target participants' actual driving trajectories in the graph, but also encode the topological context of the road network along the route through multi-hop neighborhood aggregation, laying a structured feature foundation for subsequent fusion with interactive game features and generation of intent candidate sets. It is worth noting that when the ordered node sequence length of different target traffic participants... When there is inconsistency, the dimension alignment during batch processing can be ensured by zero-padding or truncating the path-guided feature matrix to a uniform length. The padding mask will be used to shield the influence of invalid padding positions in the subsequent attention calculation stage, thereby ensuring that the accuracy of feature extraction is not affected by the difference in sequence length.

[0034] like Figure 2 The diagram shown illustrates the logical flowchart for feature extraction in interactive games.

[0035] In one optional embodiment, the motion trajectory data and interaction state attributes are time-series segmented to calculate the change in interaction geometric relationship, determine the interaction game state, and extract interaction game features, including: The motion trajectory data is divided into time windows to generate multiple time segments. Interaction state attributes corresponding to each time segment are extracted from the scene semantic graph, and a time-series mapping table between time segments and interaction state attributes is established. For each time segment, extract the position coordinates and velocity vectors of the motion trajectory data, extract the position coordinates and velocity vectors of other traffic participants for the corresponding interaction state attributes, calculate the relative displacement vector, relative velocity vector and motion direction angle between the target traffic participant and each other traffic participant, and construct the interaction geometric relationship feature group; The difference of the interactive geometric relationship feature groups of adjacent time segments is calculated to obtain the sequence of changes in interactive geometric relationship. The change in relative displacement vector is decomposed in direction to extract the radial approach rate and tangential circumduction rate. Construct an interactive game state discriminator, calculate the ratio of radial approach rate to tangential detour rate, determine the game tendency coefficient, when the game tendency coefficient is greater than the preset tendency threshold, mark the competitive game state and extract the change in relative velocity vector as the competitive feature component, otherwise mark the avoidance game state and extract the change in the angle between the motion directions as the avoidance feature component. The competitive and avoidance feature components of all time segments are concatenated in time sequence to generate interactive game features.

[0036] In one specific implementation, after obtaining the motion trajectory data and the scene semantic graph, a joint temporal analysis of the two is required to capture the dynamic game relationship between the target traffic participant and other traffic participants over time. The motion trajectory data is divided into fixed time windows, segmenting the complete trajectory sequence into several continuous and non-overlapping time segments. Each time segment contains several consecutive frames of trajectory sampling points. The length of the time window is determined based on the interaction response cycle of traffic participants in the actual scene, typically selecting a time span that covers a complete interaction action. For each time segment, the interaction state attributes corresponding to the timestamp range of that segment are retrieved from the established scene semantic graph, including structured semantic information such as the occupancy status, priority relationships, and conflict zone activation status of surrounding traffic participants. Each time segment is bound to its corresponding interaction state attribute to form a temporal mapping table. This mapping table uses the time segment index as the key and the corresponding set of interaction state attributes as the value, ensuring precise alignment of the temporal and semantic dimensions in subsequent calculations.

[0037] For each time segment, the position coordinates and velocity vectors of the target traffic participant are extracted, along with the position coordinates and velocity vectors of each other traffic participant recorded in the scene semantic graph within that time segment. Let the position coordinates of the target traffic participant within a certain time segment be... The velocity vector is The coordinates of another traffic participant are The velocity vector is Then the relative displacement vector between the two Defined as Relative velocity vector Defined as Angle between the directions of motion velocity vector of the target traffic participant Velocity vectors of other traffic participants The angle between them is obtained by taking the inverse cosine of their inner product and the product of their magnitudes. For each frame within a time segment, the above three types of geometric quantities are calculated, and the frame-level mean is taken as the representative value for that time segment, thus constructing an interactive geometric relationship feature set containing the relative displacement vector, relative velocity vector, and angle between the motion directions. When multiple other traffic participants exist in the scene, an interactive geometric relationship feature set is independently constructed for each pair of "target participant - other participants," forming a parallel description of multiple pairs of interactive relationships.

[0038] The difference between the interaction geometric relationship feature groups of adjacent time segments is calculated to obtain information on the change of interaction relationships over time. Let the first... The relative displacement vectors of the time segments are , No. The relative displacement vectors of the time segments are The change in relative displacement is .right Perform directional decomposition: project it onto the current relative displacement vector. On the unit direction vector, the radial component is obtained, which reflects the rate at which the two participants approach or move away in the radial direction, denoted as the radial approach rate. ;Will Projected to perpendicular to On the unit tangential direction vector, the tangential component is obtained. This component reflects the relative circumduction rate of the two participants in the lateral direction, and is denoted as the tangential circumduction rate. Radial approach rate Negative values ​​indicate that the two are approaching each other, while positive values ​​indicate that they are moving away from each other; tangential circumduction rate. The larger the absolute value, the more obvious the lateral avoidance action between the two participants. Similarly, the change in relative velocity vector is obtained by subtracting the angle between the relative velocity vector and the direction of motion for adjacent time segments. Change in the angle between the direction of motion and the direction of motion Together, they constitute a sequence of changes in interactive geometric relationships.

[0039] An interactive game state discriminator is constructed, the core logic of which quantifies the dominant tendency of the game behavior between two traffic participants based on the ratio of radial approach rate to tangential detour rate. A game tendency coefficient is defined. for ,in To prevent extremely small positive numbers from being divided by zero. The physical meaning is: when When the size is large, radial approach actions dominate, indicating that the main interaction mode between the two parties is to compete for the same spatial resources, corresponding to a competitive game state; when When the value is relatively small, tangential circumvention is dominant, indicating that one or both parties choose to actively avoid the situation laterally, corresponding to an avoidance game state. Compared with the preset tendency threshold Comparison: When At that time, this time segment is marked as a competitive game state, and the change in the relative velocity vector is extracted. As a competitive feature component of this segment, it is used to describe the intensity of the game between the two sides at the speed level; when At that time, this time segment is marked as an avoidance game state, and the change in the angle of the movement direction is extracted. As the avoidance feature component of this segment, it is used to describe the negotiation and courtesy behavior of both parties at the level of direction adjustment. Preset tendency threshold. It can adaptively adjust according to different traffic scenario types (such as unsignalized intersections, merging zones, pedestrian crossings, etc.) to improve the discriminator's adaptability in different scenarios.

[0040] The competitive and avoidance feature components of all time segments are concatenated sequentially. For time segments marked as competitive game states, their competitive feature components are filled into the corresponding positions; for time segments marked as avoidance game states, their avoidance feature components are filled into the corresponding positions. Components corresponding to inactive states are filled with zeros, maintaining temporal alignment. This results in an interactive game feature vector sequence that fully preserves the dynamic evolution information of the game state within each time segment from the initial moment to the current moment. It includes both the speed game intensity of competitive behavior and the directional adjustment magnitude of avoidance behavior, providing a rich and structured description of interactive behavior for subsequent intention candidate set generation and Bayesian inference network, significantly improving the accuracy and robustness of traffic intention recognition.

[0041] like Figure 3 As shown, the interaction game characteristics between the target traffic participant and other vehicles are presented within eight time segments (t1~t8), including radial approach rate (orange), tangential detour rate (purple), and game tendency coefficient γ (blue broken line, with threshold lines marked). =1.2). Radial approach rate reflects the speed at which the two vehicles approach or move away along the relative displacement direction. A positive value indicates approach, and the peak value (3.1 m / s) is reached at time t4 in the figure, indicating intense forward competition between the two vehicles. Tangential avoidance rate reflects the intensity of lateral avoidance maneuvers, reaching its highest value (2.7 m / s) at time t6, corresponding to one party actively yielding laterally. Game tendency coefficient Integrating radial and tangential game tendencies, states with γ > 1.2 are labeled as competitive game states (e.g., t3, t4, t5), while those with γ < 1.2 are labeled as avoidance game states (e.g., t6, t7, t8). In competitive states, the change in relative velocity vector is extracted as a competitive feature component, while in avoidance states, the change in the angle between the movement directions is extracted as an avoidance feature component. These two types of components are then concatenated with path guidance features for use in a Bayesian inference network to generate a candidate intent set. The presence of threshold lines in the figure clearly distinguishes the switching moments of game modes, providing highly discriminative dynamic features for intent recognition and significantly improving the accuracy and robustness of traffic intent prediction in complex intersection scenarios.

[0042] In one optional embodiment, the path guidance features and interactive game features are used to generate an intent candidate set. A Bayesian inference network is then used to solve for the posterior probability distribution of the target traffic participants' intentions, resulting in the following traffic intention recognition results: By concatenating path-oriented features and interactive game features, spatial propagation feature vectors from path-oriented features and competition and avoidance feature components from interactive game features are extracted to construct a joint feature representation. Construct an intent candidate space, generate a candidate path set based on the topological connection relationship of each node in the spatial propagation feature vector, generate an interaction strategy set based on the numerical distribution of the competition feature component and the avoidance feature component, and combine the candidate path set and the interaction strategy set to form an intent candidate set. Construct a Bayesian inference network, and calculate the path conditional probability and policy conditional probability for each intent candidate in the intent candidate set based on the joint feature representation; A probability fusion layer is constructed, and the path conditional probability and the policy conditional probability are input into the probability fusion layer for joint probability inference. The joint conditional probability is generated through the coupling propagation of path space constraints and interactive game constraints. The joint conditional probability of each intention candidate in the intention candidate set is normalized to generate the posterior probability distribution of the intention of the target traffic participant. The candidate intent with the highest probability value is selected from the posterior probability distribution of intent as the result of the passage intent recognition.

[0043] In one specific implementation, after extracting the path guidance features and interactive game features, the two types of feature vectors are concatenated along the channel dimension to form a joint feature representation. Specifically, the spatial propagation feature vectors of each node are extracted from the path guidance feature matrix, and the competition feature component and the avoidance feature component are separated from the interactive game features. The competition feature component reflects the behavioral tendency of the target traffic participant to actively rush ahead during the interaction process, and is usually obtained by aggregating the feature values ​​corresponding to the time segments where the game tendency coefficient exceeds a preset tendency threshold; the avoidance feature component corresponds to the time segments where the game tendency coefficient is lower than the preset tendency threshold, reflecting the behavioral pattern of the target traffic participant actively slowing down or detouring. The spatial propagation feature vector, the competition feature component, and the avoidance feature component are concatenated along the dimension to form a unified high-dimensional joint feature vector. This vector simultaneously encodes path topology information and interactive dynamic information, providing complete feature input for the subsequent construction of the intention candidate set and Bayesian inference.

[0044] When constructing the intent candidate space, candidate options need to be generated from both the path and policy dimensions, and then combined. In the path dimension, based on the topological connections between nodes corresponding to the spatial propagation feature vectors, all reachable paths extending along directed edges from the target traffic participant's current anchor node are enumerated. Each reachable path is represented by a sequence of nodes. During the enumeration process, candidate paths are classified and labeled according to edge semantic attributes (such as straight, left turn, right turn, merging, splitting, etc.), forming a candidate path set. In the policy dimension, the numerical distribution of competition and avoidance feature components is statistically analyzed. Interaction behaviors are divided into several categories according to numerical ranges, such as strong competition, weak competition, neutral, weak avoidance, and strong avoidance. Each category corresponds to an interaction policy, thus forming an interaction policy set. The candidate path set and the interaction policy set are combined using a Cartesian product. Each combination element consists of a candidate path and an interaction policy, representing a complete traffic intention hypothesis of the target traffic participant in the current scenario. All combination elements together constitute the intent candidate set. The size of the candidate set is determined by the product of the number of candidate paths and the number of interaction strategy categories. In practical applications, the number of candidates can be pruned according to the complexity of the topological connection relationship and computing resources, eliminating combinations that contradict the semantic attributes of the current node to reduce invalid reasoning computation.

[0045] The Bayesian inference network is constructed with each intent candidate in the intent candidate set as the inference target, and calculates the conditional probabilities of two classes respectively. Let the i-th intent candidate be the one in the intent candidate set. The intention candidates are The joint features are represented as Path conditional probability The path conditional probability is calculated by inputting the joint feature representation into a dedicated neural network branch for path matching. This branch takes the spatially propagated feature vector as its primary input and utilizes the topological information of the node sequence of candidate paths as structural constraints. It outputs the matching score of the path corresponding to the intent candidate under the current feature conditions, which is then normalized using Softmax to obtain the path conditional probability. Policy conditional probability. The calculation is then performed through another neural network branch dedicated to policy classification. This branch takes competition and avoidance feature components as its main inputs and outputs a matching score between the interaction policy category of the intention candidate and the current interaction game features. After normalization, the policy conditional probability is obtained. The two branches share the input of the joint feature representation, but each performs parameter learning independently, thus ensuring the focus and complementarity of path reasoning and policy reasoning.

[0046] The probabilistic fusion layer is responsible for coupling and fusing path conditional probabilities and policy conditional probabilities to generate joint conditional probabilities. The fusion process is not a simple multiplication or weighted averaging, but rather a structured joint inference achieved through the coupled propagation of path space constraints and interactive game constraints. Path space constraints manifest as follows: when a candidate path passes through a node of a certain semantic type (such as a conflict zone node or a merging node), the path conditional probability exerts a suppressive or enhancing effect on the policy conditional probability; for example, if a candidate path must pass through a conflict zone, the fusion weight of the avoidance strategy increases accordingly, while the fusion weight of the strong competition strategy decreases accordingly. Interactive game constraints manifest as follows: when the temporal trend of the game propensity coefficient shows a continuously increasing trend, the joint conditional probability of competitive paths receives a positive correction. The path conditional probabilities... Policy conditional probability After fusion via the coupling propagation mechanism, the first... Joint conditional probability of each intention candidate Fusion computation can be achieved through a gated attention mechanism, where path feature vectors and policy feature vectors serve as query and key-value pairs, respectively. The fusion ratio of the two types of probabilities is dynamically adjusted through attention weights, so that the joint conditional probability is simultaneously constrained by both the path topology and the dynamic state of the interaction.

[0047] After obtaining the joint conditional probabilities of all intent candidates in the intent candidate set, they are normalized to ensure that the sum of the joint probabilities of all candidates is 1, thereby generating the posterior probability distribution of the target traffic participant's intent. Normalization uses the standard normalization formula: Let the intent candidate set contain a total of... The first candidate, then the second The posterior probability of the intention of each candidate is The posterior probability distribution of intent comprehensively describes the credibility of each traffic participant's intent hypothesis under the joint conditions of the current path topology and interaction state. The sharpness or flatness of the distribution reflects the certainty of intent recognition. When the distribution is relatively flat, it indicates that the current feature information is insufficient to distinguish multiple intent candidates, which may trigger further data collection or conservative control strategies. When the distribution is concentrated on a single candidate, it indicates high confidence in intent recognition and can be directly used for traffic control decisions.

[0048] The traffic intention candidate with the highest posterior probability value is selected from the posterior probability distribution of the intention and output as the final traffic intention recognition result. This result includes both a path component (the node sequence path that the target traffic participant is most likely to travel through) and a policy component (the interaction policy type that the target traffic participant is most likely to adopt), providing direct semantic input for subsequent conflict area identification and multi-agent game strategy construction. In actual deployment, the traffic intention recognition result can also be accompanied by confidence interval labeling, outputting candidates whose posterior probability exceeds a preset confidence threshold, for subsequent control modules to perform risk assessment and multi-scenario contingency plan formulation, thereby improving the robustness and safety of the overall control system in complex traffic scenarios.

[0049] In one optional embodiment, conflict areas are identified based on the passage intention recognition results, a multi-agent game state space is constructed and an action set is defined, a cumulative reward function is constructed and the optimal game strategy is solved using policy iteration, and a passage control strategy is generated, including: Based on the traffic intention recognition results, the expected travel path of the target traffic participant is extracted, and spatial overlap detection is performed with the expected travel paths of other traffic participants to identify conflict areas. A multi-agent game state space is constructed. The position, speed, and heading of the target traffic participant in the conflict area are used as the state of the first agent, and the position, speed, and heading of other traffic participants are used as the state of the second agent. The spatial boundary constraints of the conflict area are extracted to construct the game state space. Define sets of actions for acceleration, deceleration, and maintenance for the first agent state and the second agent state respectively, and establish a transition mapping relationship between state and action; A multi-agent reinforcement learning reward mechanism is constructed, which generates a safety reward component based on the relative position vector and relative speed vector of the target traffic participant and other traffic participants, an efficiency reward component based on the travel time, and a collision penalty component based on the predicted trajectory intersection, and constructs a cumulative reward function. The optimal game strategy is solved by iteratively performing action selection and state transition in the game state space, updating the decision strategy based on the cumulative reward function, and using the strategy iteration to perform action selection and state transition in the game state space. The action sequence is extracted based on the optimal game strategy, converted into passage time adjustment instructions and spatial trajectory adjustment instructions, and a passage control strategy is generated.

[0050] In one specific implementation, based on the expected travel path in the traffic intention recognition result, overlap detection is performed on the spatial trajectories of the target traffic participant and other traffic participants. Specifically, each expected travel path is discretized into a series of path segments. For any two path segments from different participants, their spatial intersection relationship is calculated to determine whether there are areas of geometric intersection or spatial distance less than a safety threshold. If spatial overlap is detected between path segments, the spatial range of the overlapping area is marked as a conflict area, and the geometric center coordinates, spatial coverage, and expected arrival time window of the conflict area are recorded. When multiple paths form multiple independent overlapping areas, they are marked as independent conflict areas and sorted according to time priority to ensure the correct priority processing logic for conflict areas in subsequent game modeling.

[0051] After identifying the conflict zone, a multi-agent game state space is constructed. The target traffic participant is considered the first agent, and its state vector is composed of its position coordinates, velocity magnitude, and direction (heading angle) near the conflict zone. Other traffic participants are considered as second intelligent agents, and their states are constituted by the position coordinates, velocity magnitude, and direction of the corresponding participants, forming the state vector of the second intelligent agent. In addition to kinematic states, it is also necessary to extract spatial boundary constraints of the conflict zone, including the entrance boundary, exit boundary, and passable width limit of the conflict zone. These spatial constraint parameters are then integrated into an environmental constraint vector. ,and and Together they constitute a complete game state space. The game state space uses a discretized grid method to quantize the position dimension, while the velocity and heading are quantized using segmented intervals. This maps the continuous state space into a finite set of discrete states, ensuring the computability of subsequent strategy iterations.

[0052] Define action sets for the first agent state and the second agent state respectively. ,in Indicates acceleration, Indicates deceleration action, This indicates an action that maintains the current speed. Each action corresponds to a specific speed change parameter: acceleration increases speed by a preset acceleration increment, deceleration decreases speed by a preset deceleration increment, and maintaining the current speed is the final action. When establishing the state-action transition mapping, the kinematic equations are used to calculate the speed in the current state. Next action pair The next state ,in For the actions of the first intelligent agent, This is the action of the second agent. If the state transition results in any participant exceeding the spatial boundary constraints of the conflict zone or exceeding a reasonable speed range, the transition is marked as an illegal transition, and the corresponding action is given a very high penalty weight in the policy iteration to exclude it from selection.

[0053] When constructing a multi-agent reinforcement learning reward mechanism, the safety reward component Based on the relative position vector between the target traffic participant and other traffic participants and relative velocity vector Calculated. Defined as the displacement vector obtained by subtracting the position coordinates of the first agent from the position coordinates of the second agent. Defined as the difference between the velocity vectors of the two. When Exceeding the safe distance threshold hour, Take positive values ​​to encourage safe passage; when Below and When the direction points to both directions being close to each other, Negative values ​​are used to penalize dangerous approach behavior. Efficiency reward component. Based on the calculation of travel time, travel time is defined as the number of time steps taken by a target traffic participant from entering the boundary of the conflict zone to completely leaving the boundary of the conflict zone. , and The reward is inversely proportional to the travel time; that is, the shorter the travel time, the higher the reward value, thus incentivizing decision-making strategies to complete the journey as quickly as possible while ensuring safety. Collision penalty component. Based on the intersection point generation of predicted trajectories: In the current state, extend the predicted trajectories of the two agents forward by several time steps. If the predicted trajectories intersect and the time of the intersection point corresponds to the time difference between the two, then... If the value is below the collision warning threshold, a collision penalty will be triggered. Take the larger negative value; if If the warning threshold is exceeded, no penalty is triggered. Cumulative reward function. It is composed of the weighted sum of the above three components, that is ,in , , These are the weighting coefficients for safety rewards, efficiency rewards, and collision penalties, all of which are positive real numbers, and are adjusted according to the actual scenario to determine the emphasis on safety and traffic efficiency.

[0054] The strategy iteration process occurs within the constructed game state space. Execute above. Initialize the policy of the first agent. Second agent strategy The policy is uniformly random, meaning that each action in the action set is assigned an equal probability of selection in any state. The subsequent policy evaluation phase involves fixing the current policy. The value function for each state is calculated by iterating the Bellman equation. The convergence condition is that the maximum difference in the value function between adjacent iterations is lower than the accuracy threshold. Policy improvement phase: For each state, calculate the action value for each possible action of the first agent based on the current value function, and select the action that maximizes the action value to update the strategy. Similarly, update The strategy evaluation and strategy improvement are performed alternately until the strategy no longer changes within a certain number of consecutive rounds, at which point it is considered convergent, and the optimal game strategy is obtained. .

[0055] According to the optimal game strategy Starting from the current game state, actions are selected step by step according to the state transition mapping to extract the action sequence of the first agent. ,in This represents the total number of steps in the action sequence. The action sequence is parsed: if the action is acceleration or deceleration, it is converted into the corresponding speed adjustment amount, and a traffic time adjustment instruction is generated, specifying the target speed value for the traffic participant at each time point; if the action causes path deviation (when multiple paths are available), a spatial trajectory adjustment instruction is generated, specifying the adjusted travel path segment. The traffic time adjustment instruction and the spatial trajectory adjustment instruction together constitute a complete traffic control strategy, including a speed control sequence in the time dimension and a path planning sequence in the spatial dimension.

[0056] Furthermore, based on the generated traffic control strategy, when implementing traffic control for target traffic participants, the traffic time adjustment command is sent to the speed control execution unit, and the spatial trajectory adjustment command is sent to the path tracking control unit. During execution, the state observations of the target traffic participant and other traffic participants are updated in real time and aligned with the state definitions in the game's state space. If the deviation between the observed actual state and the predicted state exceeds a tolerance threshold... If this happens, the strategy re-solution process is triggered, and the strategy iteration is re-executed with the current real-time state as the initial state. The action sequence is updated and new control commands are issued to ensure the robustness and adaptability of the traffic control strategy in dynamic traffic scenarios.

[0057] In one optional embodiment, a multi-agent reinforcement learning reward mechanism is constructed, generating a safety reward component based on the relative position vector and relative velocity vector of the target traffic participant and other traffic participants, an efficiency reward component based on the travel time, and a collision penalty component based on the predicted trajectory intersection. The cumulative reward function includes: Extract the relative position vector and relative velocity vector between the target traffic participant and other traffic participants. Calculate the braking distance based on the relative velocity vector. If the magnitude of the relative position vector is greater than the braking distance, it is determined that the safe distance condition is met, and a safety reward component is generated. Record the time when the target traffic participant enters the conflict area and the time when the target traffic participant leaves the conflict area, calculate the time difference as the passage time, and generate an efficiency reward component when the passage time is less than a preset benchmark time. Calculate the predicted trajectories of the target traffic participant and other traffic participants within a future time window, determine whether the predicted trajectories intersect, and if so, generate a collision penalty component based on the time interval at the intersection. The safety reward component, the efficiency reward component, and the collision penalty component are weighted and summed to construct a cumulative reward function.

[0058] In one specific implementation, when constructing a multi-agent reinforcement learning reward mechanism, it is necessary to design independent reward components for the three dimensions of safety, traffic efficiency, and collision risk in traffic conflict scenarios, and then combine them into a complete cumulative reward function through a weighted summation method. This hierarchical design approach can ensure that the reward signals of each dimension are independent and have clear physical meanings, avoiding signal confusion between different objectives that could lead to policy degradation.

[0059] The safety reward component is calculated using the relative position vector between the target traffic participant and other traffic participants. and relative velocity vector Based on the input: relative position vector. The modulus is obtained by subtracting the position coordinates of the two participants in the same global coordinate system. Reflects the current actual spatial distance between the two. This reflects the tendency for the two participants to approach or move away, and its radial component determines the rate at which the distance between them decreases. When calculating the braking distance based on the relative velocity vector, take... exist The projected component in the direction is used as the effective approach velocity. ,Right now ,when This indicates that the two are approaching each other. The maximum deceleration of the target traffic participant is known. Under these conditions, braking distance Calculated according to the uniformly decelerated motion model, i.e. .when When the current state is determined to meet the safe distance condition, a positive safety reward component is generated. Conversely, when If the target traffic participant cannot stop before a collision even if they immediately apply maximum braking, the safety reward will be zero or a penalty will be imposed. The specific values ​​can be further determined based on and The difference is assigned a continuous value, and the larger the interval, the higher the reward, thereby guiding the strategy to maintain a sufficient safety margin rather than just satisfying the critical condition.

[0060] The generation of the efficiency reward component relies on the accurate recording of the travel time of the target traffic participants within the conflict zone. The positional status of the target traffic participants is continuously monitored at each simulation time step, and the entry time is recorded when their position coordinates first enter the pre-defined conflict zone spatial range. When its location coordinates leave the conflict zone, the departure time is recorded. The difference between the two is the travel time. ,Right now Preset base duration Determined by the length of the conflict zone and the reference speed, it represents the shortest reasonable time required to pass through the conflict zone at normal speed under undisturbed conditions. This indicates that the target traffic participant completed the passage with high efficiency, generating a positive efficiency reward component. ;when In this case, the efficiency reward component is set to zero, with no additional penalty, thus preventing the strategy from taking aggressive actions in pursuit of efficiency rewards. The magnitude of the efficiency reward component should be lower than that of the safety reward component to ensure that the strategy prioritizes safety when there is a conflict between safety and efficiency.

[0061] The generation of the collision penalty component needs to be done within a future time window. The system predicts the trajectories of the target traffic participant and other traffic participants separately. The predicted trajectories can be generated using uniform linear extrapolation or polynomial fitting based on historical motion states, with the prediction step size consistent with the simulation time step. For the two predicted trajectories, the system determines whether a spatial intersection exists by gradually increasing the predicted trajectories under time alignment conditions. The criterion is whether the Euclidean distance between the predicted positions of the two participants at the same time is less than the collision detection radius. If an intersection of predicted trajectories is detected within the time window, the predicted arrival time of the target traffic participant at the intersection is extracted. Predicted arrival times with other traffic participants Calculate the time difference of arrival for both. . The smaller the value, the more likely the two participants arrive at the intersection point simultaneously, indicating a higher risk of collision and requiring a stronger penalty signal. Collision penalty component. and There is a negative correlation when The penalty is maximized when the value approaches zero. Exceeding the safety time margin threshold The time penalty component is reduced to zero, meaning no collision penalty is triggered. If the predicted trajectories do not intersect within the time window, the collision penalty component... Take zero. Note that... The value should appear as a negative or zero value in the cumulative reward function, and its absolute value should be significantly greater than the magnitude of the safety reward and efficiency reward, so as to form a strong penalty signal to force the strategy to avoid high collision risk behaviors.

[0062] After completing the independent calculations of the above three components, the security reward component will be... Efficiency Rewards and collision penalty components Perform weighted summation to construct a cumulative reward function. As shown below: ; in , , These are the safety reward weighting coefficient, efficiency reward weighting coefficient, and collision penalty weighting coefficient, respectively, all of which are positive real numbers. The value should be set significantly greater than and To ensure that collision penalties play a dominant role in policy optimization, the weighting coefficients can be determined through cross-validation or by searching multiple sets of parameters in a simulation environment, ensuring that the policy converges to safe and efficient traffic behavior in scenarios with different traffic densities and conflict intensities.

[0063] The cumulative reward function is calculated at each time step and passed to the reinforcement learning agent as a supervisory signal for policy gradient updates or value function iterations. During policy iteration, the agent drives the policy to evolve in a direction that satisfies safety constraints and efficiency objectives by maximizing the cumulative discount reward. At the same time, the penalty term continuously suppresses the policy from making high-collision-probability behavioral decisions, ultimately enabling the two players to form a stable cooperative policy near the Nash equilibrium point.

[0064] like Figure 4The diagram illustrates the composition and dynamic changes of the cumulative reward function in multi-agent game control over eight consecutive time steps (s1~s8). The safety reward component (green) calculates the braking distance based on the relative position and velocity vectors between the target vehicle and other vehicles. A positive reward is given when the distance is greater than the braking distance. The highest safety reward (0.9) is observed at time s4, corresponding to a large safety margin. The efficiency reward component (blue) is generated based on the difference between the travel time and the baseline travel time. The efficiency rewards are relatively high at times s1 and s6 (0.25), indicating smoother travel. The collision penalty component (red) is triggered by the time interval between predicted trajectory intersections. If the time difference between the expected arrival times of the two vehicles at the conflict point is less than a safety threshold, a negative penalty is applied. The heaviest penalties (-1.2 and -1.0) are observed at times s1 and s6, suggesting a high risk of collision and forcing the strategy to prioritize deceleration or avoidance.

[0065] Looking at the cumulative total (values ​​at the top of the bars), the total rewards at times s2, s4, and s7 are relatively high, corresponding to a strategy that tends towards a "safety + efficiency" equilibrium state; while the total rewards at s1 and s6 are relatively low, mainly due to the drag from strong penalties, indicating that in these states, iterative adjustments to speed or trajectory are necessary to avoid conflict. This figure intuitively verifies the guiding role of reward function design on game strategy—the weighting of safe rewards and penalties (…). =0.6, =1.8) makes collision penalties dominant, ensuring that the strategy optimizes efficiency under safety constraints.

[0066] A second aspect of the present invention provides a traffic intention recognition and control system based on traffic scene understanding, comprising: The data acquisition unit is used to acquire the movement trajectory data and traffic scene data of the target traffic participants; The scene semantic unit is used to perform semantic segmentation on the traffic scene data to extract topological graph nodes, label the semantic attributes of the nodes, establish directed connection edges and label the semantic attributes of the edges, and generate a scene semantic graph through a graph neural network. The trajectory feature unit is used to anchor the motion trajectory data to the node sequence of the scene semantic graph, extract the spatial propagation features of the node sequence as path guidance features using a graph convolutional neural network, perform temporal segmentation of the motion trajectory data and interaction state attributes to calculate the change in interaction geometric relationship, determine the interaction game state, and extract interaction game features. The intent recognition unit is used to generate an intent candidate set from path guidance features and interactive game features, and solve the posterior probability distribution of the intent of the target traffic participants through a Bayesian inference network to obtain the traffic intent recognition result. The game strategy unit is used to identify conflict areas based on the passage intention recognition result, construct a multi-agent game state space and define an action set, construct a cumulative reward function and use strategy iteration to solve the optimal game strategy, and generate a passage control strategy. A traffic control unit is used to perform traffic control on the target traffic participants based on the traffic control strategy.

[0067] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0068] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0069] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A traffic intent recognition and control method based on traffic scenario understanding, characterized in that, include: Acquire movement trajectory data and traffic scene data of the target traffic participants; The traffic scene data is semantically segmented to extract topological graph nodes, and the semantic attributes of the nodes are labeled. Directed connecting edges are established and the semantic attributes of the edges are labeled. A scene semantic graph is generated through a graph neural network. The motion trajectory data is anchored to the node sequence of the scene semantic graph. The spatial propagation features of the node sequence are extracted as path guidance features using a graph convolutional neural network. The motion trajectory data and interaction state attributes are temporally segmented to calculate the change in interaction geometric relationship, determine the interaction game state, and extract interaction game features. The path guidance features and interactive game features are used to generate an intent candidate set. The posterior probability distribution of the intent of the target traffic participants is solved by a Bayesian inference network to obtain the traffic intent recognition result. Based on the passage intention recognition result, identify the conflict area, construct a multi-agent game state space and define the action set, construct a cumulative reward function and use strategy iteration to solve the optimal game strategy, and generate a passage control strategy. Traffic control is implemented on the target traffic participants based on the traffic control strategy.

2. The method according to claim 1, characterized in that, The process involves semantic segmentation of the traffic scene data to extract topological graph nodes, labeling node semantic attributes, establishing directed connections and labeling edge semantic attributes, and generating a scene semantic graph using a graph neural network, including: Semantic segmentation is performed on traffic scene data to identify road areas, lane areas, and intersection areas. The boundary contour lines of road areas and the center lines of lane areas are extracted. The intersection of the boundary contour lines and the center lines is used as topology graph nodes. Conflict area identifiers are marked according to whether the topology graph node is located in an intersection area. The number of associated lanes, turning type attributes, and conflict area identifiers are marked for each topology graph node to generate node semantic attributes. Detect the current position of other traffic participants in the dynamic layer, determine the topology graph node occupied by each other traffic participant, establish directed connection edges between the topology graph node occupied by other traffic participants and the adjacent topology graph node in the direction of movement, label the directed connection edges with the instantaneous speed and direction of movement angle of other traffic participants, and generate edge semantic attributes. Using node semantic attributes as the initial node feature matrix and edge semantic attributes as the adjacency relation matrix, node features are propagated through a graph neural network, enabling each topological graph node to aggregate the feature information of its neighboring nodes and generate a scene semantic graph.

3. The method according to claim 1, characterized in that, Anchoring the motion trajectory data to the node sequence of the scene semantic graph, and extracting the spatial propagation features of the node sequence as path guidance features using a graph convolutional neural network, includes: Spatial coordinate analysis is performed on the trajectory points in the motion trajectory data. The spatial distance between each trajectory point and each node in the scene semantic graph is calculated. The node with the smallest distance is selected as the anchor node, and a mapping table from trajectory points to anchor nodes is established. Based on the mapping relationship, all anchor nodes passed by the target traffic participants are extracted, and the anchor nodes are arranged in time sequence according to the timestamp to generate an ordered node sequence. The spatial connectivity attribute of each node in the ordered node sequence and the topological edge weight between adjacent nodes are extracted from the scene semantic graph. A multi-layer graph convolutional neural network is constructed, which takes an ordered sequence of nodes and their topological edge weights as the graph structure input. In the graph convolutional layer, the spatial connectivity attributes of the current node and the spatial connectivity attributes of the neighboring nodes are aggregated by message passing. The aggregated information is weighted and fused using the topological edge weights to generate the node hidden layer features of the fused neighborhood structure. Through layer-by-layer propagation of multi-layer graph convolutional layers, each node in the ordered node sequence is sequentially fused with the topological structure information of its one-hop to multi-hop neighborhood, and the spatial propagation feature vector of each node is output. The spatial propagation feature vectors of each node in the ordered node sequence are concatenated according to the temporal order to form path-guided features.

4. The method according to claim 1, characterized in that, The motion trajectory data and interaction state attributes are time-series segmented to calculate the changes in the geometric relationship of the interaction, determine the interaction game state, and extract interaction game features, including: The motion trajectory data is divided into time windows to generate multiple time segments. Interaction state attributes corresponding to each time segment are extracted from the scene semantic graph, and a time-series mapping table between time segments and interaction state attributes is established. For each time segment, extract the position coordinates and velocity vectors of the motion trajectory data, extract the position coordinates and velocity vectors of other traffic participants for the corresponding interaction state attributes, calculate the relative displacement vector, relative velocity vector and motion direction angle between the target traffic participant and each other traffic participant, and construct the interaction geometric relationship feature group; The difference of the interactive geometric relationship feature groups of adjacent time segments is calculated to obtain the sequence of changes in interactive geometric relationship. The change in relative displacement vector is decomposed in direction to extract the radial approach rate and tangential circumduction rate. Construct an interactive game state discriminator, calculate the ratio of radial approach rate to tangential detour rate, determine the game tendency coefficient, when the game tendency coefficient is greater than the preset tendency threshold, mark the competitive game state and extract the change in relative velocity vector as the competitive feature component, otherwise mark the avoidance game state and extract the change in the angle between the motion directions as the avoidance feature component. The competitive and avoidance feature components of all time segments are concatenated in time sequence to generate interactive game features.

5. The method according to claim 1, characterized in that, By generating a candidate intent set from path-oriented features and interactive game features, and solving the posterior probability distribution of the intent of the target traffic participants through a Bayesian inference network, the traffic intent recognition results are obtained, including: By concatenating path-oriented features and interactive game features, spatial propagation feature vectors from path-oriented features and competition and avoidance feature components from interactive game features are extracted to construct a joint feature representation. Construct an intent candidate space, generate a candidate path set based on the topological connection relationship of each node in the spatial propagation feature vector, generate an interaction strategy set based on the numerical distribution of the competition feature component and the avoidance feature component, and combine the candidate path set and the interaction strategy set to form an intent candidate set. Construct a Bayesian inference network, and calculate the path conditional probability and policy conditional probability for each intent candidate in the intent candidate set based on the joint feature representation; A probability fusion layer is constructed, and the path conditional probability and the policy conditional probability are input into the probability fusion layer for joint probability inference. The joint conditional probability is generated through the coupling propagation of path space constraints and interactive game constraints. The joint conditional probability of each intention candidate in the intention candidate set is normalized to generate the posterior probability distribution of the intention of the target traffic participant. The candidate intent with the highest probability value is selected from the posterior probability distribution of intent as the result of the passage intent recognition.

6. The method according to claim 1, characterized in that, Based on the traffic intention recognition results, conflict areas are identified, a multi-agent game state space is constructed, and an action set is defined. A cumulative reward function is constructed, and the optimal game strategy is solved using policy iteration. The resulting traffic control strategy includes: Based on the traffic intention recognition results, the expected travel path of the target traffic participant is extracted, and spatial overlap detection is performed with the expected travel paths of other traffic participants to identify conflict areas. A multi-agent game state space is constructed. The position, speed, and heading of the target traffic participant in the conflict area are used as the state of the first agent, and the position, speed, and heading of other traffic participants are used as the state of the second agent. The spatial boundary constraints of the conflict area are extracted to construct the game state space. Define sets of actions for acceleration, deceleration, and maintenance for the first agent state and the second agent state respectively, and establish a transition mapping relationship between state and action; A multi-agent reinforcement learning reward mechanism is constructed, which generates a safety reward component based on the relative position vector and relative speed vector of the target traffic participant and other traffic participants, an efficiency reward component based on the travel time, and a collision penalty component based on the predicted trajectory intersection, and constructs a cumulative reward function. The optimal game strategy is solved by iteratively performing action selection and state transition in the game state space, updating the decision strategy based on the cumulative reward function, and using the strategy iteration to perform action selection and state transition in the game state space. The action sequence is extracted based on the optimal game strategy, converted into passage time adjustment instructions and spatial trajectory adjustment instructions, and a passage control strategy is generated.

7. The method according to claim 6, characterized in that, A multi-agent reinforcement learning reward mechanism is constructed, generating a safety reward component based on the relative position and relative velocity vectors of the target traffic participant and other traffic participants, an efficiency reward component based on travel time, and a collision penalty component based on predicted trajectory intersections. The cumulative reward function includes: Extract the relative position vector and relative velocity vector between the target traffic participant and other traffic participants. Calculate the braking distance based on the relative velocity vector. If the magnitude of the relative position vector is greater than the braking distance, it is determined that the safe distance condition is met, and a safety reward component is generated. Record the time when the target traffic participant enters the conflict area and the time when the target traffic participant leaves the conflict area, calculate the time difference as the passage time, and generate an efficiency reward component when the passage time is less than a preset benchmark time. Calculate the predicted trajectories of the target traffic participant and other traffic participants within a future time window, determine whether the predicted trajectories intersect, and if so, generate a collision penalty component based on the time interval at the intersection. The safety reward component, the efficiency reward component, and the collision penalty component are weighted and summed to construct a cumulative reward function.

8. A traffic intent recognition control system based on traffic scene understanding, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire the movement trajectory data and traffic scene data of the target traffic participants; The scene semantic unit is used to perform semantic segmentation on the traffic scene data to extract topological graph nodes, label the semantic attributes of the nodes, establish directed connection edges and label the semantic attributes of the edges, and generate a scene semantic graph through a graph neural network. The trajectory feature unit is used to anchor the motion trajectory data to the node sequence of the scene semantic graph, extract the spatial propagation features of the node sequence as path guidance features using a graph convolutional neural network, perform temporal segmentation of the motion trajectory data and interaction state attributes to calculate the change in interaction geometric relationship, determine the interaction game state, and extract interaction game features. The intent recognition unit is used to generate an intent candidate set from path guidance features and interactive game features, and solve the posterior probability distribution of the intent of the target traffic participants through a Bayesian inference network to obtain the traffic intent recognition result. The game strategy unit is used to identify conflict areas based on the passage intention recognition result, construct a multi-agent game state space and define an action set, construct a cumulative reward function and use strategy iteration to solve the optimal game strategy, and generate a passage control strategy. A traffic control unit is used to perform traffic control on the target traffic participants based on the traffic control strategy.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images