A Traffic Operation Safety Assessment Method for Long-Tail Scenarios Based on Knowledge-Enhanced Risk Field Fusion

Abstract

This invention relates to a method for traffic operation safety assessment in long-tail scenarios based on knowledge-enhanced risk field fusion, belonging to the field of autonomous driving decision-making and planning and traffic safety assessment. The method constructs an environment based on real trajectory maps and energy field maps, building a composite risk field through physical field modeling; then, it utilizes knowledge graph retrieval to enhance and generate diverse long-tail scenarios; finally, it incorporates uncertainty into the risk field through probabilistic derivative interactive perception, outputting probabilistic collision probabilities to achieve accurate risk assessment of long-tail scenarios. This invention improves the accuracy of traffic operation risk assessment for autonomous driving systems and traffic operation management platforms in complex traffic environments.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving decision-making and planning and traffic safety assessment, and relates to a method for assessing traffic operation safety in long-tail scenarios based on knowledge-enhanced risk field fusion. Background Technology

[0002] Traffic risk modeling is the core foundation for autonomous vehicles (AVs) to make safe decisions and avoid risks, and for traffic operation management units to achieve roadside safety control of traffic flow. As autonomous driving technology and proactive traffic flow management evolve towards all scenarios and all weather conditions, accurate, continuous, and generalizable risk assessment of the complex interactions between traffic participants becomes particularly important.

[0003] Current traffic risk modeling and assessment technologies have mainly gone through the following three stages: Early research on traditional risk metrics primarily employed discrete indicators based on time or distance. The most typical is Time-to-Collision (TTC), defined as the time required for two vehicles to maintain their current speeds until a collision occurs; variations include Modified TTC and Time Headway. Additionally, Required Deceleration and Predictive Minimum Distance are also widely used to assess conflict risk. These methods are computationally simple and have clear physical meaning, but their limitations are equally significant: they only consider pairwise interactions and cannot model multi-participant coupled risks. Furthermore, they assume constant motion states, making it difficult to adapt to dynamic driving behavior. Discrete thresholds also fail to characterize the spatially continuous distribution of risk.

[0004] Artificial potential field methods and their variants were first proposed for robot obstacle avoidance to overcome the limitations of discrete indices, treating obstacles as repulsive forces and target points as attractive forces. Kinematic potential field models consider the relative speed and direction of motion of the vehicle. However, this approach uses fixed parameters, making it difficult to adapt to different driving styles and scenarios. Furthermore, it ignores behavioral uncertainties, outputting deterministic risk values. Its modeling ability for long-tailed scenarios is limited, lacking external knowledge guidance.

[0005] Recent research in probabilistic risk modeling has begun to incorporate uncertainty into risk assessment. Gaussian process-based risk prediction methods model the uncertainty of trajectory prediction. Sources of uncertainty in risk estimation include perceptual noise, intention prediction errors, and interaction uncertainties. Probability Occupancy Grid maps future trajectory distributions to spatial occupancy probabilities. Furthermore, researchers define risk as the product of the expected collision probability and the severity of conditional collisions, constructing a probabilistic risk field. Nevertheless, existing probabilistic models largely rely on data-driven approaches and have limited generalization ability to long-tail scenarios outside the training distribution.

[0006] Meanwhile, in order to verify the safety of autonomous driving systems and enable traffic operation management units to achieve roadside safety control of traffic flow, high-risk scenario generation technology has become another research hotspot.

[0007] Early research on knowledge-based scenario generation relied on expert experience to manually design test cases. The long-range airborne surveillance system established a scenario-based testing process, dividing scenarios into three levels: functional scenarios, logical scenarios, and specific scenarios, with experts defining the parameter space. A six-layer scenario model systematically describes elements such as roads, infrastructure, and traffic participants. The advantage of this approach is its strong interpretability and ability to cover known hazard patterns; however, its drawbacks include the difficulty in enumerating all possibilities and the high cost and inefficiency of manual design.

[0008] As natural driving datasets expand, researchers are beginning to mine high-risk scenarios from real-world data. Clustering algorithms identify abnormal driving behaviors from large trajectory datasets. Deep learning-based scene retrieval methods cluster similar scenarios through contrastive learning. Generative adversarial networks (GANs) can also synthesize realistic traffic accident scenes. However, data-driven methods are inherently interpolatory, making it difficult to generate entirely new scene types outside the training distribution, and their coverage of the true long tail (i.e., rare but high-risk) is limited.

[0009] To address the coverage problem, researchers have introduced adversarial concepts and optimized search techniques. Reachability analysis combined with a formalized test scenario generation framework identifies the initial state set that could lead to collisions. Reinforcement learning trains adversarial agents to proactively induce failure cases for the tested vehicle. Importance sampling reweights the original data distribution to improve the sampling efficiency for rare events. Adaptive stress testing combined with Monte Carlo tree search can discover high-risk scenarios in a continuous state space. These methods effectively explore boundary conditions, but the search space grows exponentially with scene complexity, and the generated scenarios often lack semantic interpretability.

[0010] Scene mutation and combinatorial testing, concepts from software engineering, have been incorporated into scene generation. Scene mutation operators generate variants by inserting, deleting, and modifying base scenes. The AsFault framework utilizes genetic algorithms to search for extreme road geometries that lead to lane departures. Combinatorial coverage criteria oriented towards scene elements ensure that different combinations of elements are adequately tested. However, random mutation may generate physically unreasonable scenes and lack an understanding of risk semantics.

[0011] Meanwhile, introducing external knowledge to enhance generalization ability and interpretability has become a research hotspot in recent years.

[0012] Retrieval-Augmented Generation: Retrieval-enhanced generation methods have achieved significant success in the field of natural language processing. The RAG framework improves the accuracy and richness of generated text by retrieving external knowledge bases. Retrieval-enhanced trajectory prediction methods can retrieve similar patterns from trajectory databases to assist prediction; driving scene memory databases utilize retrieval to achieve scene understanding and decision-making. However, existing work mainly focuses on perception and prediction tasks and has not yet been systematically applied to long-tail scene generation. Knowledge-Data Fusion Risk Modeling: A few studies have attempted to integrate physical knowledge with data-driven methods. Physical Information Neural Networks (PINNs) are used for vehicle dynamics modeling, using dynamic equations as loss function constraints. Methods fusing potential field models and graph neural networks can achieve interactive perception-based risk prediction. Knowledge-guided adversarial scene generation uses traffic rules as constraints to guide the search process. These studies demonstrate that combining knowledge and data can complement each other, improving the generalization ability and physical plausibility of models.

[0013] However, traditional risk models are mostly discrete indicators or fixed parameter potential fields, making it difficult to uniformly quantify the continuous spatiotemporal distribution of multi-source coupled risks such as vehicle kinetic energy, road constraints, and dynamic interactions, and they cannot adaptively adjust the contribution weights of different risk sources. Existing long-tail scenario generation relies on random sampling or manual design, with limited coverage and a lack of semantic guidance, making it difficult to generate physically reasonable and semantically rich extreme conditions. This results in insufficient testing and verification of autonomous driving systems and traffic operation management platforms in rare high-risk scenarios. Current risk estimations are mostly deterministic outputs, ignoring the inherent uncertainty of interactive behavior, which makes decision-making systems lack confidence measurement in fuzzy scenarios, easily leading to overly optimistic or conservative decisions. At the same time, the utilization of external knowledge is relatively superficial and has not been deeply integrated into risk modeling and scenario generation, limiting the model's generalization ability and interpretability. Summary of the Invention

[0014] In view of this, the purpose of this invention is to provide a traffic operation safety assessment method for long-tail scenarios based on knowledge-enhanced risk field fusion. By integrating physical field modeling, knowledge retrieval enhancement and probabilistic interactive perception, it can achieve multi-dimensional accurate characterization, efficient generation and uncertainty quantification of risks in long-tail scenarios, thereby improving the accuracy of traffic operation risk assessment by autonomous driving systems, traffic operation management platforms and other systems in complex traffic environments.

[0015] To achieve the above objectives, the present invention provides the following technical solution: A traffic operation safety assessment method for long-tail scenarios based on knowledge-enhanced risk field fusion is proposed. The method builds an environment based on real trajectory map and energy field map data, constructs a composite risk field through physical field modeling, then uses knowledge graph retrieval to enhance the generation of diverse long-tail scenarios, and finally integrates uncertainty into the risk field through probabilistic derivative interactive perception, outputting probabilistic collision probabilities to achieve accurate risk assessment of long-tail scenarios.

[0016] Furthermore, the construction of the composite risk field specifically includes: placing any position At any moment The instantaneous risk value is defined as the weighted superposition of multiple physical fields:

[0017] in, As a comprehensive risk field, Indicates the first Physics-like fields The number of physical field categories; Learnable fusion weights (satisfying) This is used to adaptively adjust the contribution of different fields to the risk composition.

[0018] Furthermore, the physical fields specifically include the following three types of fundamental physical fields: (1) Kinetic energy field: reflects the potential danger caused by the movement of traffic participants themselves; for the first Each participant, in their position With speed During motion, at a point in space Generated kinetic energy field Defined as:

[0019] in, For equivalent quality, The scale of spatial influence; For point The angle relative to the velocity direction is used to characterize the risk anisotropy in the direction of motion; (2) Road potential energy field: describes the constraining risks formed by static elements such as lane lines, curbs, and traffic signs; deviation from the center of the lane or proximity to the curb will lead to an increase in potential energy.

[0020] in, For position At any moment The road potential energy field, The distance to the nearest lane centerline. The distance to the curb. These are preset parameters; (3) Interactive perturbation field: characterizes the dynamic risk arising from the relative motion among multiple participants, with particular attention to the potential collision risk; based on relative speed and distance, the interactive potential between the two vehicles is defined. for:

[0021] in, It is a very small positive number (smoothing constant), and its function is to prevent the denominator from having a singular value of division by zero due to the distance between the two vehicles approaching 0, thus ensuring the stability of the formula value. An indicator function that satisfies: When predicting collision time hour, The interactive perturbation field is activated; the spatial distribution of the interactive perturbation field diffuses to the surrounding region through the kernel function; when hour, If this value is 0, there is no risk of interaction.

[0022] By superimposing the fields generated by all participants and static elements point by point, a comprehensive risk field is obtained. This field provides a continuous physical basis for subsequent risk assessments and scenario generation.

[0023] Furthermore, the method of leveraging knowledge graph retrieval to enhance the generation of diverse long-tail scenarios includes the following steps: S21: Construct a multimodal large model knowledge graph, wherein the nodes of the knowledge graph represent scene elements, and the edges represent the semantic relationships between scene elements; the knowledge graph contains long-tail scene variation patterns that have been extracted and matched semantically by a large language model and verified by experts. S22: Context-aware retrieval: Obtain the initial scene context and convert it into a graph structure. Use a graph neural network to encode the graph structure to generate a query vector of fixed dimensions. Calculate the similarity between the query vector and the pattern node vector in the knowledge graph. Retrieve relevant long-tail patterns to form a candidate set, where each long-tail pattern contains a set of structured change rules for modifying the initial scene. S23: Scene Derivation Fusion: The change rules in the candidate set are applied to the initial scene context. Through a kinematic model-based differential equation solver, derived scene instances and their smooth trajectories that satisfy physical constraints and dynamic consistency are optimized and generated. S24: Adversarial sampling strategy: Calculate the risk field change of derived scene instances relative to the initial scene context through a physics-driven risk field model; construct a sampling probability distribution guided by the risk field change, and perform adversarial sampling to filter out selected scene patterns with high risk gain; S25: Diversity Enhancement and Scene Library Construction: Inject random parameter noise into selected scene patterns, and / or combine and resolve conflicts between multiple patterns to generate long-tail scenes that are both diverse and high-risk, in order to construct a long-tail scene library.

[0024] Furthermore, in step S21, the construction of the multimodal large model knowledge graph specifically includes: first, constructing a multimodal knowledge graph. , where the set of nodes Representing scene elements, edge set Triple sets represent semantic relationships between elements. Store specific facts; scene elements include: static environment, dynamic participants, environmental conditions, and event patterns; Static environment: road type (such as highway, urban road, intersection), number of lanes, curvature, slope, roadside facilities (guardrails, buildings), traffic signs and markings, etc.

[0025] Dynamic participants: type (car, truck, bicycle, pedestrian), size, initial position, speed, acceleration, and intention to move (straight, left turn, lane change, etc.).

[0026] Environmental conditions: weather (sunny, rainy, snowy, foggy), lighting (daytime, nighttime, dusk), road conditions (dry, slippery, snow-covered).

[0027] Event patterns: Dangerous events (such as "ghosting out of the way", rear-end collisions, side collisions) and their typical causes (such as obstructed vision, traffic violations).

[0028] node Having attribute feature vectors For road nodes, attributes include the number of lanes and speed limit; for participant nodes, attributes include type and speed range. Relation types are defined, such as "located" (participants are located on a certain road segment), "caused" (an environmental condition causes an event), and "accompanied" (an event often accompanies another event). Triples Represents a node and There is a relationship ; The construction process of a knowledge graph consists of two steps: (1) Automatic extraction: Extract scene elements and their associations from large-scale real databases, log records and structured traffic rule documents using natural language processing (NLP) technology (named entity recognition, relation extraction); for example, extract the description "a pedestrian suddenly crosses the road from behind a parked vehicle" from the data report, which can be transformed into triples (pedestrian, located, behind the vehicle), (pedestrian, performed, crossed), (vehicle, caused, obstructed vision) etc.

[0029] (2) Expert verification: Domain experts manually verify and complete the automatically extracted triples, remove noise, merge redundancies, and add domain knowledge (safe distance calculation formula, spatiotemporal parameter range of typical conflict scenarios) to ensure the accuracy and professionalism of the map; Furthermore, a large language model is introduced for semantic enhancement extraction to improve the accuracy and coverage of extracting scene elements and their relationships from unstructured text; the specific implementation is as follows: For each input text description The design prompt template guides the output of a structured set of triples from the large model. The large model is generated label-by-label using an autoregressive approach, and its generation probability... for:

[0030] in, Indicates the first digit in the generated sequence. One tag, For the conditional probability scores of the large model; by setting a threshold Only retain High-confidence triples are used to filter noise; Furthermore, leveraging the deep semantic representation capabilities of large models, each extracted triplet is mapped to a vector. and embedding with existing pattern nodes in the knowledge graph Perform similarity comparison; define semantic matching score. for:

[0031] like Exceeding the threshold If the triplet is semantically consistent with existing patterns, it can be assigned to the same node or its conflict resolved; otherwise, it is considered a new candidate pattern and expanded into the knowledge graph after expert verification. This mechanism effectively integrates the open-domain knowledge of large models, enabling the knowledge graph to cover a wider range of long-tail scene patterns.

[0032] Ultimately, knowledge graph It contains thousands of nodes and tens of thousands of edges, covering both common and rare long-tail scenario patterns.

[0033] Furthermore, in step S22, the context-aware retrieval specifically includes: given an initial scene context First, convert it into a graph structure. The nodes and edges correspond to respectively The elements and their relationships are identified; then, a graph neural network (GNN) is used to... Encode into a fixed-dimensional query vector Specifically, this method employs a three-layer graph convolutional network (GCN), with each layer performing the following aggregation operations:

[0034] in It is a node In the Hidden representation of layers, Represents a node The neighborhood group, and These are trainable parameters. Initial node features. (Node attribute vector). After... After the layers, this method obtains a graph-level representation through global average pooling:

[0035] in For the image The set of nodes.

[0036] Then, in the knowledge graph Search and query vectors Related long-tail patterns; due to Since the graph itself is a graph, direct graph matching computation would be too complex. Therefore, this method pre-selects each pattern node in the knowledge graph. (Representing a typical long-tail event) is also encoded into a vector using a graph neural network (GNN). The query vector is then indexed using all pattern vectors; during retrieval, the cosine similarity between the query vector and all pattern vectors is calculated. :

[0037] Select the one with the highest similarity The candidate set consists of several patterns. Each candidate pattern Includes a set of scene element change rules Each rule It describes the modification operations on a certain part of the initial scene; the rules are represented in a structured manner: The operation types include "add", "delete", and "modify", and the association constraints specify the spatiotemporal relationship with other elements.

[0038] Furthermore, in step S23, the scene derivation and fusion specifically includes: employing a differentiable scene renderer. This maps the rules to the initial states and trajectory perturbations of the participants; Define the initial scene The state of a group of participants and static environment map composition;

[0039] in, For the first The two-dimensional position coordinates of each traffic participant at the initial moment; For the first The velocity vector of each traffic participant at the initial moment; For the first The heading angle of each traffic participant at the initial moment; For the first Identify the type of traffic participant, such as motor vehicle, non-motor vehicle, pedestrian, etc. This represents the total number of traffic participants in the initial scenario.

[0040] Given a set of rules The renderer first updates the static environment, then generates new participants or modifies the state of existing participants; for new participants, the renderer automatically generates reasonable positions and velocities based on the association constraints in the rules. The renderer is essentially a differential equation solver based on a kinematic model, capable of generating smooth trajectories according to the initial state and intent. For modified participants, optimization methods are used to adjust their trajectories to satisfy physical constraints and maintain consistency with other participants. Let the modified initial state be... Renderer output future The trajectory of the steps This can be obtained by solving the following optimization problem:

[0041] The first term ensures continuity of motion, the second term smooths out speed changes, and the third term is a collision penalty term. As weight; The preset safe distance threshold represents the minimum safe distance that traffic participants should maintain to avoid collisions. This optimization problem can be solved using a differentiable programming layer, allowing the renderer to be trained end-to-end.

[0042] Furthermore, in step S24, the adversarial sampling strategy specifically includes: for each candidate pattern Calculate the change in the risk field after applying this model. Let the risk field of the original scene be... The risk field of derivative scenarios is The change in risk is defined as:

[0043] The integral measure of this formula is in the space-time domain. The total increase in internal risks; The larger the value, the more likely the pattern is to induce high-risk situations; Then, according to Define the sampling probability distribution:

[0044] in, The temperature parameter controls the randomness of the sampling. By sampling from this distribution, this method tends to select scene patterns that significantly increase risk, thereby achieving adversarial generation.

[0045] Furthermore, the probabilistic derivative interaction perception specifically includes: First, based on historical observation trajectories... Using conditional variational autoencoders (CVAEs) or diffusion models, predict the future of each participant. Multiple possible trajectories for each step, and assign probability weights to them:

[0046] in, For the number of modes, For modal probabilities, These are the trajectory distribution parameters; For the first Given a historical trajectory, the posterior probability distribution of the trajectory of a traffic participant in the next τ steps. For the first The first participant The trajectory modes are governed by a Gaussian distribution.

[0047] Then, the uncertainty in trajectory prediction is propagated to the risk field; for each mode Calculate the corresponding risk field Then the total risk field Defined as the expectation of each modal risk field:

[0048] in, It is the first The first participant The risk field contributed by each modal trajectory; In addition, the variance of the risk field is defined. As a measure of uncertainty:

[0049] This variance plot can identify areas where risk estimates are unreliable due to interaction uncertainties, providing an additional safety margin for the decision-making system; To apply the aforementioned probabilistic risk perception to closed-loop decision-making, a further interactive risk indicator is derived—the probabilistic collision probability (PCP):

[0050] in, The area occupied by the car's future trajectory. This is the risk aversion coefficient. This indicator takes into account both expected risk and uncertainty, enabling the planner to adopt a more conservative strategy in long-tail scenarios.

[0051] The beneficial effects of this invention are as follows: This invention proposes a knowledge-enhanced traffic risk field fusion method for long-tail scenarios. It constructs a composite risk field through multi-physics field weighted superposition, introducing learnable weights to adaptively adjust the contributions of different risk sources, thus more accurately characterizing traffic risks than traditional fixed-parameter potential fields. It enhances scene generation using knowledge graph retrieval, combining large-model semantic extraction and adversarial sampling to efficiently cover diverse long-tail high-risk scenarios that traditional methods struggle to discover. Innovatively, it propagates trajectory prediction uncertainty to the risk field, outputting probabilistic collision probabilities and quantifying risk assessment confidence, providing an additional safety margin for the decision-making system. This method improves the accuracy of risk perception and system robustness of autonomous driving in complex long-tail scenarios, effectively enhancing driving safety.

[0052] The knowledge-enhanced long-tail scenario traffic risk field fusion framework proposed in this invention innovatively proposes: (1) constructing a unified risk representation model driven by a physical field, incorporating multiple risk sources into the unified framework; (2) designing a knowledge retrieval-enhanced long-tail scenario generation mechanism, using a multimodal large-scale model knowledge graph to guide the semantic variation and combination of scenarios; and (3) probabilistic derivative interactive perception method, propagating trajectory prediction uncertainty to the risk field, and realizing risk assessment based on uncertainty perception. Through the above three points, the organic integration of knowledge guidance and data-driven approaches is achieved, improving the accuracy, comprehensiveness, and interpretability of risk modeling in long-tail scenarios.

[0053] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0054] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is an overall flowchart of the traffic risk assessment method for long-tail scenarios based on knowledge-enhanced risk field fusion of the present invention. Figure 2 This is a detailed flowchart of the traffic risk assessment method for long-tail scenarios based on knowledge-enhanced risk field fusion, as described in this invention. Detailed Implementation

[0055] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0056] Please see Figures 1-2 This invention provides a method for assessing traffic operation safety in long-tail scenarios based on knowledge-enhanced risk field fusion. The method builds an environment based on real trajectory maps and energy field maps, constructs a composite risk field through physical field modeling, and then uses knowledge graph retrieval to enhance the generation of diverse long-tail scenarios. Finally, it integrates uncertainty into the risk field through probabilistic interactive perception and outputs probabilistic collision probabilities to achieve accurate risk assessment of long-tail scenarios.

[0057] The method specifically includes the following steps: Step 1: Unified Representation of Physics-Driven Risk Modeling Traffic risk is essentially a product of the coupling of multiple physical fields, including the kinetic energy field generated by vehicle motion, the potential energy field of road geometric constraints, the repulsive force field of static obstacles, and the dynamic interactive disturbance field. To achieve unified quantification of risk, this method proposes a composite risk field model, which considers the risk at any location... At any moment The instantaneous risk value is defined as the weighted superposition of multiple physical fields:

[0058] in Indicates the first Physics-like fields Learnable fusion weights (satisfying) This is used to adaptively adjust the contribution of different fields to the risk composition.

[0059] Specifically, the following three types of fundamental physical fields are constructed: (1) Kinetic energy field: reflects the potential dangers caused by the movement of traffic participants themselves. For the first... Each participant, in their position With speed During motion, at a point in space The generated kinetic energy field is defined as:

[0060] in, For equivalent quality, For spatial influence scale, For point The angle relative to the velocity direction is used to characterize the risk anisotropy in the direction of motion.

[0061] (2) Road potential energy field: describes the constraining risks formed by static elements such as lane lines, curbs, and traffic signs. Deviating from the center of the lane or approaching the curb will increase the potential energy.

[0062] in, The distance to the nearest lane centerline. The distance to the curb. These are preset parameters.

[0063] (3) Interactive perturbation field: Characterizes the dynamic risk arising from the relative motion among multiple participants, with particular attention to the potential collision risk. Based on relative speed and distance, the interactive potential between the two vehicles is defined as follows:

[0064] Where TTT (Time to Threat) is the predicted collision time, when it is less than a threshold. The field is activated at that time. The spatial distribution of this field diffuses to the surrounding region through the kernel function.

[0065] By superimposing the fields generated by all participants and static elements point by point, a comprehensive risk field is obtained. This field provides a continuous physical basis for subsequent risk assessments and scenario generation.

[0066] Step 2: Generation of Long-Tail Scenarios Enhanced by Knowledge Graph Retrieval Long-tail scenarios refer to low-probability but high-risk traffic situations. Traditional methods rely on random sampling or manual design, which struggles to cover diverse and rare cases. To address this, this paper proposes a knowledge retrieval-enhanced scenario generation mechanism that utilizes external knowledge bases (including real datasets, expert experience rules, and log records) to guide scenario construction and evolution. The core idea of ​​this mechanism is to treat long-tail scenarios as structured variations of regular scenarios and explicitly model the variation patterns through knowledge graphs, thereby achieving efficient, controllable, and diverse long-tail scenario generation.

[0067] 2.1 Construction of a Multimodal Large-Scale Knowledge Graph First, construct a multimodal knowledge graph. , where the set of nodes Representing scene elements, edge set Triple sets represent semantic relationships between elements. Store specific facts. Scene elements encompass the following dimensions: Static environment: road type (such as highway, urban road, intersection), number of lanes, curvature, slope, roadside facilities (guardrails, buildings), traffic signs and markings, etc.

[0068] Dynamic participants: type (car, truck, bicycle, pedestrian), size, initial position, speed, acceleration, and intention to move (straight, left turn, lane change, etc.).

[0069] Environmental conditions: weather (sunny, rainy, snowy, foggy), lighting (daytime, nighttime, dusk), road conditions (dry, slippery, snow-covered).

[0070] Event patterns: Dangerous events (such as "ghosting out of the way", rear-end collisions, side collisions) and their typical causes (such as obstructed vision, traffic violations).

[0071] node Having attribute feature vectors For road nodes, attributes include the number of lanes and speed limit; for participant nodes, attributes include type and speed range. Relation types are defined, such as "located" (participants are located on a certain road segment), "caused" (an environmental condition causes an event), and "accompanied" (an event often accompanies another event). Triples Represents a node and There is a relationship .

[0072] The construction process of a knowledge graph consists of two steps: (1) Automatic extraction: Natural language processing (NLP) techniques (named entity recognition, relation extraction) are used to extract scene elements and their relationships from large-scale real databases, log records, and structured traffic rule documents. For example, the description "a pedestrian suddenly crosses the road from behind a parked vehicle" can be extracted from a data report and transformed into triples such as (pedestrian, located, behind the vehicle), (pedestrian, performed, crossed), and (vehicle, caused, obstructed vision).

[0073] (2) Expert verification: Domain experts manually verify and complete the automatically extracted triples, remove noise, merge redundancies, and add domain knowledge (safe distance calculation formula, spatiotemporal parameter range of typical conflict scenarios) to ensure the accuracy and professionalism of the map.

[0074] This invention further introduces a large language model for semantic enhancement extraction to improve the accuracy and coverage of extracting scene elements and their relationships from unstructured text. The specific implementation is as follows: For each input text description The design prompt template guides the output of a structured set of triples from the large model. The large model is generated label-by-label using an autoregressive approach, and its generation probability is:

[0075] in, , indicating the th in the generated sequence One tag, This represents the conditional probability distribution for a large model. A threshold is set... Only retain High-confidence triples are used to filter noise.

[0076] Furthermore, leveraging the deep semantic representation capabilities of large models, each extracted triplet is mapped to a vector. and embedding with existing pattern nodes in the knowledge graph Perform a similarity comparison. Define the semantic matching score:

[0077] like Exceeding the threshold If the triplet is semantically consistent with existing patterns, it can be assigned to the same node or its conflict resolved; otherwise, it is considered a new candidate pattern and expanded into the knowledge graph after expert verification. This mechanism effectively integrates the open-domain knowledge of large models, enabling the knowledge graph to cover a wider range of long-tail scene patterns.

[0078] Ultimately, knowledge graph It contains thousands of nodes and tens of thousands of edges, covering both common and rare long-tail scenario patterns.

[0079] 2.2 Context-Aware Retrieval Given an initial scene context This method first converts it into a graph structure. The nodes and edges correspond to respectively The elements and their relationships within the graph. Then, a graph neural network (GNN) is used to... Encode into a fixed-dimensional query vector Specifically, this method employs a three-layer graph convolutional network (GCN), with each layer performing the following aggregation operations:

[0080] in, It is a node In the Hidden representation of layers, Represents a node The neighborhood group, and These are trainable parameters. Initial node features. (Node attribute vector). After... After the layers, this method obtains a graph-level representation through global average pooling:

[0081] in, For the image The set of nodes.

[0082] Next, in the knowledge graph Search and query Related long-tail patterns. Due to Since the graph itself is a graph, direct graph matching computation would be too complex. Therefore, this method pre-selects each pattern node in the knowledge graph. (Representing a typical long-tail event) is also encoded into a vector using GNN. And construct an index from all pattern vectors. During retrieval, calculate the cosine similarity between the query vector and all pattern vectors:

[0083] Select the one with the highest similarity The candidate set consists of several patterns. Each mode Includes a set of scene element change rules Each rule This describes modifications to a portion of the initial scene. The rules are represented in a structured format: The operation types include "add", "delete", and "modify", and the association constraints specify the spatiotemporal relationship with other elements.

[0084] 2.3 Scene Derivation and Integration For each candidate mode This method needs to be concretized into an executable scenario instance. This requires the rule set Applied to the initial scene This approach ensures both the physical plausibility and dynamic consistency of the new scene. To this end, a differentiable scene renderer is introduced. It maps rules to the initial state and trajectory perturbations of the participants.

[0085] Define the initial scene The state of a group of participants

[0086] and static environment map Composition. Given a set of rules The renderer first updates the static environment, then generates new participants or modifies the state of existing participants. For new participants, the renderer automatically generates appropriate positions and velocities based on the association constraints in the rules.

[0087] The core of the renderer is a kinematic model-based differential equation solver capable of generating smooth trajectories based on the initial state and intent. For modified participants, this method utilizes optimization techniques to adjust their trajectories, ensuring they satisfy physical constraints and remain consistent with other participants. Let the modified initial state be... Renderer output future The trajectory of the steps This can be obtained by solving the following optimization problem:

[0088] The first term ensures continuity of motion, the second term smooths out speed changes, and the third term is a collision penalty term. The weights are denoted as . This optimization problem can be solved using differentiable programming layers, allowing the renderer to be trained end-to-end.

[0089] 2.4 Adversarial Sampling Strategy To focus on high-risk scenarios, this method employs an adversarial sampling strategy to select the final generated pattern. For each candidate pattern... Calculate the change in the risk field after applying this model. The risk field is driven by the physical field model defined in step 1. Given. Let the risk field of the original scene be... The risk field of derivative scenarios is The change in risk is defined as:

[0090] This integral measures the performance in the space-time domain. The total increase in internal risks. The larger the value, the more likely the pattern is to induce high-risk situations.

[0091] Then, according to Define the sampling probability distribution:

[0092] in, The temperature parameter controls the randomness of the sampling. By sampling from this distribution, this method tends to select scene patterns that significantly increase risk, thereby achieving adversarial generation.

[0093] 2.5 Diversity Enhancement and Scene Library Construction To avoid generating overly homogenous scenes, random noise injection and combined mutation operations are further introduced. For the selected mode... During the concretization process, Gaussian noise is added to its parameters:

[0094] in, The new values ​​for the scene parameters generated after adding noise; These are the baseline values ​​of the original scene parameters determined by the scene change rules; This is a random noise term that follows a Gaussian distribution, used to perturb parameters within a reasonable range and enhance scene diversity.

[0095] Furthermore, it allows multiple patterns to be combined and applied to the same initial scenario, forming a composite long-tail scenario. The combination operation is defined as the Cartesian product of the pattern sets, and conflicting rules are merged. The combined rule set is... In case of conflict, the higher risk priority shall prevail.

[0096] By repeating the sampling, generation, and mutation processes described above, a large-scale long-tail scene library can be constructed. This scenario library not only includes rare scenarios that are rarely seen in the original data distribution, but also ensures high-risk coverage through adversarial sampling, providing rich material for subsequent model training, validation, and closed-loop testing.

[0097] Step 3: Probabilistic Derivation of Interaction Perception In real-world traffic, participant behavior inherently possesses uncertainty, especially in complex interaction scenarios. Therefore, risk estimation should not be a deterministic scalar, but rather a probability distribution to reflect the degree of confidence in potential risks. This method proposes a probabilistic derivative interaction perception module, integrating the uncertainty of interaction prediction into the risk field.

[0098] First, based on historical observation trajectories Using conditional variational autoencoders (CVAEs) or diffusion models, predict the future of each participant. Multiple possible trajectories for each step, and assign probability weights to them:

[0099] in, For the number of modes, For modal probabilities, These are the trajectory distribution parameters.

[0100] Next, the uncertainty in trajectory prediction is propagated to the risk field. For each mode... Calculate the corresponding risk field The total risk field is then defined as the expectation of the risk fields of each modality:

[0101] in, It is the first The first participant The risk field contributed by each modal trajectory.

[0102] Furthermore, the variance of the risk field is defined as a measure of uncertainty:

[0103] This variance plot can identify areas where risk estimates are unreliable due to interactive uncertainties, providing an additional safety margin for the decision-making system.

[0104] To apply the aforementioned probabilistic risk perception to closed-loop decision-making, a further interactive risk indicator is derived—Probabilistic Collision Probability (PCP):

[0105] in, The area occupied by the car's future trajectory. This is the risk aversion coefficient. This indicator takes into account both expected risk and uncertainty, enabling the planner to adopt a more conservative strategy in long-tail scenarios.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for assessing traffic operation safety in long-tail scenarios based on knowledge-enhanced risk field fusion, characterized in that, An environment is built based on real trajectory maps and energy field maps. A composite risk field is constructed through physical field modeling. Then, knowledge graph retrieval is used to enhance the generation of diverse long-tail scenarios. Finally, uncertainty is integrated into the risk field through probabilistic derivative interactive perception, and probabilistic collision probabilities are output.

2. The traffic operation safety assessment method for long-tail scenarios according to claim 1, characterized in that, The construction of the composite risk field specifically includes: placing any position At any moment The instantaneous risk value is defined as the weighted superposition of multiple physical fields: in, As a comprehensive risk field, Indicates the first Physics-like fields The number of physical field categories; These are learnable fusion weights used to adaptively adjust the contribution of different fields to the risk composition.

3. The traffic operation safety assessment method for long-tail scenarios according to claim 2, characterized in that, The physical fields specifically include the following three types of fundamental physical fields: (1) Kinetic energy field: reflects the potential danger caused by the movement of traffic participants themselves; for the first Each participant, in their position With speed During motion, at a point in space Generated kinetic energy field Defined as: in, For equivalent quality, The scale of spatial influence; For point The angle relative to the velocity direction is used to characterize the risk anisotropy in the direction of motion; (2) Road potential energy field: describes the constrained risk formed by static elements; deviation from the center of the lane or proximity to the curb will lead to an increase in potential energy: in, For position At any moment The road potential energy field, The distance to the nearest lane centerline. The distance to the curb. These are preset parameters; (3) Interactive perturbation field: characterizes the dynamic risk arising from relative motion among multiple participants; based on relative speed and distance, the interactive potential between two vehicles is defined. for: in, This is a smoothing constant, its function is to prevent the denominator from having a division-by-zero singular value when the distance between the two vehicles approaches 0, and to ensure the stability of the formula's numerical value; Let the indicator function satisfy: when the collision time is predicted hour, The interactive perturbation field is activated, and its spatial distribution diffuses to the surrounding region through the kernel function; when hour, No interaction risks arise; By superimposing the fields generated by all participants and static elements point by point, a comprehensive risk field is obtained. .

4. The traffic operation safety assessment method for long-tail scenarios according to claim 2, characterized in that, The method of leveraging knowledge graph retrieval to enhance the generation of diverse long-tail scenarios includes the following steps: S21: Construct a multimodal large model knowledge graph, wherein the nodes of the knowledge graph represent scene elements, and the edges represent the semantic relationships between scene elements; the knowledge graph contains long-tail scene variation patterns that have been extracted and matched semantically by a large language model and verified by experts. S22: Context-aware retrieval: Obtain the initial scene context and convert it into a graph structure. Use a graph neural network to encode the graph structure to generate a query vector of fixed dimensions. Calculate the similarity between the query vector and the pattern node vector in the knowledge graph. Retrieve relevant long-tail patterns to form a candidate set, where each long-tail pattern contains a set of structured change rules for modifying the initial scene. S23: Scene Derivation Fusion: The change rules in the candidate set are applied to the initial scene context. Through a kinematic model-based differential equation solver, derived scene instances and their smooth trajectories that satisfy physical constraints and dynamic consistency are optimized and generated. S24: Adversarial sampling strategy: Calculate the risk field change of derived scene instances relative to the initial scene context through a physics-driven risk field model; construct a sampling probability distribution guided by the risk field change, and perform adversarial sampling to filter out selected scene patterns with high risk gain; S25: Diversity Enhancement and Scene Library Construction: Inject random parameter noise into selected scene patterns, and / or combine and resolve conflicts between multiple patterns to generate long-tail scenes that are both diverse and high-risk, in order to construct a long-tail scene library.

5. The traffic operation safety assessment method for long-tail scenarios according to claim 4, characterized in that, In step S21, the construction of the multimodal large model knowledge graph specifically includes: First, constructing a multimodal knowledge graph. , where the set of nodes Representing scene elements, edge set Triple sets represent semantic relationships between elements. Store specific facts; scene elements include: static environment, dynamic participants, environmental conditions, and event patterns; node Having attribute feature vectors ,side The relation type and triples are defined. Represents a node and There is a relationship ; The construction process of a knowledge graph consists of two steps: (1) Automatic extraction: Extract scene elements and their relationships from large-scale real databases, log records and structured traffic rule documents using natural language processing technology; (2) Expert verification: Domain experts manually verify and complete the automatically extracted triples, remove noise, merge redundancies, and add domain knowledge to ensure the accuracy and professionalism of the map; Furthermore, a large language model is introduced for semantic enhancement extraction to improve the accuracy and coverage of extracting scene elements and their relationships from unstructured text; the specific implementation is as follows: For each input text description The design prompt template guides the output of a structured set of triples from the large model. The large model is generated label-by-label using an autoregressive approach, and its generation probability... for: in, Indicates the first digit in the generated sequence One tag, For the conditional probability scores of the large model; by setting a threshold Only retain High-confidence triples are used to filter noise; Furthermore, leveraging the deep semantic representation capabilities of large models, each extracted triplet is mapped to a vector. and embedding with existing pattern nodes in the knowledge graph Perform similarity comparison; define semantic matching score. for: like Exceeding the threshold If the triplet is semantically consistent with the existing pattern, it is considered to be assigned to the same node or the conflict is resolved; otherwise, it is considered a new candidate pattern and expanded into the knowledge graph after expert verification.

6. The traffic operation safety assessment method for long-tail scenarios according to claim 4, characterized in that, In step S22, the context-aware retrieval specifically includes: given an initial scene context First, convert it into a graph structure. The nodes and edges correspond to respectively The elements and their relationships within the graph; then, a graph neural network is used to... Encode into a fixed-dimensional query vector ; Then, in the knowledge graph Search and query vectors Related long-tail patterns; each pattern node in the knowledge graph is pre-defined. Encode into vectors using a graph neural network. The query vector is then indexed using all pattern vectors; during retrieval, the cosine similarity between the query vector and all pattern vectors is calculated. : Select the one with the highest similarity The candidate set consists of several patterns. Each candidate pattern Includes a set of scene element change rules Each rule It describes the modification operations on a certain part of the initial scene; the rules are represented in a structured manner: The operation types include "add", "delete", and "modify", and the association constraints specify the spatiotemporal relationship with other elements.

7. The method for assessing traffic operation safety in long-tail scenarios according to claim 4 or 6, characterized in that, In step S23, the scene derivation and fusion specifically includes: employing a differentiable scene renderer. This maps the rules to the initial states and trajectory perturbations of the participants; Define the initial scene The state of a group of participants and static environment map composition; in, For the first The two-dimensional position coordinates of each traffic participant at the initial moment; For the first The velocity vector of each traffic participant at the initial moment; For the first The heading angle of each traffic participant at the initial moment; For the first Type identification of each traffic participant; This represents the total number of traffic participants in the initial scenario. Given a set of rules The renderer first updates the static environment, then generates new participants or modifies the state of existing participants; for new participants, the renderer automatically generates reasonable positions and velocities based on the association constraints in the rules. The renderer is essentially a differential equation solver based on a kinematic model, capable of generating smooth trajectories according to the initial state and intent. For modified participants, optimization methods are used to adjust their trajectories to satisfy physical constraints and maintain consistency with other participants. Let the modified initial state be... Renderer output future The trajectory of the steps This is obtained by solving the following optimization problem: The first term ensures continuity of motion, the second term smooths out speed changes, and the third term is a collision penalty term. As weight, The preset safe distance threshold represents the minimum safe distance that traffic participants should maintain to avoid collisions; when the spatial distance between any two vehicles at time t is less than the threshold... At this time, the collision penalty term is activated, thereby forcibly avoiding dangerous approaches during trajectory optimization.

8. The traffic operation safety assessment method for long-tail scenarios according to claim 4, characterized in that, In step S24, the adversarial sampling strategy specifically includes: for each candidate pattern Calculate the change in the risk field after applying this model. Let the risk field of the original scene be... The risk field of derivative scenarios is The change in risk is defined as: The integral measure of this formula is in the space-time domain. The total increase in internal risks; The larger the value, the more likely the pattern is to induce high-risk situations; Then, according to Define the sampling probability distribution: in, The temperature parameter controls the randomness of the sampling.

9. The traffic operation safety assessment method for long-tail scenarios according to claim 1, characterized in that, The probabilistic derived interaction perception specifically includes: First, based on historical observation trajectories Using conditional variational autoencoders or diffusion models, predict the future of each participant. Multiple possible trajectories for each step, and assign probability weights to them: in, For the number of modes, For modal probabilities, These are the trajectory distribution parameters; For the first Given a historical trajectory, the posterior probability distribution of the trajectory of a traffic participant in the next τ steps. For the first The first participant The Gaussian distribution followed by each trajectory mode; Then, the uncertainty in trajectory prediction is propagated to the risk field; for each mode Calculate the corresponding risk field Then the total risk field Defined as the expectation of each modal risk field: in, It is the first The first participant The risk field contributed by each modal trajectory; In addition, the variance of the risk field is defined. As a measure of uncertainty: Variance plots can identify areas where risk estimates are unreliable due to interaction uncertainties, providing an additional safety margin for decision-making systems; To apply probabilistic risk perception to closed-loop decision-making, a further interactive risk indicator is derived—Probabilistic Collision Probability (PCP): in, The area occupied by the car's future trajectory. This represents the risk aversion coefficient.

Images