Vehicle trajectory prediction method, device and system for right-of-way conflict scenarios

Abstract

The application relates to the technical field of automatic driving, and particularly discloses a vehicle trajectory prediction method, device and system for a right-of-way conflict scene, which comprises the following steps: acquiring map data with semantic information and historical trajectory data of a target vehicle respectively; performing static scene understanding according to the map data with semantic information; performing dynamic element feature extraction according to the historical trajectory data of the target vehicle; performing interactive intention recognition according to road traffic rule information and target vehicle motion characteristics, so as to obtain an interactive intention feature vector of the target vehicle and a lane; and performing hierarchical trajectory prediction according to the interactive intention feature vector of the target vehicle and the lane and the target vehicle motion characteristics, so as to obtain a multi-modal driving intention of the target vehicle and a probability trajectory of each modal driving intention. The vehicle trajectory prediction method provided by the application can realize accurate and robust trajectory prediction for a complex traffic scene without sacrificing fidelity.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a vehicle trajectory prediction method, a vehicle trajectory prediction device, and a vehicle trajectory prediction system for right-of-way conflict scenarios. Background Technology

[0002] To achieve safe autonomous driving, vehicles must be able to accurately predict the future trajectories of other road users in the vicinity. This predictive capability is the foundation for downstream planning and control modules to make safe and efficient decisions. With the development of deep learning technology, the field of trajectory prediction has evolved from simple physical models to complex interactive perception models. However, existing technologies still have significant limitations in representing environmental information, capturing interaction relationships, and achieving long-term predictions.

[0003] In the early stages of trajectory prediction, researchers primarily employed physics-based models, such as uniform velocity models, uniform acceleration models, or more complex vehicle kinematics / dynamics models. These methods utilize physical laws to infer future motion states. Meanwhile, classic machine learning methods, such as the Kalman filter and its variants, were also widely used for state estimation and prediction based on observational data. The fundamental limitation of these methods lies in their simplistic assumptions: that the movements of traffic participants follow pre-defined, relatively simple rules. Therefore, they perform adequately only in short-term predictions (e.g., within 1-2 seconds) and cannot effectively model complex, non-linear driving behaviors, such as sudden lane changes, emergency braking, or interactive game-like situations in dense traffic. In complex scenarios requiring an understanding of driver intent and multi-vehicle interactions, the predictive accuracy of these models drops sharply, making them unsuitable for the safety requirements of high-level autonomous driving systems.

[0004] To overcome the limitations of physical models, deep learning methods have been introduced into the field of trajectory prediction. Among them, Recurrent Neural Networks (RNNs) and their variants, especially Long Short-Term Memory (LSTM) networks, are widely used due to their inherent advantages in processing time-series data. These models treat the historical trajectory of a vehicle as a time series, capturing the long-term temporal dependencies of motion through their internal memory units.

[0005] While LSTMs excel at capturing the temporal dynamics of individual agents, they inherently struggle with explicitly modeling complex spatial interactions. To address this, researchers have proposed mechanisms such as "social pooling," which aggregates spatial information by aggregating the hidden states of neighboring agents. However, this pooling operation typically aggregates information in a generic, unstructured manner, making it difficult to distinguish between different types of interactions and failing to fully utilize the rich structured information provided by high-resolution maps. Furthermore, the sequential computational nature of RNN / LSTM models makes parallelization during training and inference difficult, resulting in low computational efficiency, especially in scenarios involving a large number of traffic participants.

[0006] Another mainstream approach represents the driving scene as a rasterized image from a bird's-eye view (BEV). In this paradigm, scene elements such as road layout and vehicle position are "drawn" onto a two-dimensional pixel grid, and then the powerful image feature extraction capabilities of convolutional neural networks (CNNs) are used to learn the scene representation. This method intuitively transforms the trajectory prediction problem into a computer vision task. However, this method has a fundamental drawback: the rasterization process itself. It converts structured, vector map information (such as precise lane centerlines and topological connections between lanes) into a discrete pixel grid, inevitably leading to information loss and quantization errors. Precise geometric information (such as lane curvature) and topological information (such as whether a lane is preceding, succeeding, or adjacent) become blurred after rasterization. The model needs to indirectly learn this information, which should be explicitly given, from the pixel patterns, increasing the learning difficulty. Furthermore, to achieve sufficient accuracy, high-resolution raster images are required, which incurs huge computational and memory overhead, limiting its application in resource-constrained in-vehicle systems.

[0007] To more naturally model the interactions between entities in traffic scenarios, Graph Neural Networks (GNNs) have been introduced into the field. In GNNs, traffic participants can be represented as nodes in a graph, and their interactions are represented as edges. This method can directly model interactions in non-Euclidean space, avoiding the information loss caused by rasterization.

[0008] Although GNN represents the right direction, early GNN applications still have the following significant shortcomings: (1) Coarse-grained graph representation: Some methods treat the entire lane centerline (polyline) as a single graph node. This representation is too coarse, ignoring the geometric details inside the lane and failing to perform fine-grained interaction modeling. For example, the interaction between a vehicle and the lane should be different at different positions on the lane segment, and treating the entire lane as a single node fails to capture this difference. (2) General graph convolution operations: These methods encode road networks using simple, undirected, or fully connected graph structures. This general design cannot encode the rich topological semantics specific to road networks. For example, it cannot distinguish whether a lane is the successor, predecessor, or left-neighbor lane of another lane, and these different topological relationships have drastically different effects on driving behavior. (3) Inefficient long-range dependency modeling: Standard GCN propagates information by stacking layers. To capture long-range dependencies (e.g., predicting turning behavior at an intersection tens of meters away), information needs to be propagated multiple steps across the graph, which means a very deep network is required.

[0009] In summary, existing technologies in trajectory prediction face a core trade-off: the contradiction between representational fidelity and computational feasibility. CNN rasterization methods sacrifice map information fidelity in order to utilize mature image processing architectures. While early GNN methods improved interactive modeling capabilities, their graph structures and operator designs were too general, failing to fully capture the unique, high-fidelity topology of road networks. Therefore, how to achieve accurate and robust trajectory prediction for complex traffic scenarios without sacrificing fidelity has become a pressing technical problem for those skilled in the art. Summary of the Invention

[0010] This invention provides a vehicle trajectory prediction method, a vehicle trajectory prediction device, and a vehicle trajectory prediction system for right-of-way conflict scenarios, solving the problem in related technologies that trajectory prediction cannot be performed accurately and robustly without sacrificing fidelity in complex traffic scenarios.

[0011] As a first aspect of the present invention, a method for predicting vehicle trajectories in right-of-way conflict scenarios is provided, comprising:

[0012] Acquire map data with semantic information and historical trajectory data of the target vehicle, respectively;

[0013] Static scene understanding is performed based on the map data containing semantic information to obtain road traffic rule information;

[0014] Dynamic feature extraction is performed on the historical trajectory data of the target vehicle to obtain the motion characteristics of the target vehicle.

[0015] Based on the road traffic rules information and the target vehicle's motion characteristics, interactive intent recognition is performed to obtain the interactive intent feature vector between the target vehicle and the lane.

[0016] Based on the interactive intent feature vector between the target vehicle and the lane and the motion features of the target vehicle, hierarchical trajectory prediction is performed to obtain the multimodal driving intent of the target vehicle and the probability trajectory of each modal driving intent.

[0017] Furthermore, static scene understanding is performed based on the map data containing semantic information to obtain road traffic rule information, including:

[0018] Based on the map data with semantic information, graph nodes are defined, node features are parameterized, and graph edges are defined and semantically labeled to obtain a traffic knowledge graph containing semantic information.

[0019] A semantically gated graph convolution operator is constructed and applied to the traffic knowledge graph containing semantic information to assist the traffic knowledge graph containing semantic information in information dissemination with logical reasoning ability.

[0020] Furthermore, a semantically gated graph convolution operator is constructed, including:

[0021] Determine the semantic features of the connection edges formed when information is transmitted between neighboring nodes and the current node;

[0022] A semantic gate value is determined based on the semantic features of the connected edges, wherein the expression for the semantic gate value is:

[0023]

[0024] Among them, g ji,k The semantic gate value is represented by σ, which represents the Sigmoid function. Indicate the semantic features of the connecting edges;

[0025] Modulated aggregated information is obtained by modulating the information stream according to semantic gating values, wherein the expression of the modulated aggregated information is:

[0026]

[0027] Where, m ji,k This represents the modulated aggregated information. This represents the characteristics of neighboring nodes at layer l. The symbol represents the weight matrix, and ⊙ represents the Hadamard product;

[0028] Based on the modulated aggregation information, a semantically gated graph convolution operator is determined, and the expression of the semantically gated graph convolution operator is as follows:

[0029]

[0030] in, This represents the complete update rule of the current node at level l+1. This represents the feature of the current node at layer l. This represents the weight matrix.

[0031] Furthermore, dynamic feature extraction is performed based on the historical trajectory data of the target vehicle, including:

[0032] The historical trajectory data of the target vehicle is represented as a historical displacement vector;

[0033] The historical displacement vector is mapped onto a temporal convolutional network of a preset dimension to obtain the dynamic feature vector of the target vehicle.

[0034] Furthermore, interactive intent recognition is performed based on the road traffic rule information and the target vehicle's motion characteristics, including:

[0035] Based on the road traffic rules information and the target vehicle motion characteristics, an iterative interactive fusion is performed to obtain the interactive intent feature vector of the target vehicle and the lane. In each iteration, an interactive process of information injection, information propagation, information feedback and interactive modeling is performed.

[0036] Furthermore, each iteration involves an interactive process of information injection, information propagation, information feedback, and interactive modeling, including:

[0037] The dynamic feature vector of the target vehicle is injected into the traffic knowledge graph so that the static road traffic rule information in the traffic knowledge graph is updated to dynamic context information including real-time traffic situation.

[0038] The semantically gating graph convolution operator is invoked to propagate dynamic traffic information on the updated traffic knowledge graph;

[0039] Each dynamic element feature on the updated traffic knowledge graph reads the propagated traffic dynamic information to enhance the scene context information of its own features;

[0040] Interaction modeling is performed based on dynamic feature characteristics following all enhanced scene context information.

[0041] Further, hierarchical trajectory prediction is performed based on the interactive intent feature vector between the target vehicle and the lane, as well as the motion features of the target vehicle, including:

[0042] Driving intentions are identified and classified based on the interactive intention feature vector between the target vehicle and the lane, as well as the motion characteristics of the target vehicle.

[0043] For each driving intention to be determined, a trajectory probability prediction for that driving intention is performed to obtain the probabilistic trajectory under the driving intention.

[0044] Construct a target loss function based on intent classification and trajectory generation;

[0045] The target loss function is trained to obtain the multimodal driving intentions of the target vehicle and the probability trajectory of each modal driving intention.

[0046] Furthermore, the expression for the target loss function is:

[0047]

[0048] in, Describes the target loss function. Indicates the intention classification loss, λ cls The weights represent the intended classification loss. λ represents the trajectory generation loss. reg The weights represent the trajectory generation loss.

[0049] Among them, intention classification loss The expression is:

[0050]

[0051] Trajectory generation loss The expression is:

[0052]

[0053] As another aspect of the present invention, a vehicle trajectory prediction device for right-of-way conflict scenarios is provided, for implementing the vehicle trajectory prediction method for right-of-way conflict scenarios described above, wherein the device includes:

[0054] The acquisition module is used to acquire map data with semantic information and historical trajectory data of the target vehicle, respectively.

[0055] The static scene understanding module is used to perform static scene understanding based on the map data with semantic information to obtain road traffic rule information;

[0056] The dynamic feature extraction module is used to extract dynamic feature characteristics based on the historical trajectory data of the target vehicle to obtain the motion characteristics of the target vehicle.

[0057] An interactive intent recognition module is used to perform interactive intent recognition based on the road traffic rule information and the target vehicle motion characteristics to obtain the interactive intent feature vector of the target vehicle and the lane.

[0058] The hierarchical trajectory prediction module is used to perform hierarchical trajectory prediction based on the interactive intention feature vector between the target vehicle and the lane and the motion features of the target vehicle, so as to obtain the multimodal driving intention of the target vehicle and the probability trajectory of each modal driving intention.

[0059] As another aspect of the present invention, a vehicle trajectory prediction system for right-of-way conflict scenarios is provided, comprising: an input device for inputting map data and historical trajectory data, an output device for outputting multimodal prediction results, and the vehicle trajectory prediction device for right-of-way conflict scenarios described above, wherein the vehicle trajectory prediction device for right-of-way conflict scenarios is communicatively connected to the input device and the output device respectively.

[0060] This invention provides a vehicle trajectory prediction method for right-of-way conflict scenarios. It acquires map data with semantic information and historical trajectory data of the target vehicle. Static scene understanding is performed based on the semantically information-rich map data, and dynamic feature extraction is performed based on the target vehicle's historical trajectory data. The results of static scene understanding and dynamic feature extraction are then used for interactive intent identification. Finally, hierarchical trajectory prediction is performed based on the results of interactive intent identification to obtain the target vehicle's multimodal driving intent and the probability trajectory of each modality. This vehicle trajectory prediction method for right-of-way conflict scenarios addresses the lack of semantic information in static environment representation through static scene understanding. The extraction of dynamic features allows for the fusion of the obtained dynamic features with the static scene results, enabling logical reasoning on a knowledge graph. This fundamentally improves the reliability and security of prediction results in complex, rule-intensive scenarios. Finally, hierarchical trajectory prediction provides a clear causal reasoning path from intent to behavior, enhancing the interpretability of the prediction results. Therefore, the vehicle trajectory prediction method for right-of-way conflict scenarios provided by this invention can fully capture high-fidelity topology results, and improve the robustness, interpretability and verifiability of the prediction results while ensuring high accuracy of trajectory prediction. Attached Figure Description

[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof.

[0062] Figure 1 The flowchart shows the vehicle trajectory prediction method for right-of-way conflict scenarios provided by this invention.

[0063] Figure 2 A flowchart illustrating the static scene understanding method provided by this invention.

[0064] Figure 3 A flowchart illustrating the specific implementation of static scene understanding provided by this invention.

[0065] Figure 4 The flowchart of the dynamic feature extraction method provided by the present invention.

[0066] Figure 5 The flowchart of the hierarchical trajectory prediction method provided by the present invention is shown.

[0067] Figure 6 This is a structural block diagram of the vehicle trajectory prediction device for right-of-way conflict scenarios provided by the present invention.

[0068] Figure 7 This is a structural block diagram of the vehicle trajectory prediction system for right-of-way conflict scenarios provided by the present invention. Detailed Implementation

[0069] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0070] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. 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 should fall within the scope of protection of the present invention.

[0071] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0072] This embodiment provides a vehicle trajectory prediction method for right-of-way conflict scenarios. Figure 1 This is a flowchart of a vehicle trajectory prediction method for right-of-way conflict scenarios provided by an embodiment of the present invention, as shown below. Figure 1 As shown, it includes:

[0073] S100: Acquire map data with semantic information and historical trajectory data of the target vehicle, respectively;

[0074] In this embodiment of the invention, high-definition map data with semantic information can be obtained through some map providers (such as Baidu Maps, Google Maps, etc.); in addition, the historical trajectory data of the target vehicle can be obtained through the vehicle's onboard sensors, etc.

[0075] S200: Perform static scene understanding based on the map data with semantic information to obtain road traffic rule information;

[0076] In this embodiment of the invention, the original vectorized high-definition map and its accompanying semantic annotation layer are parsed and constructed into a structured traffic knowledge graph. This traffic knowledge graph not only represents the geometric shape and topological connectivity of roads, but also explicitly encodes deep traffic semantics such as traffic rules and lane functional attributes.

[0077] S300: Extract dynamic feature elements based on the historical trajectory data of the target vehicle to obtain the motion features of the target vehicle;

[0078] In this embodiment of the invention, kinematic features that can characterize the inherent, context-free kinematic features of traffic participants, such as historical speed, acceleration, and turning angle, are extracted from their historical time-series data.

[0079] S400. Based on the road traffic rule information and the target vehicle motion characteristics, perform interactive intent recognition to obtain the interactive intent feature vector between the target vehicle and the lane.

[0080] Specifically, through iterative information fusion, the kinematic features of dynamic elements are deeply integrated with the scene rules and context in the traffic knowledge graph. The aim is to generate a highly contextualized final feature representation for each participant to support subsequent intention decisions.

[0081] S500: Perform hierarchical trajectory prediction based on the interactive intention feature vector between the target vehicle and the lane and the motion features of the target vehicle to obtain the multimodal driving intention of the target vehicle and the probability trajectory of each modal driving intention.

[0082] Specifically, based on the obtained interactive intention feature vector of the target vehicle and the lane, as well as the motion features of the target vehicle, a hierarchical prediction process is performed: First, the high-level driving intentions of the target vehicle (such as left turn, lane change) are explicitly classified; then, based on the intention, a probability distribution of the future trajectory is generated, thereby realizing the structured quantification of uncertainty and producing interpretable prediction results.

[0083] In summary, the vehicle trajectory prediction method for right-of-way conflict scenarios provided by this invention acquires map data with semantic information and historical trajectory data of the target vehicle. It performs static scene understanding based on the semantically information-rich map data and dynamic feature extraction based on the target vehicle's historical trajectory data. The results of static scene understanding and dynamic feature extraction are then used for interactive intent identification. Finally, hierarchical trajectory prediction is performed based on the results of interactive intent identification to obtain the target vehicle's multimodal driving intent and the probability trajectory of each modality. This vehicle trajectory prediction method for right-of-way conflict scenarios addresses the lack of semantic information in static environment representation through static scene understanding. The extraction of dynamic features allows for the fusion of the obtained dynamic features with the static scene results, enabling logical reasoning on a knowledge graph. This fundamentally improves the reliability and security of prediction results in complex, rule-intensive scenarios. Finally, hierarchical trajectory prediction provides a clear causal reasoning path from intent to behavior, enhancing the interpretability of the prediction results. Therefore, the vehicle trajectory prediction method for right-of-way conflict scenarios provided by this invention can fully capture high-fidelity topology results, and improve the robustness, interpretability and verifiability of the prediction results while ensuring high accuracy of trajectory prediction.

[0084] In this embodiment of the invention, static scene understanding is performed on map data with semantic information. Specifically, the static vectorized high-definition map containing rich traffic rules and functional definitions is transformed into a structured traffic knowledge graph that can be used by a deep learning model for rule reasoning and behavioral constraint analysis.

[0085] Specifically, static scene understanding is performed based on the map data containing semantic information to obtain road traffic rule information, such as... Figure 2 As shown, it includes:

[0086] S210. Based on the map data with semantic information, define graph nodes, parameterize node features, define graph edges and semantic annotations to obtain a traffic knowledge graph containing semantic information.

[0087] In this embodiment of the invention, to avoid information loss and explicitly encode traffic rules (specifically, to treat traffic rules as a structured, directly readable data feature, rather than having the model passively and implicitly learn from massive amounts of behavior), the invention can directly construct a semantically rich traffic knowledge graph from vectorized data through the following process. This knowledge graph can be formally defined as an attributed, directed, heterogeneous graph.

[0088] It should be noted that, This represents the basic physical unit that constitutes the entire road network. In this embodiment of the invention, each node... This does not represent an entire abstract lane, but rather a specific, directed, and refined physical line segment that constitutes the centerline of a lane. It has a definite start point, end point, and physical form; ε represents the logical relationship between these basic physical units (lane segments), where edges define the possibilities and ways in which a vehicle can travel from one segment to another, and are a mathematical expression of the road topology. For example, the edges in the Asuc set represent all legal connections that "proceed along the lane." It represents the traffic function and regulatory attributes carried by each basic physical unit (lane segment). This is a direct digitization of the rules of the physical world. For example, the semantic attributes of a node can clearly indicate that it is "a road with a speed limit of 30km / h" or "it belongs to a pedestrian crossing area". This represents the right-of-way rules that different physical units need to follow when interacting. If the semantics of a node is "what this place is", then the semantics of an edge is "what to do when going from one place to another". For example, the semantics of an edge can clearly stipulate that when entering lane B from lane A through this connection, "vehicles in lane B need to give way".

[0089] Specifically, the map data with semantic information is vectorized high-definition map data and its accompanying semantic annotation layer, which includes information such as lane line type, speed limit, traffic signs, and functional areas (such as intersections and merging lanes).

[0090] Graph node definition Each node in the heterogeneous graph G This represents a directed line segment between any two consecutive points constituting the lane centerline. This line segment can be further refined with geometric and semantic information. It should be noted that the edge types in the heterogeneous graph are heterogeneous. In this embodiment of the invention, at least four edge types with completely different physical and logical meanings are explicitly defined: suc (suc), pre (predecessor), left (left adjacent), and right (right adjacent). The node types are also heterogeneous: "lane segment nodes" and "dynamic element nodes (i.e., vehicles, pedestrians, etc.)".

[0091] Node feature parameterization: the initial feature vector h for each node i It is designed to comprehensively represent information across three dimensions: geometry, location, and traffic semantics. Its mathematical expression is:

[0092] h i =Concat(h i,geo ,h i,loc ,h i,sem ),

[0093] Among them, geometric features (h) i,geo ) indicates that through a multilayer perceptron (MLP) geo Difference between the starting and ending coordinates of a line segment Encode it to capture its physical shape, such as length and orientation.

[0094] Location features (h) i,loc Specifically, this means through MLP loc coordinates of the center point of the line segment Encode it and anchor it in the scene's global coordinate system. It should be noted that MLP... loc Specifically, it can be understood as a position encoder, MLP loc Specifically, this means converting a simple, absolute physical coordinate (e.g., [x = 123.4, y = 567.8]) into a "position feature vector".

[0095] Semantic features (h) i,sem Specifically, this means through MLP sem The original semantic label vectors extracted from the map semantic layer that describe the functional attributes of nodes Encode the semantic tags The functional attributes of the lane segment are clearly described, including: whether it is located inside an intersection, the speed limit, whether it is a solid line boundary prohibiting overtaking, whether there is a stop line, and the lane function type (e.g., dedicated left-turn lane). It should be noted that MLP... sem Specifically, it can be understood as a semantic encoder, MLP semThis means converting a set of discrete, multimodal raw semantic labels (e.g., [speed limit = 60, type = left turn lane, stop line = yes]) into a "semantic feature vector".

[0096] Graph edge definition and semantic annotation Specifically, the edges of the graph are defined as four types with explicit topological semantics, represented by four independent adjacency matrices {A}. suc A pre A left A right} represents the successor, predecessor, left adjacent, and right adjacent relationships, respectively. Each edge e ji (from node v) j to v i It is also assigned a semantic feature vector. Its dimension is P, which describes the traffic rules of this connection. An adjacent connection edge of a lane merging operation is labeled "Yield required for vehicles on the main road"; a subsequent connection edge from the straight lane to the left-turn lane is labeled "Controlled by the left-turn signal".

[0097] S220. Construct a semantically gated graph convolution operator and apply the semantically gated graph convolution operator to the traffic knowledge graph containing semantic information to assist the traffic knowledge graph containing semantic information in information dissemination with logical reasoning ability.

[0098] In this embodiment of the invention, after constructing a traffic knowledge graph, a semantically-gated graph convolution operator is specifically constructed to perform information dissemination with logical reasoning capabilities on the traffic knowledge graph.

[0099] In this embodiment of the invention, constructing a semantically gated graph convolution operator includes:

[0100] 1) Determine the semantic features of the connecting edges formed when information is transmitted between neighboring nodes and the current node;

[0101] 2) Determine the semantic gate value based on the semantic features of the connected edges, wherein the expression for the semantic gate value is:

[0102]

[0103] Among them, g ji,k The semantic gate value is represented by σ, which represents the Sigmoid function. Indicate the semantic features of the connecting edges;

[0104] 3) Modulate the information stream according to the semantic gating value to obtain modulated aggregated information, wherein the expression of the modulated aggregated information is:

[0105]

[0106] Where, m ji,k This represents the modulated aggregated information. This represents the characteristics of neighboring nodes at layer l. The symbol represents the weight matrix, and ⊙ represents the Hadamard product;

[0107] 4) Determine the semantically gated graph convolution operator based on the modulated aggregation information. The expression of the semantically gated graph convolution operator is as follows:

[0108]

[0109] in, This represents the complete update rule of the current node at level l+1. This represents the feature of the current node at layer l. Let k represent the weight matrix. k∈{suc,pre,...} represents the current feature from node i. Which information to retain, enhance, or transform in order to effectively combine with incoming neighbor information, i.e., multiply the corresponding features by a weight.

[0110] In this embodiment of the invention, the subscript *suc* denotes a successor relationship (denoted as *suc*), which defines the direct connectivity of nodes in the graph along the legal driving direction. If lane segment node *vj* can be directly reached from lane segment node *vi*, then *vi* is said to be a successor node of *vj*. The successor relationship chain constitutes the main extension path of the lane and is the basis for the model to predict longitudinal behavior (such as following other vehicles and going straight). The subscript *pre* denotes a predecessor relationship (denoted as *pre*), which is the reverse relationship of the successor relationship. If lane segment node *vj* can be directly reached from lane segment node *vi*, then *vj* is said to be a predecessor node of *vi*. The predecessor relationship defines the upstream source of traffic flow and helps the model understand the history and context of the scene. The left neighbor relationship (denoted as left) defines the lane segment node that is spatially closest to the left side of the current lane and has the potential to run parallel or change lanes. The right neighbor relationship (denoted as right) defines the lane segment node that is spatially closest to the right side of the current lane and has the potential to run parallel or change lanes. The left and right neighbor relationships are the key basis for the model to predict lateral behaviors (such as lane changing and avoidance).

[0111] It should be noted that traditional graph convolution treats all connections equally, but the flow of information on the graph should strictly adhere to traffic rules. For example, the influence of information from a lane that needs to yield should be dynamically adjusted according to the current situation. Based on this, this invention constructs a semantically gating graph convolution operator to make decisions more consistent with human driver logic.

[0112] Specifically, the semantic gating mechanism implements the above idea by introducing a semantic gating unit. For example... Figure 3 As shown, when information comes from neighbor node v j Passed to v i At that time, the gating unit will read the connection edge e ji semantic features And generate a gate value g ji,k To modulate the information stream. The modulated aggregated information m ji,k The calculation is as follows:

[0113]

[0114] in, Indicates v j Features in layer l Let represent the learnable weight matrix, and ⊙ represent the Hadamard product. The semantic gate value g. ji,k It is itself calculated by a small gated network based on the semantic features of the edges:

[0115]

[0116] Here, σ represents the Sigmoid function, which constrains the gate value to the interval (0,1). Finally, node v... i The complete update rule at level l+1 is:

[0117]

[0118] In one specific embodiment, when an edge is marked as "needs to yield", its corresponding gating network can learn to output a gating value close to 0 under specific conditions, thereby effectively suppressing the information flow from that direction. This makes the model's decision-making process more in line with the risk avoidance logic of human drivers.

[0119] Multi-scale and long-range dependency specifically means that the semantic gating operator can be seamlessly integrated with lane-drilled convolution and multi-scale fusion modules, enabling the model to not only understand local traffic rules, but also to infer long-range strategic driving paths determined by both road network structure and traffic rules.

[0120] It should be understood that the implementation of scene understanding and information dissemination in this embodiment of the invention is illustrated using a graph information dissemination mechanism based on traffic semantic perception as an example. Its essence is to utilize the semantic features of edges to gate or modulate the information flow between nodes. Besides the specific implementation given in this embodiment, other graph neural network architectures capable of achieving similar functions can also be used to achieve the same goal of dynamically adjusting information weights according to traffic rules.

[0121] In embodiments of the present invention, such as Figure 4 As shown, dynamic feature extraction is performed based on the historical trajectory data of the target vehicle, including:

[0122] S310. Represent the historical trajectory data of the target vehicle as a historical displacement vector;

[0123] S320. Map the historical displacement vector to a temporal convolutional network of a preset dimension to obtain the dynamic feature vector of the target vehicle.

[0124] In this embodiment of the invention, the core objective of processing the historical trajectories of all dynamic elements (i.e., traffic participants) in the scene is to extract a feature vector from these time-series data that can characterize their inherent kinematic patterns and short-term behavioral tendencies. This feature is a quantitative description of the motion history of the dynamic element itself without considering the complex semantics of the environment.

[0125] Specifically, the observed trajectory of the target dynamic element i over the past T time steps is represented as a series of displacement vectors {Δp}. -(T-1) ,…,Δp0}。 In this embodiment of the invention, a one-dimensional temporal convolutional network (1D-CNN) is used as the backbone network.

[0126] The process of extracting dynamic feature elements can be formally described as a function f. dyn It maps the trajectory sequence to a fixed-dimensional feature vector. This output vector is defined as the Initial Dynamic Behavior Feature Vector, denoted as...

[0127]

[0128] in, D dyn This represents the dimension of the feature vector. It is worth noting that... It primarily encodes the context-free motion information of this element: historical velocity, acceleration, and angular rate.

[0129] When extracting dynamic features, an initial dynamic behavior feature vector is output for each dynamic feature i in the scene. This vector will serve as the initial input representing the dynamic element in subsequent interactive driving intent recognition.

[0130] It should be understood that the embodiments of the present invention specifically use networks such as 1D-CNN to extract features for dynamic / static feature extractors. However, any neural network model that can effectively process time-series data or point cloud data, such as recurrent neural networks (RNN / LSTM / GRU), can be used to replace the corresponding feature extraction module. This does not affect the core process of hierarchical prediction in the present invention.

[0131] In this embodiment of the invention, interactive intent recognition is performed based on the road traffic rule information and the target vehicle motion characteristics, including:

[0132] Based on the road traffic rules information and the target vehicle motion characteristics, an iterative interactive fusion is performed to obtain the interactive intent feature vector of the target vehicle and the lane. In each iteration, an interactive process of information injection, information propagation, information feedback and interactive modeling is performed.

[0133] In this embodiment of the invention, the interactive intent recognition process deeply integrates "static scene knowledge" containing rich traffic rules and "dynamic element behavior" extracted from dynamic elements through an iterative interactive process. The ultimate goal is to generate a fully contextualized behavioral feature vector for each dynamic element, designed to directly support subsequent modules in performing complex high-level intent judgments.

[0134] Specifically, scene knowledge graph feature set Each A feature vector represents a lane segment node, which contains the node's geometry, location, and key traffic semantic information.

[0135] Initial dynamic behavior feature set Each This represents the initial feature vector of a dynamic element.

[0136] Specifically, each iteration involves an interactive process of information injection, information propagation, information feedback, and interactive modeling, including:

[0137] (1) Inject the dynamic feature vector of the target vehicle into the traffic knowledge graph so that the road traffic rule information represented as static in the traffic knowledge graph is updated to dynamic context information including real-time traffic situation.

[0138] In this embodiment of the invention, the iterative interactive fusion process is specifically composed of M identical interactive blocks stacked together. In the l-th interactive block, information is processed and evolved strictly according to four logical steps: information injection, information propagation, information feedback, and interactive modeling. The graph node features at the beginning of the l-th cycle are denoted as... Dynamic element characteristics are denoted as

[0139] Specifically, the information injection (Actor-to-Lane, A2L) step injects the real-time motion state of each dynamic element into the traffic knowledge graph. Through a spatial attention mechanism, each graph node j aggregates the information of its nearby dynamic elements. Features are thus updated from purely static descriptions to dynamic contextual information including real-time traffic conditions. The expression for information injection is:

[0140]

[0141] (2) Call the semantically gating graph convolution operator to propagate traffic dynamic information on the updated traffic knowledge graph;

[0142] Specifically, the Lane-to-Lane (L2L) information propagation step disseminates traffic dynamics information across the updated knowledge graph. Crucially, the information propagation operator invoked here is the graph information propagation operator based on traffic semantic awareness constructed earlier. This ensures that the flow of information follows traffic rules; for example, congestion information propagates along "successor" relationships without easily crossing the "no passage" semantic boundary. The specific expression for information propagation is:

[0143]

[0144] Among them, f L2L Internally, the aforementioned semantically gated graph convolution is performed.

[0145] (3) Each dynamic element feature on the updated traffic knowledge graph reads the disseminated traffic dynamic information to enhance the scene context information of its own features;

[0146] In this embodiment of the invention, the Lane-to-Actor (L2A) step enables each dynamic element to "read" and "understand" the scene state, rich in dynamic and rule-based information, after global propagation. Through a spatial attention mechanism, each dynamic element i aggregates the latest features of its surrounding graph nodes. This allows its own features to be enhanced by the scene context.

[0147] (4) Perform interactive modeling based on the dynamic feature characteristics after all enhanced scene context information.

[0148] In this embodiment of the invention, after all dynamic elements have fully perceived the scene context, the interaction modeling (Actor-to-Actor, A2A) step performs the final modeling of their direct interactions. Since the features of each element at this point have incorporated rich scene semantics and rule information, this interaction modeling is based on behavioral game theory with a high level of understanding of the scene. The specific expression for interaction modeling is:

[0149]

[0150] After M iterations, this interaction process outputs a final contextualized behavior feature vector for each dynamic element i, denoted as .

[0151]

[0152] The vector It was designed to highly condense all the information needed to support high-level decision-making and to serve directly as input for subsequent hierarchical trajectory generation.

[0153] In this embodiment of the invention, hierarchical trajectory prediction is performed based on the interactive intent feature vector between the target vehicle and the lane, and the motion features of the target vehicle, such as... Figure 5 As shown, it includes:

[0154] S510. Based on the interactive intention feature vector between the target vehicle and the lane and the motion features of the target vehicle, driving intention is identified and classified.

[0155] In this embodiment of the invention, the traditional "black box" multimodal trajectory regression method is completely abandoned, and a hierarchical generative model that is more in line with human driving decision-making logic is adopted.

[0156] (1) High-level driving intention identification and classification.

[0157] Objective: To determine the final behavioral feature vector based on previously obtained, deeply integrated scene elements. Explicitly infer the high-level driving intentions of the target vehicle.

[0158] Intent set definition: Based on the scenario type, a discrete set of driving intentions is predefined.

[0159] Intent Classifier: An Intent Classification Network (MLP) intent Received feature vector It outputs a K-dimensional confidence score (logits) z. i ,

[0160]

[0161] This K-dimensional vector is transformed into a probability distribution c using the SoftMax function. i =[c i,1 ,…,c i,K ], where each component This represents the confidence level of the model in predicting that the vehicle will perform the k-th intention.

[0162] c i =Softmax(z i ).

[0163] S520. For each driving intention to be determined, perform trajectory probability prediction for that driving intention to obtain the probabilistic trajectory under the driving intention.

[0164] In this embodiment of the invention, a specific execution path is generated for each possible driving intention. The generated path is not a single deterministic trajectory, but rather a probability distribution that describes the uncertainty of the execution.

[0165] Probabilistic representation of the trajectory: the vehicle's position (x) at the t-th time step in the future. (t) ,y (t) It is modeled as a bivariate Gaussian distribution, described by five parameters: location mean. Standard deviation and correlation coefficient ρ (t) ,

[0166] Intent conditionalization: Each intent I k It is encoded into a learnable embedding vector e k Embed this intention with vehicle features Concatenate the vectors to form an intention-conditional input vector.

[0167] Probabilistic Decoder: A Trajectory Decoding Network traj Receive the conditional input and regress the intent I. k The corresponding set of distributed parameters Ψ for all future time steps k :

[0168]

[0169] S530. Construct a target loss function based on intent classification and trajectory generation;

[0170] In this embodiment of the invention, the expression for the target loss function is:

[0171]

[0172] in, Describes the target loss function. Indicates the intention classification loss, λ cls The weights represent the intended classification loss. λ represents the trajectory generation loss. reg The weights represent the trajectory generation loss.

[0173] Among them, intention classification loss The expression is:

[0174]

[0175] In this embodiment of the invention, the intention is to classify the loss. Specifically, it can be understood as the confidence score (or predicted probability) given by the prediction model for the "true driving intention" of the target vehicle i. i,gt Indicates from probability vector c i Extract the probability value of the item corresponding to the "true intention". For example, if "turn left" represents the second intention, and the model predicts a probability of c... i = [0.1, 0.7, 0.2], and the vehicle did indeed turn left (i.e., its true intention was to "turn left"), then c i,gt The value is 0.7.

[0176] Trajectory generation loss The expression is:

[0177]

[0178] It should be understood that, given the vehicle's true intentions, the actual location of the vehicle is not reflected in the trajectory probability "cloud" predicted by the model. The likelihood of it occurring (i.e., probability density); The actual coordinates of the vehicle at time t in the future are a precise point. The model predicts not a line, but a probability distribution (described by a bivariate Gaussian distribution). Therefore... This indicates the probability that the actual location point falls within the predicted distribution.

[0179] Specifically, in this embodiment of the invention, a specially designed combination loss function is used. For end-to-end training, this function forces the model to learn the intrinsic relationship between intent and trajectory:

[0180]

[0181] Intended classification loss The standard cross-entropy loss is used, aiming to ensure the model correctly classifies the baseline (Ground Truth). gt The expression for the intention classification loss is:

[0182]

[0183] Trajectory generation loss Negative log-likelihood loss (NLL) is employed. This loss only applies to the baseline intention I. gt The matching prediction distribution Ψ gt Its goal is to adjust the parameters of the distribution so that the baseline trajectory Y gt To maximize the probability of occurrence under this distribution, the expression for the trajectory generation loss is:

[0184]

[0185] S540. Train the target loss function to obtain the multimodal driving intentions of the target vehicle and the probability trajectory of each modal driving intention.

[0186] In this embodiment of the invention, the final output is a structured set of prediction results. For each target vehicle, K tuples (intent probability, trajectory distribution parameters) are output, i.e.

[0187] It should be understood that, regarding the representation of trajectory probability distribution, the embodiments of this invention employ a binary Gaussian distribution to parameterize the trajectory uncertainty. To model more complex trajectory distributions, it can also be replaced with a Mixture Density Network (MDN) to output the parameters of the Gaussian mixture model; although these methods differ in specific form, their fundamental purpose is to achieve conditional trajectory distribution prediction based on specific intentions, consistent with the core objective of this invention.

[0188] In summary, the vehicle trajectory prediction method for right-of-way conflict scenarios provided by this invention constructs and utilizes a traffic knowledge graph for semantic reasoning, and adopts a hierarchical intent anchoring framework for interpretable prediction. Specifically, (1) Scene understanding and semantic reasoning based on traffic knowledge graph. The accurate trajectory prediction of this invention relies not only on the perception of scene geometry, but also on a deep understanding of traffic rules. To this end, this invention proposes a complete set of scene understanding and semantic reasoning methods. The key is: First, a high-definition map containing semantic information such as traffic functions and traffic rules is constructed into a structured traffic knowledge graph, thereby reducing the gap between the physical world and model cognition. Second, a semantic-based gating information propagation mechanism is constructed, enabling the model to perform logical reasoning on the knowledge graph and dynamically and selectively process information flow according to traffic rules (such as right-of-way and yielding relationships). This transforms the prediction model from a simple pattern recognizer into a reasoning engine with basic rule compliance capabilities, fundamentally improving the reliability and security of prediction results in complex, rule-intensive scenarios. (2) A structured trajectory prediction process that integrates semantic reasoning and intent decision-making. To address the black-box problem of the integrated "feature fusion-trajectory generation" in existing prediction models, a logically transparent and clearly structured prediction process is proposed. The complex prediction task is constructed as a complete information processing and decision-making chain, from independent feature extraction to deep interactive reasoning and hierarchical intent decision-making. First, this process decouples heterogeneous information sources in the scene, extracting the intrinsic kinematic features of dynamic elements and the traffic knowledge graph features of the static environment in parallel, forming two independent initial information flows with clear physical and semantic meanings. Then, an iterative interactive fusion method deeply and cyclically integrates these two sets of features. This process simulates the driver's cognitive process: injecting dynamic vehicle information into the semantic map, performing information propagation and reasoning according to traffic rules, feeding back scene information rich in global context to the vehicle, and finally modeling a fully "contextualized" vehicle-to-vehicle interaction. The output of this step is a highly condensed final behavioral feature vector with a deep understanding of the scene. Finally, this process does not directly use this final feature for trajectory regression, but rather for a hierarchical decision-making process: First, the feature is used to explicitly classify the vehicle's high-level driving intentions; then, using the identified intentions as strong constraints, a probability distribution of the future trajectory under that intention is generated. Therefore, the vehicle trajectory prediction method for right-of-way conflict scenarios provided by this invention, by constructing a computational framework that simulates the structured reasoning paradigm of human expert drivers, can improve the robustness, interpretability, and verifiability of the prediction results while ensuring high accuracy in trajectory prediction.

[0189] As another embodiment of the present invention, a vehicle trajectory prediction device 100 for right-of-way conflict scenarios is provided, which implements the vehicle trajectory prediction method for right-of-way conflict scenarios described above, wherein, as Figure 7 As shown, it includes:

[0190] The acquisition module 110 is used to acquire map data with semantic information and historical trajectory data of the target vehicle, respectively.

[0191] The static scene understanding module 120 is used to perform static scene understanding based on the map data with semantic information to obtain road traffic rule information;

[0192] The dynamic feature extraction module 130 is used to extract dynamic feature characteristics based on the historical trajectory data of the target vehicle to obtain the motion characteristics of the target vehicle.

[0193] The interactive intent recognition module 140 is used to perform interactive intent recognition based on the road traffic rule information and the target vehicle motion characteristics to obtain the interactive intent feature vector of the target vehicle and the lane.

[0194] The hierarchical trajectory prediction module 150 is used to perform hierarchical trajectory prediction based on the interactive intention feature vector between the target vehicle and the lane and the motion features of the target vehicle, so as to obtain the multimodal driving intention of the target vehicle and the probability trajectory of each modal driving intention.

[0195] This invention provides a vehicle trajectory prediction device for right-of-way conflict scenarios. It acquires map data with semantic information and historical trajectory data of the target vehicle. It performs static scene understanding based on the semantically information-rich map data and dynamic feature extraction based on the target vehicle's historical trajectory data. The results of static scene understanding and dynamic feature extraction are then used for interactive intent recognition. Finally, hierarchical trajectory prediction is performed based on the interactive intent recognition results to obtain the target vehicle's multimodal driving intent and the probability trajectory of each modality. This vehicle trajectory prediction method for right-of-way conflict scenarios addresses the lack of semantic information in static environment representation through static scene understanding. The extraction of dynamic features allows for the fusion of the obtained dynamic features with the static scene results, enabling logical reasoning on a knowledge graph. This fundamentally improves the reliability and security of prediction results in complex, rule-intensive scenarios. Finally, hierarchical trajectory prediction provides a clear causal reasoning path from intent to behavior, enhancing the interpretability of the prediction results. Therefore, the vehicle trajectory prediction device for right-of-way conflict scenarios provided by the present invention can fully capture high-fidelity topology results, and improve the robustness, interpretability and verifiability of the prediction results while ensuring high accuracy of trajectory prediction.

[0196] The specific working principle of the vehicle trajectory prediction device for right-of-way conflict scenarios provided by this invention can be referred to the specific description of the vehicle trajectory prediction method for right-of-way conflict scenarios above, and will not be repeated here.

[0197] As another embodiment of the present invention, a vehicle trajectory prediction system 10 for right-of-way conflict scenarios is provided, wherein, as Figure 7 As shown, it includes: an input device 200 for inputting map data and historical trajectory data, an output device 300 for outputting multimodal prediction results, and the vehicle trajectory prediction device 100 for right-of-way conflict scenarios mentioned above. The vehicle trajectory prediction device 100 for right-of-way conflict scenarios is communicatively connected to the input device 200 and the output device 300, respectively.

[0198] The vehicle trajectory prediction system for right-of-way conflict scenarios provided by this invention, by employing the aforementioned output device for outputting multimodal prediction results, addresses the semantic information gap in static environment representation through static scene understanding. It also integrates the obtained dynamic feature extraction with the static scene results, enabling logical reasoning on a knowledge graph. This fundamentally improves the reliability and security of prediction results in complex, rule-intensive scenarios. Finally, hierarchical trajectory prediction yields a clear causal reasoning path from intent to behavior, enhancing the interpretability of the prediction results. Therefore, the vehicle trajectory prediction system for right-of-way conflict scenarios provided by this invention can fully capture high-fidelity topological results, improving the robustness, interpretability, and verifiability of prediction results while ensuring high accuracy in trajectory prediction.

[0199] The specific working principle of the vehicle trajectory prediction system for right-of-way conflict scenarios provided by this invention can be found in the detailed description of the vehicle trajectory prediction method for right-of-way conflict scenarios above, and will not be repeated here.

[0200] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting vehicle trajectories in right-of-way conflict scenarios, characterized in that, include: Acquire map data with semantic information and historical trajectory data of the target vehicle, respectively; Static scene understanding is performed based on the map data containing semantic information to obtain road traffic rule information; Dynamic feature extraction is performed on the historical trajectory data of the target vehicle to obtain the motion characteristics of the target vehicle. Based on the road traffic rules information and the target vehicle's motion characteristics, interactive intent recognition is performed to obtain the interactive intent feature vector between the target vehicle and the lane. Based on the interactive intention feature vector between the target vehicle and the lane and the motion features of the target vehicle, hierarchical trajectory prediction is performed to obtain the multimodal driving intention of the target vehicle and the probability trajectory of each modal driving intention. Specifically, static scene understanding is performed based on the map data containing semantic information to obtain road traffic rule information, including: Based on the map data with semantic information, graph nodes are defined, node features are parameterized, and graph edges are defined and semantically labeled to obtain a traffic knowledge graph containing semantic information. A semantically gated graph convolution operator is constructed and applied to the traffic knowledge graph containing semantic information to assist the traffic knowledge graph containing semantic information in information dissemination with logical reasoning ability. The hierarchical trajectory prediction based on the interactive intent feature vector between the target vehicle and the lane and the motion features of the target vehicle includes: Driving intentions are identified and classified based on the interactive intention feature vector between the target vehicle and the lane, as well as the motion characteristics of the target vehicle. For each driving intention to be determined, a trajectory probability prediction for that driving intention is performed to obtain the probabilistic trajectory under the driving intention. Construct a target loss function based on intent classification and trajectory generation; The target loss function is trained to obtain the multimodal driving intentions of the target vehicle and the probability trajectory of each modal driving intention.

2. The vehicle trajectory prediction method for right-of-way conflict scenarios according to claim 1, characterized in that, Constructing a semantically gated graph convolution operator, including: Determine the semantic features of the connection edges formed when information is transmitted between neighboring nodes and the current node; A semantic gate value is determined based on the semantic features of the connected edges, wherein the expression for the semantic gate value is: ， in, Indicates semantic gating value, This represents the Sigmoid function. Indicate the semantic features of the connecting edges; Modulated aggregated information is obtained by modulating the information stream according to semantic gating values, wherein the expression of the modulated aggregated information is: ， in, This represents the modulated aggregated information. Indicates the neighbor node at the th Features of the layer Represents the weight matrix. It represents the Hadamardi (or Hadama) stack; Based on the modulated aggregation information, a semantically gated graph convolution operator is determined, and the expression of the semantically gated graph convolution operator is as follows: ， in, Indicates the current node is at the . The complete update rules for the layer, Indicates the current node is at the . Features of the layer This represents the weight matrix.

3. The vehicle trajectory prediction method for right-of-way conflict scenarios according to claim 1, characterized in that, Dynamic feature extraction is performed based on the historical trajectory data of the target vehicle, including: The historical trajectory data of the target vehicle is represented as a historical displacement vector; The historical displacement vector is mapped onto a temporal convolutional network of a preset dimension to obtain the dynamic feature vector of the target vehicle.

4. The vehicle trajectory prediction method for right-of-way conflict scenarios according to claim 1, characterized in that, Interactive intent recognition is performed based on the road traffic rule information and the target vehicle's motion characteristics, including: Based on the road traffic rules information and the target vehicle motion characteristics, an iterative interactive fusion is performed to obtain the interactive intent feature vector of the target vehicle and the lane. In each iteration, an interactive process of information injection, information propagation, information feedback and interactive modeling is performed.

5. The vehicle trajectory prediction method for right-of-way conflict scenarios according to claim 4, characterized in that, Each iteration involves an interactive process of information injection, information propagation, information feedback, and interactive modeling, including: The dynamic feature vector of the target vehicle is injected into the traffic knowledge graph so that the static road traffic rule information in the traffic knowledge graph is updated to dynamic context information including real-time traffic situation. The semantically gating graph convolution operator is invoked to propagate dynamic traffic information on the updated traffic knowledge graph; Each dynamic element feature on the updated traffic knowledge graph reads the propagated traffic dynamic information to enhance the scene context information of its own features; Interaction modeling is performed based on dynamic feature characteristics following all enhanced scene context information.

6. The vehicle trajectory prediction method for right-of-way conflict scenarios according to claim 1, characterized in that, The expression for the target loss function is: ， in, Describes the target loss function. Indicates the intention to classify losses. The weights represent the intended classification loss. This represents the trajectory generation loss. The weights represent the trajectory generation loss. Among them, intention classification loss The expression is: ， Trajectory generation loss The expression is: ， in, This represents the true coordinates of the vehicle at time t in the future. This represents the predicted distribution that matches the true value intention.

7. A vehicle trajectory prediction device for right-of-way conflict scenarios, used to implement the vehicle trajectory prediction method for right-of-way conflict scenarios as described in any one of claims 1 to 6, characterized in that, include: The acquisition module is used to acquire map data with semantic information and historical trajectory data of the target vehicle, respectively. The static scene understanding module is used to perform static scene understanding based on the map data with semantic information to obtain road traffic rule information; The dynamic feature extraction module is used to extract dynamic feature characteristics based on the historical trajectory data of the target vehicle to obtain the motion characteristics of the target vehicle. An interactive intent recognition module is used to perform interactive intent recognition based on the road traffic rule information and the target vehicle motion characteristics to obtain the interactive intent feature vector of the target vehicle and the lane. The hierarchical trajectory prediction module is used to perform hierarchical trajectory prediction based on the interactive intention feature vector between the target vehicle and the lane and the motion features of the target vehicle, so as to obtain the multimodal driving intention of the target vehicle and the probability trajectory of each modal driving intention.

8. A vehicle trajectory prediction system for right-of-way conflict scenarios, characterized in that, include: An input device for inputting map data and historical trajectory data, an output device for outputting multimodal prediction results, and a vehicle trajectory prediction device for right-of-way conflict scenarios as described in claim 7, wherein the vehicle trajectory prediction device for right-of-way conflict scenarios is communicatively connected to the input device and the output device respectively.

Images