Dynamic trajectory prediction method and system based on space-time causal atlas

Abstract

The invention discloses a dynamic trajectory prediction method and system based on a space-time causal atlas, and relates to the crossing field of artificial intelligence, geographic information systems and edge computing. According to the method, a space-time causal map including physical connection, function association and causal relationship is constructed, and road network topological data and a causal rule base are fused and dynamically updated; utilizing a multi-scale space-time sensing module to extract features from microcosmic, mesoscopic and macroscopic scales respectively; by means of a differentiable causal reasoning module, trajectory prediction is decomposed into a structural causal effect, an environment regulation effect and an individual heterogeneous effect, and robustness is enhanced in combination with anti-factual reasoning; and realizing cross-city knowledge migration and distributed training based on a federated migration learning framework. The system comprises a space-time causal map construction module, a multi-scale feature extraction module, a causal reasoning prediction module, a federal transfer learning module and a result interpretation module. According to the method, the problems of precision attenuation, insufficient interpretability and poor cross-domain adaptability of a traditional method are solved, the precision, interpretability and cross-domain adaptability of trajectory prediction are improved, and the method is suitable for scenes such as intelligent traffic and automatic driving.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, geographic information systems (GIS) and edge computing, and particularly relates to a dynamic trajectory prediction method and system fusing spatio-temporal causal reasoning and knowledge graph, which is suitable for real-time trajectory prediction scenarios such as intelligent transportation, autonomous driving, emergency response, etc. BACKGROUND

[0002] In the field of intelligent transportation, autonomous driving, etc., trajectory prediction is a key technology, and its accuracy and reliability directly affect the effect of related applications. However, the existing trajectory prediction technology has many defects, which is difficult to meet the high requirements of practical application. Traditional trajectory prediction methods such as LSTM, Transformer, etc. can only capture sequence correlation, but ignore the physical constraints and causal relationships of urban road networks. For example, road construction events will cause vehicles to detour, but these methods cannot effectively model this causal relationship, resulting in a large deviation between the prediction results and the actual situation. Existing technologies are mostly black box models, which cannot provide prediction basis, which is a serious problem in safety-critical fields such as autonomous driving, cannot meet the relevant compliance requirements, and hinders the practical application of technology. Moreover, the road network structure of different cities has great differences, and the existing model needs to be retrained when migrating to a new city, which has high generalization cost and is not conducive to the rapid promotion and application of technology. Therefore, a new trajectory prediction method and system is needed to solve the problems existing in the above-mentioned prior art. SUMMARY

[0003] The purpose of the present application is to propose a trajectory prediction framework based on spatio-temporal causal graph, which solves the problems of precision decay and lack of interpretability of traditional methods in complex scenarios by explicitly modeling the causal relationship of road network and introducing a differentiable causal reasoning module, improves the precision, interpretability and cross-domain adaptability of trajectory prediction, and meets the application requirements in the field of intelligent transportation, autonomous driving, etc.

[0004] Technical scheme: In order to achieve the above application purpose, the present application provides a dynamic trajectory prediction method based on spatio-temporal causal graph, the specific steps are as follows:

[0005] Step one: build a spatio-temporal causal graph containing physical connection relationship, functional association relationship and causal relationship, the graph nodes contain geometric features , semantic features , causal features and historical statistical features ; fuse road network topology data and defined causal rule base through knowledge injection module , dynamically update the event causal chain in the graph based on Bayesian network, calculate the event propagation probability considering the node distance and influence radius .

[0006] Among them, nodes during static graph construction The edges include physical connection edges, functional association edges, and causal relationship edges. Physical connection edges are road segments with attributes such as length, speed limit, and number of lanes; functional association edges are determined based on POI semantic relationships, and their weights are determined by co-occurrence frequency; causal relationship edges are based on knowledge definitions, and their causal strength is calibrated using historical data. During knowledge injection, the road network topology is obtained through OpenStreetMap, and an initial causal rule base is constructed by combining it with a traffic domain knowledge graph. .

[0007] During dynamic graph updates, traffic states are integrated in real time, and the traffic state tensor is introduced through dynamic adjustment of edge weights. ,in , This represents the real-time average vehicle speed. To limit the rate. Simultaneously, the propagation probability of the causal chain of events is calculated based on Bayesian networks. , The distance between nodes. For the event The radius of influence.

[0008] Step 2: Through the multi-scale spatiotemporal perception module, features of micro-level individual behavior, meso-level regional interaction, and macro-level global pattern are extracted respectively. The micro-scale uses a spatiotemporal capsule network, the meso-level uses a graph attention network, and the macro-level uses a Transformer with causal masking.

[0009] The microscale spatiotemporal capsule network extracts motion features through a two-layer capsule network and uses a dynamic routing mechanism to iteratively update the coupling coefficients between capsules. The specific calculation is as follows: , , Behavioral intention reasoning is based on a policy network developed through reinforcement learning. By maximizing cumulative rewards Training, reward function , To improve the accuracy of location prediction, For the sake of time rationality, For environmental consistency.

[0010] Mesoscale graph attention networks employ attention coefficients Update node characteristics, Attention coefficient The spatiotemporal attention mechanism introduces spatiotemporal attention weights. Focus on key areas.

[0011] Macro-scale spatiotemporal causal Transformer introduces causal masks Ensure that predictions rely on historical information. Self-attention is calculated as Periodic patterns are extracted using Fourier transform. The periodic characteristics are reconstructed by preserving the main periodic components.

[0012] Step 3: Using the differentiable causal reasoning module, the trajectory prediction is decomposed into structural causal effects, environmental regulation effects, and individual heterogeneity effects. Counterfactual reasoning is then used to enhance the model's robustness. The causal effect decomposer represents the predicted trajectory as... ,in To enable learnable weights, the counterfactual inference engine generates counterfactual trajectories by intervening in the causal graph, based on a differentiable causal model. Calculate the intervention effect.

[0013] Structural causal effects are based on shortest path estimation using road network topology; environmental moderating effects consider the influence of real-time environmental factors; and individual heterogeneous effects reflect personalized user preferences. Weights are learned through causal intervention experiments and are used for counterfactual reasoning regarding interventions. The counterfactual result is , This is the subgraph obtained after removing all edges pointing to X.

[0014] Based on the federated transfer learning framework, cross-city knowledge transfer is achieved through meta-causal graphs, and distributed model training and inference are realized by combining edge computing.

[0015] Cross-city knowledge transfer extracts the common structure of causal graphs from different cities using meta-causal graphs. Meta-learning (MAML) is used to quickly adapt to new cities and optimize model parameters based on a small amount of labeled data. During edge collaborative training, the gradient is calculated locally on the edge device. The server aggregates gradients through federated learning. Weight Dynamically adjust based on data volume and model performance.

[0016] This application also provides a dynamic trajectory prediction system based on spatiotemporal causal graphs. By constructing spatiotemporal causal graphs and integrating technologies such as multi-scale spatiotemporal perception, differentiable causal reasoning, and federated transfer learning, high-precision and interpretable trajectory prediction can be achieved.

[0017] The spatiotemporal causal graph construction module is the system's basic data processing unit, responsible for constructing and dynamically updating the spatiotemporal causal graph.

[0018] The multi-scale feature extraction module, as the core unit of feature processing, receives graph data from the spatiotemporal causal graph construction module and user historical trajectory data, and extracts features from the micro, meso and macro scales respectively.

[0019] The causal reasoning prediction module, the core of the system's reasoning, receives features at various scales output by the multi-scale feature extraction module and performs differentiable causal reasoning.

[0020] The federated transfer learning module is a cross-domain adaptation and distributed training unit that supports cross-city model transfer and edge collaborative training.

[0021] The results explanation module, as the system's results output unit, receives the prediction results from the causal reasoning prediction module and generates a prediction explanation report containing the causal path.

[0022] This application has at least the following beneficial effects: By innovatively constructing a spatiotemporal causal graph, it achieves a deep integration of road network physical constraints, functional correlations, and event causal relationships, breaking through the limitations of traditional models that rely solely on the correlation of historical trajectory sequences; by leveraging multi-scale spatiotemporal perception and differentiable causal reasoning mechanisms, it can not only capture features from microscopic individual behavior to macroscopic global patterns in a hierarchical manner, but also explicitly decompose structural causality, environmental regulation, and individual heterogeneous effects in trajectory prediction, solving the problem of insufficient interpretability of black-box models; the innovatively designed federated transfer learning framework extracts cross-domain common structures through meta-causal graphs, significantly reducing the training cost and time for deployment in new scenarios, overcoming the shortcomings of poor cross-domain adaptability of traditional models, thus achieving trajectory prediction with high accuracy, strong interpretability, and high generalization ability in complex dynamic environments, providing core technical support for fields such as intelligent transportation and autonomous driving. Attached Figure Description

[0023] To further illustrate the technical solutions and advantages in the embodiments or prior art, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 is a flowchart of the steps of a dynamic trajectory prediction method based on spatiotemporal causal graph provided in an embodiment of this application;

[0025] Figure 2 is a schematic diagram of the structure of a dynamic trajectory prediction system based on spatiotemporal causal graph provided in an embodiment of this application. Detailed Implementation

[0026] To further illustrate the technical means and effects adopted by this application to achieve the intended purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, describes the specific implementation, structure, features and effects of a dynamic trajectory prediction method and system based on spatiotemporal causal graphs proposed in this application. The following embodiments are only used to more clearly illustrate the technical solution of this application and should not be used to limit the scope of protection of this application.

[0027] All attributes used herein have the meaning commonly understood by those skilled in the art, unless otherwise defined. It should be noted that attributes used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0028] Obtain the road network topology from OpenStreetMap and predict the detour trajectory of an autonomous vehicle during heavy rain. The vehicle's current location coordinates on the highway are: The time was 18:30 during the evening rush hour, the real-time environment was heavy rain (rainfall 10mm / h, visibility 150 meters), and there was a flood warning at K32+500 on the expressway. The vehicle type was a sedan, and the safety level was set to high.

[0029] Step 1: Construct a spatiotemporal causal graph encompassing the highway segment and its surrounding road network. Nodes include highway entrances, toll stations, and surrounding POIs (such as gas stations and service areas). Node features include latitude and longitude, function type, and influence radius. Edges include physical connections to the highway segment with attributes such as length and speed limit; functional association edges, such as the connection between highway exits and surrounding urban roads; and causal relationship edges, such as "heavy rain → road flooding" and "road flooding → vehicle detour." Causal strength is calibrated using historical data. The knowledge injection module integrates road network data from OpenStreetMap and a traffic domain knowledge graph to construct an initial causal rule base, such as "rainfall > 5mm / h → road friction coefficient ↓ (confidence 0.85)." During dynamic graph updates, edge weights are adjusted based on real-time traffic data, and the propagation probability of the "flooding warning" event is calculated using a Bayesian network, considering the distance to each node and the influence radius.

[0030] Step Two: At the microscale, the spatiotemporal capsule network analyzes the vehicle's historical trajectory, extracting motion features such as deceleration patterns. It updates the coupling coefficients through a dynamic routing mechanism and, combined with a reinforcement learning policy network, infers the driver's intention to drive cautiously. At the mesoscale, the graph attention network calculates the node features of the road network within a 3km radius of the current location. Attention coefficients are used to focus on key nodes such as the K35 exit, and spatiotemporal attention weights ensure the model prioritizes areas with less traffic, such as Zhongshan Road, connected to this exit. At the macroscale, the spatiotemporal causal Transformer combines historical evening rush hour data, introducing a causal mask to ensure predictions are based on historical information. Fourier transform is used to extract the periodic patterns of traffic flow during the evening rush hour, predicting that urban roads are more efficient than highways at this time.

[0031] Step 3: The causal effect decomposer decomposes the predicted trajectory into structural causal effects, environmental conditioning effects, and individual heterogeneous effects. Structural causal effects estimate traffic disruptions caused by highway flooding based on road network topology; environmental conditioning effects consider the impact of heavy rain on driving safety; and individual heterogeneous effects reflect the high safety level and risk-avoidance characteristics of vehicles. Through calculation... The predicted trajectory is obtained. The counterfactual reasoning engine conducts an intervention experiment. If we assume that we continue driving, it is calculated that there is a 78% probability of encountering a flooded section of road, with an average delay of 15 minutes. If we assume that we exit the highway at K35 and detour via Zhongshan Road → Xihu Road, the probability of delay is reduced to 12%, and the estimated additional distance traveled is 2.5km.

[0032] Step 4: Utilize meta-causal graphs to extract common knowledge about trajectory prediction in other cities during rainstorms, and rapidly adapt to the local road network characteristics through meta-learning. Edge devices locally preprocess vehicle trajectory data and extract features, calculate gradients, and upload them to the server. The server aggregates gradients and updates the model through federated learning, while employing differential privacy to protect vehicle trajectory data. Based on the federated transfer learning framework, distributed model training and inference are achieved by combining edge computing.

[0033] Based on the same inventive concept as the above method, this application also provides a dynamic trajectory prediction system based on spatiotemporal causal graphs. This system includes: a spatiotemporal causal graph construction module, a multi-scale feature extraction module, a causal inference prediction module, a federated transfer learning module, a path search module, and a result interpretation module. It should be noted that each module in the above system corresponds to a specific step of the method provided in this application embodiment, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this application embodiment.

[0034] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A dynamic trajectory prediction method based on spatiotemporal causal graphs, characterized in that, include: Step 1: Construction of a spatiotemporal causal graph, which includes physical connectivity, functional associations, and causal relationships. The graph nodes contain geometric features. semantic features Causal characteristics and historical statistical characteristics ; The knowledge injection module integrates road network topology data with a defined causal rule base. Based on the dynamic updating of the event causal chain in the graph using a Bayesian network, the event propagation probability is calculated. When considering node distance and radius of influence . Step 2: Through the multi-scale spatiotemporal perception module, features of micro-level individual behavior, meso-level regional interaction, and macro-level global pattern are extracted respectively. The micro-scale uses a spatiotemporal capsule network, the meso-level uses a graph attention network, and the macro-level uses a Transformer with causal masking. Step 3: Using the differentiable causal reasoning module, the trajectory prediction is decomposed into structural causal effects, environmental regulation effects, and individual heterogeneity effects. Counterfactual reasoning is then used to enhance the model's robustness. The causal effect decomposer represents the predicted trajectory as... ,in To enable learnable weights, the counterfactual inference engine generates counterfactual trajectories by intervening in the causal graph, based on a differentiable causal model. Calculate the intervention effect. Step 4: Based on the federated transfer learning framework, cross-city knowledge transfer is achieved through meta-causal graphs, and distributed model training and inference are realized by combining edge computing.

2. The method according to claim 1, characterized in that, In the spatiotemporal causal graph construction step, the static graph construction node... Edges include physical connection edges, functional association edges, and causal relationship edges; During knowledge injection, the road network topology is obtained through OpenStreetMap, and an initial causal rule base is constructed by combining it with a knowledge graph from the transportation domain. .

3. The method according to claim 1, characterized in that, In the multi-scale spatiotemporal perception step, the microscale spatiotemporal capsule network extracts motion features through a two-layer capsule network and iteratively updates the coupling coefficients between capsules using a dynamic routing mechanism. Behavioral intention reasoning based on a policy network using reinforcement learning Training is performed by maximizing cumulative rewards; mesoscale graph attention networks employ attention coefficients. Update node features; introduce spatiotemporal attention weights through the spatiotemporal attention mechanism. Focus on key areas; Macro-scale spatiotemporal causal Transformer introduces causal masks To ensure that predictions rely on historical information, periodic patterns are extracted through Fourier transform.

4. The method according to claim 1, characterized in that, In the federated transfer learning step, cross-city knowledge transfer extracts the common structure of causal graphs from different cities through meta-causal graphs. Meta-learning (MAML) is used to quickly adapt to new cities and optimize model parameters based on a small amount of labeled data. During edge collaborative training, edge devices compute gradients locally, and the server aggregates the gradients and weights through federated learning. Dynamically adjust based on data volume and model performance.

5. A dynamic trajectory prediction system based on spatiotemporal causal graphs, characterized in that, include: Spatiotemporal causal graph construction module: used to construct and update the spatiotemporal causal graph as described in claim 2. Multi-scale feature extraction module: Implements micro, meso and macro feature extraction as described in claim 3. Causal reasoning prediction module: performs differentiable causal reasoning as described in claim 1. Federated Transfer Learning Module: Supports cross-city model transfer and edge collaborative training. Results Interpretation Module: Generates a predictive interpretation report containing causal paths, including the contribution of structural causal effects, the contribution of environmental moderating effects, and the contribution of individual heterogeneous effects.

6. The system according to claim 5, characterized in that, The system is deployed on edge computing devices and supports the following operations: local data preprocessing and feature extraction, using spatiotemporal data dimensionality reduction algorithms to reduce cloud data transmission; real-time inference based on local causal graphs, achieving model compression through quantized perception training, with a single request response time of less than 50ms; knowledge exchange with cloud models through federated learning, using differential privacy to protect user trajectory data, and meeting GDPR compliance requirements.

Images