Traffic Scene Reuse Method Based on Causality and Uncertainty Fusion

Abstract

This disclosure relates to intelligent transportation and autonomous driving technologies, specifically a traffic scene reuse method based on the fusion of causality and uncertainty. The method includes constructing a probabilistic dynamic scene graph sequence; adaptively updating each probabilistic dynamic scene graph based on edge features to obtain an updated scene graph sequence; distinguishing between true causal relationships and spurious correlations on the edges of each updated scene graph to obtain a causally valid scene graph sequence; and identifying recurring traffic scene interaction patterns based on the causally valid scene graph sequence. This solution can address the problem that existing scene graph construction methods in the field of autonomous driving struggle to reflect the uncertainties brought about by perceptual noise and the diversity of driving behavior, and that the inability to distinguish between true causal relationships and spurious correlations easily leads to prediction failures in out-of-distribution scenarios. This allows for better extraction and reuse of typical traffic scene interaction patterns, improving the safety and reliability of autonomous driving systems in complex environments.

Description

Technical Field

[0001] This disclosure relates to intelligent transportation and autonomous driving technologies, and in particular to a method for reusing traffic scenarios based on the fusion of causality and uncertainty. Background Technology

[0002] With the development of autonomous driving technology, vehicles need to perform real-time perception, prediction, and decision-making in complex and ever-changing traffic environments. Although traditional end-to-end methods can directly map sensor inputs to control outputs, they lack explicit modeling of the various traffic participants and their interactions in the scenario, resulting in insufficient generalization and safety in complex interactions and rare, high-risk scenarios.

[0003] In recent years, scene graph modeling has been gradually introduced into the field of autonomous driving, representing nodes (vehicles, pedestrians, traffic lights, etc.) and edges (spatial relationships, semantic relationships, etc.) in traffic scenes in a structured way, providing an intermediate layer for downstream decision-making. However, existing methods still have the following shortcomings: First, most scene graph construction methods only output a single deterministic relationship, which is difficult to reflect the uncertainty brought about by perceptual noise and the diversity of driving behavior; second, traditional relationship modeling mainly relies on statistical correlation and lacks a causal inference mechanism, making it impossible to distinguish between true causal relationships and spurious correlations, thus easily leading to prediction failures in out-of-distribution scenarios. Summary of the Invention

[0004] To address the aforementioned issues in existing technologies, this disclosure proposes a traffic scene reuse method based on the fusion of causality and uncertainty. This method takes complex traffic environments as the research object and unfolds at three levels: scene abstraction, relationship optimization, and knowledge extraction. It gradually forms a scene representation and knowledge base with robustness, interpretability, and knowledge reuse capabilities, which can provide safer and more reliable support for end-to-end autonomous driving.

[0005] Specifically, a traffic scene reuse method based on the fusion of causality and uncertainty includes: constructing a probabilistic dynamic scene graph sequence. t represents the time marker, T represents the total number of times, and V represents the set of nodes, where nodes represent traffic participants. The edges between nodes represent relationships between traffic participants. Edge features include geometric, topological, semantic, and uncertain relationships. Geometric relationships include the distance and relative speed between two nodes. Topological relationships include whether two nodes are in the same lane, adjacent lanes, or conflicting lanes. Semantic relationships include whether two nodes are bound by traffic rules. Uncertain relationships are defined as edges satisfying independent and identically distributed (ICD). Each probabilistic dynamic scene graph is adaptively updated based on edge features to obtain an updated scene graph sequence. For each updated scene graph, edges are distinguished between true causal relationships and false correlations to obtain a causally valid scene graph sequence. Based on the causally valid scene graph sequence, recurring traffic scene interaction patterns are identified.

[0006] In one embodiment of the above technical solution, each probabilistic dynamic scene graph is adaptively updated based on edge features, including: using a relational graph attention network, assigning independent attention heads to geometric relations, topological relations, semantic relations and uncertain relations respectively, to obtain attention weights representing different relations between each node and its neighboring nodes; and updating the node based on the attention weights of different relations between each node and its neighboring nodes.

[0007] In one embodiment of the above technical solution, the attributes of the node include the location, speed, acceleration, orientation angle, category, and detection confidence of the traffic participant.

[0008] In one embodiment of the above technical solution, the probability distribution of the edge is: , the mean As an estimate of the relationship strength between the i-th node and the j-th node, the standard deviation The relationship between the i-th node and the j-th node is uncertain.

[0009] In one embodiment of the above technical solution, distinguishing between real causal relationships and spurious correlations for each updated scene graph edge includes: based on the updated scene graph Predicted output If for a certain edge The prediction result obtained after deletion is The importance of an edge is then quantified as ,when Exceeding the threshold At that time, it is considered that the edge Edges that have a causal effect on downstream predictions are retained as causally valid edges; otherwise, they are deleted. .

[0010] In one embodiment of the above technical solution, based on a causally valid scene graph sequence, a frequent subgraph mining algorithm is used to identify recurring high-order subgraph patterns, and recurring interaction scene patterns can be obtained through these high-order subgraph patterns.

[0011] In one embodiment of the above technical solution, the construction step of the probabilistic dynamic scene graph includes: smoothing the initial scene graph constructed based on the acquired image data, wherein the smoothing process updates the nodes and edges in the initial scene graph through an exponential moving average mechanism.

[0012] In one embodiment of the above technical solution, the traffic participants include vehicles, surrounding vehicles, pedestrians, cyclists, traffic lights, stop lines, lane lines, road boundaries, zebra crossings, and static obstacles.

[0013] The beneficial technical effects of this disclosure are: This solution can better extract and reuse typical traffic scenario interaction patterns, improve the decision-making efficiency of autonomous driving systems, and thus enhance the safety and reliability of autonomous driving systems in complex environments.

[0014] Based on the above technical solutions, a risk-driven method for constructing a traffic scenario knowledge base is further proposed. This method calculates the probability of a violation occurring when a high-order subgraph pattern is identified. , The subgraph corresponding to the identified higher-order subgraph pattern; the identified higher-order subgraph patterns are stored as structured knowledge entries: ,in, Represents the frequency of occurrence of higher-order subgraph patterns. The scene label represents the scenario, and k is the knowledge item identifier. In one embodiment of the above technical solution, the scene label includes intersection, lane merging, and following. This technical solution is specifically designed for rapid identification of risky scenarios, which facilitates rapid risk avoidance decisions and improves the safety of autonomous driving. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 , one A schematic diagram of a method for dynamic graph relationship modeling and high-order pattern extraction based on uncertainty and causal reasoning. Detailed Implementation

[0017] As can be seen from the background technology, the existing scene graph modeling technology has the following problems: (1) a single deterministic relationship is difficult to perceive noise and is difficult to meet the needs of diverse driving behaviors; (2) due to the reliance on statistical correlation modeling, it is impossible to distinguish between real causal relationships and spurious correlations, which makes it easy to fail to predict in real scenarios. The spurious correlation: also known as false correlation or pseudo-correlation, refers to the phenomenon that there is a statistical correlation between two variables but no actual causal relationship.

[0018] Based on this, this disclosure proposes a method for scene modeling, relation learning, and high-order pattern extraction in complex dynamic traffic environments. This method introduces uncertainty modeling and causal reasoning mechanisms to structurally represent and optimize multiple participants and their interactions in traffic scenarios. This highlights key interactions and filters out spurious correlations, thereby forming a causally enhanced dynamic graph structure. This provides support for perception, prediction, and decision-making in end-to-end autonomous driving, enabling noise perception, meeting the diverse needs of driving behavior, and improving prediction accuracy. Furthermore, by extracting and filtering high-order subgraph patterns, a general knowledge base is constructed to improve the decision-making efficiency of autonomous driving systems, providing safer and more reliable support for end-to-end autonomous driving.

[0019] The following description, in conjunction with the accompanying drawings, clearly and completely describes how the technical solution of this case is implemented. Obviously, the described embodiments are only a part of the embodiments of this case, and not all of them. Based on the embodiments in this case, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0020] See Figure 1 The flowchart shown is a method for dynamic graph relationship modeling and high-order pattern extraction based on uncertainty and causal reasoning, which includes three parts: scene graph modeling, relationship learning, and knowledge extraction and integration.

[0021] (a) Scene Graph Modeling

[0022] To address the shortcomings of existing scene graph modeling methods, which mostly output only single deterministic relationships and fail to reflect perceptual noise or accommodate the uncertainties arising from diverse driving behaviors, this disclosure proposes a probabilistic dynamic scene modeling method. This method transforms continuous data perceived by multimodal sensors into a unified structured scene representation. Building upon entity extraction, node construction, and edge construction, this method introduces additional uncertainty modeling for nodes and edges, converting continuous perceptual data into a dynamic graph structure that supports subsequent relationship learning and knowledge extraction.

[0023] This public node refers to traffic participants, including vehicles, surrounding vehicles, pedestrians, cyclists, traffic lights, stop lines, lane lines, road boundaries, crosswalks, and static obstacles.

[0024] Specifically, multimodal data collected by cameras, LiDAR, millimeter-wave radar, etc., is used as visual input to acquire all traffic participants at the current moment. In one implementation, traffic participants are treated as targets, and detection results are obtained through target detection. A unique ID is assigned to each detected target, and a Transformer-based multi-target tracking network is used to ensure that the IDs maintain trajectory continuity in the time dimension.

[0025] Next, node modeling is performed, with each traffic participant modeled as a node. The node attribute vector is represented as:

[0026]

[0027] in, The position of the node. They are velocity and acceleration, respectively. This is the heading angle (or heading angle). Categories (such as vehicles, pedestrians, traffic lights, etc.) Calculate the detection confidence score for the node. Use the node attribute value as the perceived observation result for that node.

[0028] This process forms a node set. This is used to describe all traffic participants at the current moment.

[0029] In one implementation, the attribute vector of a vehicle node is represented as {23.5, 17.8, 12.6, 0.8, 92.4, 1, 0.05}, representing that the vehicle is located at coordinates (23.5 m, 17.8 m), has a speed of 12.6 m / s, an acceleration of 0.8 m / s², a heading angle of 92.4°, is classified as a surrounding vehicle, and has an uncertainty variance of 0.05. The attribute vector of a pedestrian node is represented as follows: , where j is the pedestrian identifier. The attribute vector of the traffic light node is represented as: , k represents the traffic light symbol.

[0030] Edge modeling is performed based on the set of nodes. First, the set of relationships between nodes is defined. Nodes have positional relationships, including front, back, left, right, left-front, right-front, left-back, and right-back. A scene graph can be constructed based on these positional relationships. Each edge... This represents the interaction relationship between two nodes. This disclosure uses three dimensions—geometric, topological, and semantic—to describe edge features. Geometric relationships include nodes... and Euclidean distance and relative velocity , , They are nodes and Location, = , = , , They are nodes and Speed. Topological relationships. Used to indicate nodes and Whether they are in the same lane, adjacent lane, or conflicting lane, semantic relationship Used to indicate nodes and Whether it is subject to traffic rules (such as right-of-way, priority, red light restrictions). These features are combined into edge vectors:

[0031]

[0032] in, It is a multilayer perceptron.

[0033] This completes the construction of the scene graph. The scene graph can be represented as follows: V represents the set of nodes, and E represents the set of edges.

[0034] In one implementation, the visibility or obstruction of a node is determined based on the positional relationship between the nodes.

[0035] In one implementation, the distance relationship between nodes is determined based on the calculated Euclidean distance, including close distance and extremely close distance. The judgment ranges for close distance and extremely close distance can be preset. For example, the preset judgment range for close distance is 5-12 meters, and the judgment range for extremely close distance is less than 5 meters.

[0036] To better reflect the diversity of perceived noise and traffic behavior, this disclosure extends the edge relationship to a probability distribution form:

[0037]

[0038] Each edge satisfies independent and identically distributed characteristics. The mean... As an estimate of the relationship strength between the i-th node and the j-th node, the standard deviation The relationship between the i-th node and the j-th node is uncertain. The detection confidence of the nodes. The variance propagated to the edge, through Calculation, where It is a weighted average.

[0039] In terms of temporal dimension, to ensure the continuity and robustness of the dynamic graph, the initial scene graph constructed based on the acquired image data is smoothed. The smoothing process updates the nodes and edges in the initial scene graph through an exponential moving average mechanism.

[0040] The node update formula is:

[0041]

[0042] in, The sensory observation results at the current time t, is the momentum factor of the node.

[0043] The edge update formula is:

[0044]

[0045] This method can smooth out fluctuations between consecutive frames and reduce abrupt changes in relationships caused by noise in a single frame. The quantity factor of the edge.

[0046] The above process enables the creation of a scene graph from visual input, ultimately resulting in a probabilistic dynamic scene graph with uncertain relationships:

[0047]

[0048] Both nodes and edges possess uncertainties and maintain consistency across the temporal dimension, providing an input foundation for subsequent causal relationship learning and higher-order pattern extraction. T represents the total number of time points.

[0049] (ii) Relationship Learning

[0050] After obtaining the probabilistic dynamic scene graph, it is necessary to further learn the interaction relationships between nodes. This disclosure proposes an enhanced relationship learning method that combines a relationship graph attention mechanism with causal reasoning to highlight key interactions and filter out spurious correlations, thereby forming a causally enhanced dynamic graph structure.

[0051] Based on the probabilistic dynamic scene graph, a relational graph attention network is introduced, assigning independent attention heads to different types of relations, such as geometric, topological, semantic, and uncertain relations. For nodes... with neighboring nodes The influence relationship between them, and the formula for calculating attention weights are:

[0052]

[0053] in, Represents relation types, including geometric relations, topological relations, semantic relations, and uncertain relations. , These are node features, The characteristic distribution of the edges, , For transformation matrices of correspondence type, This is the attention parameter vector. It is an activation function used in deep learning. For nodes The set of neighboring nodes. "" indicates the vector concatenation operation.

[0054] Based on this weight, the node update formula is:

[0055]

[0056] This mechanism can adaptively highlight the most important relationship features in different scenarios.

[0057] In the attention-weighted relation graph, this disclosure further utilizes a causal intervention method to examine the importance of edges. Specifically, consider the scenario graph. Predicted output If for a certain edge (Indices i and j are the identifiers of the two nodes on this edge) The prediction result obtained after deletion is The importance of an edge is then quantified as follows:

[0058]

[0059] when Exceeding the threshold At that time, it was believed that the edge Edges that have a causal effect on downstream predictions are retained as causal valid edges. This method can effectively distinguish between true causal relationships and spurious correlations, improving the reliability of relationship modeling in out-of-distribution scenarios.

[0060] To ensure the stability of causal relationships throughout the overall learning process, this disclosure introduces a causal consistency constraint into the optimization objective. The overall loss function is defined as:

[0061]

[0062] in, For the monitoring loss of downstream forecasting tasks, This is the causal consistency loss, used to constrain the consistency between the edges retained by the model and the results of causal interventions. As a balancing factor, by introducing causal consistency constraints for screening and optimization, we can not only ensure prediction accuracy but also enhance the rationality of the causal structure, ultimately obtaining a causally effective scenario graph with a dynamic graph structure that enhances causality.

[0063] (III) Knowledge Extraction and Integration

[0064] Based on the dynamic graph structure with enhanced causality, this disclosure further proposes a method for extracting higher-order subgraph patterns (motifs) and constructing a knowledge base, which is used to preserve and reuse typical traffic interaction knowledge.

[0065] First, the causal-enhanced scene graph sequence As input, a frequent subgraph mining algorithm is used to identify recurring high-order subgraph patterns. These high-order subgraph patterns reveal recurring interaction scene patterns. Each high-order subgraph pattern is defined as a subgraph. ,in The extracted high-order subgraph patterns contain several nodes and relationships, such as the typical conflict pattern of "vehicles turning left, oncoming traffic going straight, and pedestrians crossing." Through frequent subgraph mining, a set of high-order subgraph patterns can be obtained. .

[0066] To highlight patterns that directly impact the safety of autonomous driving, this disclosure proposes a risk-driven high-order subgraph pattern selection mechanism. A risk index is applied to each high-order subgraph pattern. Risk measurement is performed, defining risk indicators as the probability of a violation occurring when they occur:

[0067]

[0068] Violations include collisions and various traffic violations, such as collisions, running red lights, crossing stop lines, driving in the wrong direction, speeding, and crossing lane lines. This information can be automatically identified and statistically analyzed through scene semantic tags.

[0069] Risk indicators This comprehensively reflects the danger and illegality of the model; when its value exceeds a set threshold... When a pattern is selected, it is marked as a critical higher-order subgraph pattern; otherwise, it is classified as a common higher-order subgraph pattern. Through this risk selection mechanism, the knowledge base not only covers frequently occurring traffic scenario patterns but also prioritizes high-risk, high-violation-rate critical patterns, thereby better serving the safety decisions of autonomous driving.

[0070] Finally, the extracted risk-driven higher-order subgraph patterns are stored as structured knowledge entries:

[0071]

[0072] Where k is the knowledge entry identifier. Represents the frequency of occurrence of higher-order subgraph patterns. Representing typical scenario tags, such as intersections, lane merging, and following other vehicles, the high-order subgraph pattern knowledge base of traffic scenarios formed in this way can support rapid matching and retrieval in new scenarios and provide general knowledge support for the safety decisions of autonomous driving.

[0073] In one implementation, a warning is issued by utilizing dangerous scenarios identified in traffic scene patterns.

[0074] In one implementation, frequent traffic scene patterns are used to perform time-series scene prediction, thereby improving the intelligence of autonomous driving.

[0075] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0076] For example, a traffic scene reuse method based on the fusion of causality and uncertainty includes: constructing a probabilistic dynamic scene graph sequence. t represents the time marker, T represents the total number of times, and V represents the set of nodes, where nodes represent traffic participants. The edges between nodes represent relationships between traffic participants. Edge features include geometric, topological, semantic, and uncertain relationships. Geometric relationships include the distance and relative speed between two nodes. Topological relationships include whether two nodes are in the same lane, adjacent lanes, or conflicting lanes. Semantic relationships include whether two nodes are bound by traffic rules. Uncertain relationships are defined as edges satisfying a probability distribution. Each probabilistic dynamic scene graph is adaptively updated based on its edge features to obtain an updated scene graph sequence. For each updated scene graph, edges are distinguished between true causal relationships and false correlations to obtain a causally valid scene graph sequence. Based on the causally valid scene graph sequence, recurring traffic scene interaction patterns are identified.

[0077] For example, a risk-driven method for constructing a traffic scenario knowledge base involves using a frequent subgraph mining algorithm to identify recurring high-order subgraph patterns in the causally valid scenario graph sequence obtained by the above method, and calculating the probability of a violation occurring when the identified high-order subgraph patterns appear.

[0078]

[0079] The subgraph corresponding to the identified higher-order subgraph pattern;

[0080] The identified higher-order subgraph patterns are stored as structured knowledge entries: ,in, Represents the frequency of occurrence of higher-order subgraph patterns. 'k' represents the scene tag, and 'k' is the knowledge entry identifier.

[0081] Those skilled in the art can implement the corresponding system based on the above method, and also know that a dedicated computer-readable storage medium can be used to store a computer program that can be executed on a processor to implement the above method.

[0082] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0083] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0084] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, Python, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of this disclosure.

[0085] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0086] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0087] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions that execute on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0088] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation in a combination of software and hardware are equivalent.

[0089] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of this disclosure is defined by the appended claims.

Claims

1. A traffic scene reuse method based on causality and uncertainty fusion, characterized in that, Comprise: Constructing probabilistic dynamic scene graph sequence , t is time stamp, T is total number of time, V is node set, node is traffic participant, , edge is relationship between traffic participants, edge feature includes geometric relationship, topological relationship, semantic relationship and uncertainty relationship, the geometric relationship includes distance and relative speed between two nodes, the topological relationship includes whether two nodes are in the same lane, adjacent lane or conflict lane, the semantic relationship includes whether two nodes are constrained by traffic rules, the uncertainty relationship is that edge satisfies independent and identically distributed, and the probability distribution of the edge is: , the mean value is taken as relationship strength estimation between the ith node and the jth node, and the standard deviation is taken as relationship uncertainty between the ith node and the jth node; Adaptive node update of each probabilistic dynamic scene graph based on edge features to obtain an updated scene graph sequence; For each updated scene graph, distinguish real causal relationships from spurious correlations, including: based on the prediction output of the updated scene graph , if the prediction result obtained after deleting a certain edge is , then the importance of the edge is quantified as: When exceeds the threshold , it is considered that the edge has a causal effect on downstream prediction, and thus is retained as a causally effective edge, otherwise the edge is deleted , and thus a causally effective scene graph sequence is obtained;​ Based on the causally effective scene graph sequence, a frequent subgraph mining algorithm is used to identify the high-order subgraph patterns that repeatedly appear in it, and a traffic scene interaction pattern that repeatedly appears is obtained through the high-order subgraph pattern.

2. The method of claim 1, wherein, Adaptive node update of each probabilistic dynamic scene graph based on edge features, comprising: Using a relational graph attention network, independent attention heads are assigned for geometric relationships, topological relationships, semantic relationships, and uncertain relationships to obtain attention weights representing different relationships between each node and its neighbor nodes; Update the node based on the attention weights between each node and its neighbor nodes.

3. The method of claim 1, wherein, The attributes possessed by the node include the position, speed, acceleration, orientation angle, category, and detection confidence of the traffic participant.

4. The method of claim 1, wherein, The construction steps of the probabilistic dynamic scene graph include smoothing the initial scene graph constructed based on the collected image data, and the smoothing updates the nodes and edges in the initial scene graph through an exponential moving average mechanism.

5. The method of claim 1, wherein, The traffic participants include the ego vehicle, surrounding vehicles, pedestrians, cyclists, traffic lights, stop lines, lane lines, road boundaries, zebra crossings, and static obstacles.

6. The method of claim 1, wherein, Using the identified high-order subgraph patterns to construct a risk-driven traffic scene knowledge base, comprising: Calculate the probability of violating rules when it appears: a subgraph corresponding to the identified high-order subgraph pattern; The identified high-order subgraph patterns are stored as structured knowledge items: ， wherein, represents the frequency of occurrence of the high-order subgraph pattern, represents a scene label, k is a knowledge item identification.

7. The method of claim 6, wherein, The scene labels include intersections, merging, and following.

Images