A multi-traffic agent trajectory prediction method based on conflict perception and dynamic model switching
By introducing a multi-traffic subject trajectory prediction method that incorporates potential conflict identification and interaction intensity modeling, the problems of potential conflict identification and dynamic model adjustment in complex traffic scenarios are solved, achieving more efficient and safer trajectory prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2026-03-02
- Publication Date
- 2026-07-24
AI Technical Summary
Existing multi-traffic subject trajectory prediction methods fail to effectively identify potential conflict relationships in complex traffic scenarios and are difficult to dynamically adjust the model according to the traffic scenario, resulting in insufficient stability and security of the prediction results.
By introducing a potential conflict identification mechanism and interaction intensity modeling, a dynamically adjustable trajectory prediction strategy is constructed. The prediction model is selected or adjusted according to the interaction complexity of the traffic scenario. Potential conflicts are calculated using a local motion reference frame and two-dimensional time distance. The model library is switched in combination with the global interaction intensity index to achieve the identification and efficient prediction of key interaction subjects.
It improves the safety, stability, and adaptability of trajectory prediction in complex traffic environments, reduces computational overhead, and enhances the accuracy and reliability of prediction results.
Smart Images
Figure CN122452287A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent transportation technology, specifically to a method for predicting the trajectory of multiple traffic entities in complex traffic scenarios, applicable to applications such as autonomous driving systems, intelligent driver assistance systems, and traffic operation analysis. Background Technology
[0002] With the continuous development of intelligent transportation systems, vehicle-road cooperative technologies, and autonomous driving technologies, accurately predicting the future movement of traffic entities has become a crucial foundation for ensuring traffic safety and improving traffic efficiency. By analyzing the historical movement trajectories of traffic entities such as vehicles and pedestrians, predicting their future positions and movement trends can provide key support for path planning, behavioral decision-making, and risk warning for autonomous vehicles.
[0003] In real-world traffic scenarios, the movement of traffic subjects is often influenced by a variety of surrounding factors, especially at traffic hubs, urban road intersections, and areas with high traffic volume, where complex and dynamically changing interactions exist among multiple traffic subjects. The mutual influence between traffic subjects is not only reflected in spatial proximity but also in speed, direction of travel, and intentions, making trajectory prediction a highly complex and uncertain problem.
[0004] Existing multi-traffic subject trajectory prediction methods typically rely on historical trajectory information of traffic subjects, modeling the future trajectory of the target subject by incorporating the motion states of surrounding traffic subjects. However, in complex traffic environments, some existing technologies fail to effectively distinguish potential conflict relationships between traffic subjects, often treating all neighboring traffic subjects equally and uniformly incorporating them into the interaction modeling process. This approach, to some extent, ignores the differences in the degree of influence of different traffic subjects on the target subject, easily introducing irrelevant interaction information with low correlation to the prediction results, thereby reducing the stability and reliability of the prediction results.
[0005] On the other hand, in high-risk or highly interactive scenarios, there may be potential conflict trends among traffic entities, such as intersections, merging, or close proximity of driving trajectories. If these potential conflicts are not identified and modeled in advance during the prediction process, the prediction results may fail to reflect the true risk status, thereby affecting the safety of subsequent decision-making and control modules.
[0006] Furthermore, existing trajectory prediction methods typically employ prediction models with fixed structures and computational complexity, lacking the ability to adapt to dynamic changes in traffic scenarios. In scenarios with a small number of traffic entities and simple interactions, using highly complex models may result in unnecessary consumption of computational resources. Conversely, in scenarios with dense traffic entities, complex interactions, or numerous potential conflicts, using models with limited computational power will fail to adequately characterize the interaction features between traffic entities, thereby affecting prediction accuracy and reliability.
[0007] In summary, existing technologies for predicting the trajectories of multiple traffic entities still have the following shortcomings: First, they lack an effective mechanism for identifying and distinguishing potential conflict relationships, making it difficult to highlight key interacting entities; second, the prediction models struggle to dynamically adjust according to changes in the complexity of interactions within the traffic scenario, failing to balance prediction performance and computational efficiency; and third, the security and stability of the prediction results in complex traffic environments still need further improvement. Therefore, it is necessary to propose a new method for predicting the trajectories of multiple traffic entities to overcome these technical problems. Summary of the Invention
[0008] The technical problem to be solved by this invention is to provide a multi-traffic subject trajectory prediction method for complex traffic scenarios. This method analyzes the historical motion trajectories and interaction relationships of multiple traffic subjects in a traffic scenario, revealing the motion evolution characteristics of traffic subjects under different interaction conditions. By introducing a potential conflict relationship identification mechanism, a trajectory prediction analysis model oriented towards the interaction behavior of multiple traffic subjects is established, providing a theoretical basis for the identification and prediction of key interactive subjects in complex traffic environments. Furthermore, a trajectory prediction strategy that can be dynamically adjusted according to the interaction complexity of the traffic scenario is constructed, providing methodological support for improving the security, stability, and adaptability of multi-traffic subject trajectory prediction results.
[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0010] A method for predicting the trajectories of multiple traffic entities in complex traffic scenarios includes the following steps:
[0011] Step 1: Obtain historical motion trajectory data of multiple traffic entities within a preset time window in the traffic scenario, and determine the target traffic entity;
[0012] Step 2: Using the target traffic subject as a reference, construct a local motion reference frame, and convert the motion trajectories of other traffic subjects to the local motion reference frame for representation;
[0013] Step 3: Based on the relative motion relationship between the target traffic subject and other traffic subjects, calculate the longitudinal time distance and lateral time distance for each traffic subject, and identify traffic subjects that have potential conflicts with the target traffic subject according to preset conflict determination conditions;
[0014] Step 4: For the identified potential conflicting traffic entities, calculate the interaction strength between them and the target traffic entity, and construct a traffic entity interaction strength representation model.
[0015] Step 5: Based on the interaction intensity, adaptively select or adjust the prediction model used for trajectory prediction to adapt to the interaction complexity of the current traffic scenario.
[0016] Step 6: Based on the selected or adjusted prediction model, predict the future movement trajectory of the target traffic subject and output the prediction results in the form of a probability distribution.
[0017] As a preferred embodiment of the present invention, the historical motion trajectory data includes at least one of the position information, speed information and motion direction information of multiple traffic entities within a continuous time step.
[0018] As a preferred embodiment of the present invention, the local motion reference system takes the current position of the target traffic subject as the origin of the coordinate system, and dynamically updates the coordinate axis direction according to the real-time movement direction of the target traffic subject.
[0019] As a preferred embodiment of the present invention, the longitudinal time interval and the lateral time interval are used to characterize the potential spatiotemporal proximity of the target traffic subject to other traffic subjects in the current motion state;
[0020] When the longitudinal and lateral time intervals meet the preset threshold conditions, the corresponding traffic entities are identified as potential conflict traffic entities.
[0021] As a preferred embodiment of the present invention, the interaction intensity is calculated comprehensively based on at least two of the relative distance, relative speed, relative acceleration, or longitudinal time distance and lateral time distance between the target traffic subject and the potential conflicting traffic subject.
[0022] The interaction strength is used to characterize the interaction complexity of a traffic scenario. The interaction strength is calculated using two-dimensional time-distance (2D-TTC) as follows: ,in, Longitudinal time interval, This refers to the horizontal time interval.
[0023] As a preferred embodiment of the present invention, traffic scenarios are divided into low-interaction-complexity scenarios and high-interaction-complexity scenarios based on the interaction complexity, and different prediction models or prediction strategies are adopted for different scenarios, specifically including:
[0024] Introducing a global interaction strength index This is used to quantify the overall interaction complexity of a traffic scenario and serves as the basis for decision-making regarding the dynamic optimization and switching of the subsequent model library. Its calculation method is as follows:
[0025]
[0026] in, This refers to the number of conflicting subject pairs within the scene, where each conflicting subject pair is a combination of a key interactive subject and a target subject. The number of pairwise combinations of all subjects in the scene. The total number of subjects in the scene;
[0027] The interaction intensity range is set to [0%, 100%], and the interaction intensity range is divided into several risk level ranges, which represent the change process from low-risk scenarios to high-risk scenarios.
[0028] Global interaction strength index The lower the value, the lower the interaction complexity; conversely, the higher the value, the higher the interaction complexity. According to the metrics... The value is marked within the risk level range to determine whether the scenario is a low-interaction complexity scenario or a high-interaction complexity scenario, which is used to select the best-performing prediction model under different interaction intensities.
[0029] As a preferred embodiment of the present invention, a model library consisting of multiple trajectory prediction models is pre-constructed based on the differences in structural complexity, interactive modeling capabilities and computational overhead of different models.
[0030] Set the interaction complexity range , These represent the upper and lower limits of the threshold, and the interaction strength, respectively. Below ,Right now Defined as a low-interaction-complexity scenario; interaction intensity satisfy Defined as a scenario with medium interaction complexity; interaction intensity Higher than ,Right now This is defined as a scenario with high interaction complexity;
[0031] For scenarios with high interaction complexity, an enhanced trajectory prediction model is used; for scenarios with medium interaction complexity, a standard trajectory prediction model is used; and for scenarios with low interaction complexity, a lightweight trajectory prediction model is used.
[0032] During the prediction process, based on the interaction intensity level of the current scene, the corresponding prediction model is selected from the model library, or the model parameters are dynamically adjusted to achieve a balance between prediction accuracy and computational efficiency.
[0033] As a preferred embodiment of the present invention, the selection or adjustment of the prediction model includes adjusting at least one of the following: model size, computational accuracy, prediction time range, or input feature dimension.
[0034] As a preferred embodiment of the present invention, the prediction result is represented in the form of a probability distribution, which is used to describe the uncertainty of the future motion state of the target traffic subject; the probability distribution includes a joint description of the mean information and the degree of dispersion information of the future position of the target traffic subject.
[0035] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:
[0036] This invention effectively highlights key interacting entities that have a real impact on the target traffic entity by introducing a potential conflict identification mechanism and a traffic subject interaction intensity modeling method, thereby reducing the interference of irrelevant traffic entities on trajectory prediction results. By dynamically adjusting the trajectory prediction strategy according to the interaction complexity of the traffic scenario, unnecessary computational overhead is reduced while ensuring prediction accuracy. Through probabilistic trajectory prediction, the ability of trajectory prediction results to represent risks and uncertainties in complex traffic environments is improved, thereby enhancing the safety, stability, and adaptability of trajectory prediction methods in autonomous driving and intelligent transportation systems. Attached Figure Description
[0037] Figure 1 This is a schematic diagram of the overall process of the multi-traffic subject trajectory prediction method of the present invention.
[0038] Figure 2 This is a schematic diagram illustrating the interaction relationships and potential conflicts among traffic entities in this invention.
[0039] Figure 3 This is a schematic diagram of the dynamic prediction strategy based on interaction complexity in this invention. Detailed Implementation
[0040] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Examples of the embodiments are shown in the drawings. The embodiments described below with reference to the accompanying drawings are only used to explain the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.
[0041] like Figure 1 As shown, this invention proposes a multi-traffic subject trajectory prediction method based on conflict perception and dynamic model switching. The method includes the following steps:
[0042] Step 1: Acquisition and preprocessing of trajectory data of multiple traffic entities. First, acquire the historical movement trajectory data of traffic entities in the scene.
[0043] This study analyzes the motion characteristics of multiple traffic subjects in traffic scenarios, selects typical traffic scenarios as research objects, obtains historical motion trajectory data of traffic subjects in the scenarios, analyzes the motion state of traffic subjects under different traffic environments, and abstracts the motion relationship of multiple traffic subjects into a traffic subject interaction network.
[0044] Taking the Waymo Open Perception Dataset (WOPD) as an example, real-world road scenes from multiple cities were obtained, including traffic subjects such as vehicles (Veh), non-motorized vehicles (Bic), and pedestrians (Ped). Each scene lasts approximately 20 seconds, with a sampling frequency of 10Hz. In the official experimental dataset, 798 scenes were selected as the training set, 202 scenes as the validation set, and 150 scenes as the test set.
[0045] However, since some test set trajectories in this dataset are not publicly available, this invention uses a validation set instead of a test set for evaluation. Simultaneously, to improve the model's generalization performance and robustness, a cross-validation strategy is employed during training: a subset of samples in the training set is used as an internal validation subset for dynamically adjusting hyperparameters and early stopping criteria.
[0046] In the data preprocessing stage, this invention sets the observation time window to... (i.e., 20 frames), the prediction time window is (i.e., 30 frames), but to improve training efficiency, sampling is done at 4 frames and 6 frames.
[0047] Then, this invention constructs a dataset suitable for trajectory prediction modeling based on the original trajectory data. Specifically, a sliding window approach is used to organize the continuous time-series trajectory data into a sample structure as follows: The tensor form, in which Indicates the number of samples. Indicates the time step (including the observation segment and the prediction segment). The feature dimensions (location, type) are represented. This structure allows the model to capture both time-dependent and subject-interaction features simultaneously. The processed dataset partitioning and basic statistical information are shown in Table 1. The number of samples, trajectories, and category proportions in the validation and test sets are approximately distributed similarly to those in the training set, ensuring the consistency of data distribution.
[0048] Table 1
[0049]
[0050] Step 2: Based on the historical motion trajectory and relative motion relationship of the traffic subject, construct a local motion reference frame centered on the target traffic subject, analyze the relative position, relative speed and motion direction relationship between the target traffic subject and the surrounding traffic subjects, and provide a unified reference framework for subsequent interactive analysis.
[0051] In this embodiment, a local motion reference frame is constructed, that is, a local coordinate system is established with the initial position of the target body A as the origin and the initial driving direction as the x-axis. This eliminates the interference of absolute position differences on the interaction judgment and makes the calculation of the interaction relationship more focused on the relative motion characteristics.
[0052] Let the initial state of target A in the global coordinate system be ( ); The initial state of the surrounding main body B is ( ).in, Let A be the horizontal and vertical coordinates of subject A in the global coordinate system at time 0. Let A be the orientation angle of subject A in the global coordinate system at time 0. Let B be the horizontal and vertical coordinates of the subject B in the global coordinate system at time 0. Let be the orientation angle of subject B in the global coordinate system at time 0. The global coordinate system is transformed to the local coordinate system through rotation and translation transformations, as shown in the following formulas:
[0053]
[0054]
[0055] in,( )and( ) are the position coordinates of moving entities A and B at time t, in the local coordinate system of the initial position core of A. )and( Let A and B be the position coordinates of the moving entities A and B at time t in the initial global coordinate system. As defined above, is the orientation angle of A in the global coordinates at time 0.
[0056] Step 3: Analyze the influencing factors of potential conflict relationships between traffic entities, establish a potential conflict identification model based on the relative motion state between traffic entities, and screen and identify potential conflict traffic entities that have a significant impact on the target traffic entity.
[0057] In this application, "significant impact" refers to the degree of interaction that substantially constrains the motion decision-making and trajectory prediction of the target traffic subject. Two-dimensional longitudinal time interval (2D-TTC) is used as the quantitative indicator for determining "significant impact." When the 2D-TTC value between the target traffic subject and a surrounding traffic subject is less than a preset threshold, the surrounding subject is deemed to have a "significant impact" on the target traffic subject. This threshold can be determined using empirical methods and experimental calibration methods.
[0058] 1. Empirical method: Use a commonly used empirical threshold, such as 3 seconds;
[0059] 2. Experimental calibration: Experiments were conducted on the dataset to statistically analyze indicators such as collision recognition accuracy and trajectory prediction accuracy under different thresholds, and the optimal threshold was calibrated through methods such as cross-validation.
[0060] Based on the identification of potential conflict subjects in two-dimensional time distance, the interaction relationship and potential conflict between traffic subjects need to be combined with the quantitative determination of relative motion state. Typical scenarios include... Figure 2 As shown in the figure, the light blue vehicle is the target traffic subject, the green vehicle is the surrounding traffic subject, and the dashed line is the movement trajectory of the two. Trajectory 1 and trajectory 2 are both possible movement paths of the surrounding subjects. Trajectory 1 has a closer intersection distance with the target subject's movement trajectory and a smaller relative speed difference, which is a relatively strong interaction relationship and has a higher potential conflict risk. Trajectory 2 has a lower spatial proximity to the target subject and a weaker interaction intensity.
[0061] Based on this interaction scenario, this step uses two-dimensional time distance (2D-TTC) calculation to filter out subjects with strong interaction and high conflict similar to trajectory 1.
[0062] First, based on the constructed local motion reference frame, the interaction relationship between two objects is determined by calculating the two-dimensional time distance. Specifically, the longitudinal, lateral and average time distances between the target subject A and the surrounding subject B are calculated to quantify the potential conflict risk. The calculation formula and parameter definition are shown in Table 2.
[0063] Table 2
[0064]
[0065] in Let A be the length of the target body A. Let A be the width. and The distribution represents the velocities of subjects A and B along the x-axis in the current coordinate system. and The distribution represents the y-axis velocity of subjects A and B in the current coordinate system. This indicator takes into account both the probability of collision in the driving direction and the lateral direction, avoiding misjudgment of risk caused by a single-dimensional assessment.
[0066] In this embodiment, the threshold is calibrated through preliminary experiments, and the average time interval is... The threshold is set to 3 seconds (different values need to be used for different test scenarios). When target A and surrounding subject B satisfy the condition for at least one frame within the observation window... If the interaction is successful, the surrounding entity is identified as a key interactive entity; otherwise, it is considered an irrelevant entity and filtered out during subsequent prediction modeling.
[0067] Step 4: Global interaction intensity calculation. To comprehensively characterize the interaction complexity of the traffic scenario, a global interaction intensity index is introduced. This is used to quantify the conflict complexity of a traffic scenario and serves as the basis for decision-making in the dynamic optimization and switching of the subsequent model library. Its calculation method is as follows:
[0068] ;
[0069] in, The number of conflicting subject pairs within the scene (the number of combinations of key interactive subjects and target subjects). The number of pairwise combinations of all subjects in the scene ( (Total number of subjects in the scene).
[0070] Based on the identified potential conflicting traffic entities, the interaction characteristics between traffic entities are analyzed, and a traffic entity interaction intensity representation model is constructed (the input of this model is all trajectories in this scenario and the conflict determination threshold; the output is the interaction complexity I of the scenario), which is used to quantify the degree of interaction influence between traffic entities and reflect the interaction complexity of the current traffic scenario.
[0071] To select the best-performing prediction model under different interaction intensities, this invention divides the interaction intensity range [0%, 100%] into 10 risk level intervals (each interval being 10%), representing the change process from low-risk scenarios to extremely high-risk scenarios.
[0072] Step 5: Dynamic prediction model selection and switching based on interaction intensity. Based on the traffic scene interaction intensity level calculated in Step 4, adaptively select or adjust the prediction model used for trajectory prediction to adapt to traffic scenes with different interaction complexities.
[0073] In this invention, the selection of the trajectory prediction model can be achieved through the following two technical approaches, both of which are based on dynamic decision-making based on the scene interaction complexity (I) characterized by interaction intensity (2D-TTC):
[0074] 1. Hierarchical Adaptation: Predefine threshold ranges for different complexity levels, compare the quantified value of interaction intensity with the threshold, and thus adapt the corresponding prediction model according to the hierarchy.
[0075] For example, for highly complex scenarios (interaction intensity is higher than a high threshold, such as I>=60%), an enhanced trajectory prediction model is adopted, such as a multi-layer deep neural network or a complex network structure that incorporates attention mechanisms, in order to fully capture the strong interaction constraints between multiple subjects.
[0076] For medium-complexity scenarios (interaction intensity between high and low thresholds, such as 30%<=I<60%), standard trajectory prediction models are used, such as a hybrid architecture that combines LSTM-CNN temporal modeling and GNN interaction modeling, which ensures prediction accuracy while taking into account computational efficiency.
[0077] For low-complexity scenarios (interaction intensity below a low threshold, such as I < 30%), a lightweight trajectory prediction model, such as a simplified LSTM-CNN-Transformer temporal modeling network, is used to achieve fast inference.
[0078] 2. Model Library Optimization: A trajectory prediction model library with various structures and complexities is pre-built and offline tested on public or self-built datasets to evaluate the prediction accuracy, inference latency, and other performance metrics of each model under different interaction complexities, forming an "interaction complexity - model performance" mapping table. During the actual prediction phase, based on the interaction intensity value of the current scene, this mapping table is consulted to select the best-performing model for trajectory prediction.
[0079] The core logic of this step can be achieved through... Figure 3 The key to this intuitive demonstration lies in the pre-built model library consisting of various trajectory prediction models (LSTM, Social-LSTM, Social-GAN, Traffic-Predict, GATraj, etc.). These different models vary in structural complexity, interaction modeling capabilities, and computational overhead. For example, in low-interaction-complexity scenarios, a lightweight prediction model with a simpler structure can be used to improve computational efficiency; in medium-complexity scenarios, commonly used standard models can be used for prediction; and in high-interaction-complexity scenarios, an enhanced prediction model capable of fully modeling the interactions between multiple entities is employed to improve prediction accuracy.
[0080] In the actual prediction process, based on the interaction intensity level of the current scene, the corresponding prediction model is selected from the model library, or the model parameters are dynamically adjusted, thereby achieving a balance between prediction accuracy and computational efficiency.
[0081] Taking model library optimization as an example, a trajectory prediction model library containing various structures and complexities (such as Social-LSTM, Social-GAN, Traffic-Predict, and GATraj) is pre-built. Offline testing is performed on public or self-built datasets to evaluate the prediction accuracy, inference latency, and other performance metrics of each model under different interaction complexity scenarios, forming an "interaction complexity - model performance" mapping table. During the actual prediction phase, based on the interaction intensity value of the current scenario, this mapping table is consulted to select the best-performing model for trajectory prediction.
[0082] Step 6, Future Trajectory Prediction: Based on the prediction model selected or adjusted in Step 5, the future trajectory of the target traffic vehicle is predicted. The prediction results can be output in the form of a probability distribution to characterize the uncertainty of the future trajectory.
[0083] The probability distribution may include the mean, variance, and correlation information in different directions of the predicted location, thereby providing a reliable basis for subsequent path planning, risk assessment, or decision support modules.
[0084] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solutions based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.
Claims
1. A multi-traffic subject trajectory prediction method based on conflict perception and dynamic prediction, characterized in that, Includes the following steps: Step 1: Obtain historical motion trajectory data of multiple traffic entities within a preset time window in the traffic scenario, and determine the target traffic entity; Step 2: Using the target traffic subject as a reference, construct a local motion reference frame, and convert the motion trajectories of other traffic subjects to the local motion reference frame for representation; Step 3: Based on the relative motion relationship between the target traffic subject and other traffic subjects, calculate the longitudinal time distance and lateral time distance for each traffic subject, and identify traffic subjects that have potential conflicts with the target traffic subject according to preset conflict determination conditions; Step 4: For the identified potential conflicting traffic entities, calculate the interaction strength between them and the target traffic entity, and construct a traffic entity interaction strength representation model. Step 5: Based on the interaction intensity, adaptively select or adjust the prediction model used for trajectory prediction to adapt to the interaction complexity of the current traffic scenario. Step 6: Based on the selected or adjusted prediction model, predict the future movement trajectory of the target traffic subject and output the prediction results in the form of a probability distribution.
2. The multi-traffic entity trajectory prediction method according to claim 1, characterized in that, The historical motion trajectory data includes at least one of the following: position information, speed information, and direction of motion information of multiple traffic entities within a continuous time step.
3. The multi-traffic entity trajectory prediction method according to claim 1, characterized in that, The local motion reference system uses the current position of the target traffic subject as the origin of the coordinate system, and dynamically updates the coordinate axis direction according to the real-time movement direction of the target traffic subject.
4. The multi-traffic entity trajectory prediction method according to claim 1, characterized in that, The longitudinal and lateral time intervals are used to characterize the potential spatiotemporal proximity of the target traffic subject to other traffic subjects in their current motion state. When the longitudinal and lateral time intervals meet the preset threshold conditions, the corresponding traffic entities are identified as potential conflict traffic entities.
5. The multi-traffic entity trajectory prediction method according to claim 1, characterized in that, The interaction intensity is calculated based on at least two of the following: the relative distance, relative speed, relative acceleration, or the longitudinal time distance and the lateral time distance between the target traffic subject and the potential conflicting traffic subject. The interaction strength is used to characterize the interaction complexity of a traffic scenario. The interaction strength is calculated using two-dimensional time-distance (2D-TTC) as follows: ,in, Longitudinal time interval, This refers to the horizontal time interval.
6. The multi-traffic entity trajectory prediction method according to claim 5, characterized in that, Based on the interaction complexity, traffic scenarios are divided into low-interaction-complexity scenarios and high-interaction-complexity scenarios, and different prediction models or prediction strategies are adopted for different scenarios, specifically including: Introducing a global interaction strength index This is used to quantify the overall interaction complexity of a traffic scenario and serves as the basis for decision-making regarding the dynamic optimization and switching of the subsequent model library. Its calculation method is as follows: ， in, This refers to the number of conflicting subject pairs within the scene, where each conflicting subject pair is a combination of a key interactive subject and a target subject. The number of pairwise combinations of all subjects in the scene. The total number of subjects in the scene; The interaction intensity range is set to [0%, 100%], and the interaction intensity range is divided into several risk level ranges, which represent the change process from low-risk scenarios to high-risk scenarios. Global interaction strength index The lower the value, the lower the interaction complexity; conversely, the higher the value, the higher the interaction complexity. According to the metrics... The value is marked within the risk level range to determine whether the scenario is a low-interaction complexity scenario or a high-interaction complexity scenario, which is used to select the best-performing prediction model under different interaction intensities.
7. The multi-traffic entity trajectory prediction method according to claim 1, characterized in that, The trajectory prediction model is selected in two ways, both of which are based on dynamic decision-making on the scene interaction complexity I represented by the interaction intensity: (1) Graded adaptation: Predefine threshold ranges for different complexity levels, compare the quantified value of interaction intensity with the threshold, and adapt the corresponding prediction model accordingly. (2) Model library selection: A trajectory prediction model library containing various structures and complexities is pre-built and offline testing is completed on public or self-built datasets; Evaluate the performance metrics of each model under different interaction complexities and form an interaction complexity-model performance mapping table; During the prediction phase, based on the interaction intensity value of the current scene, the mapping table is queried, and the model with the best performance is selected for trajectory prediction.
8. The multi-traffic entity trajectory prediction method according to claim 1 or 7, characterized in that, Based on the differences in structural complexity, interactive modeling capabilities, and computational overhead among different models, a model library consisting of multiple trajectory prediction models is pre-built; Set the interaction complexity range , These represent the upper and lower limits of the threshold, and the interaction strength, respectively. Below ,Right now Defined as a low-interaction-complexity scenario; interaction intensity satisfy Defined as a scenario with medium interaction complexity; interaction intensity Higher than ,Right now This is defined as a scenario with high interaction complexity; For scenarios with high interaction complexity, an enhanced trajectory prediction model is used; for scenarios with medium interaction complexity, a standard trajectory prediction model is used; and for scenarios with low interaction complexity, a lightweight trajectory prediction model is used. During the prediction process, based on the interaction intensity level of the current scene, the corresponding prediction model is selected from the model library, or the model parameters are dynamically adjusted to achieve a balance between prediction accuracy and computational efficiency.
9. The multi-traffic entity trajectory prediction method according to claim 6, characterized in that, The selection or adjustment of the prediction model includes adjusting at least one of the following: model size, computational accuracy, prediction time range, or input feature dimension.
10. The multi-traffic entity trajectory prediction method according to claim 1, characterized in that, The prediction results are expressed in the form of a probability distribution, which is used to describe the uncertainty of the future motion state of the target traffic subject; The probability distribution includes a joint description of the mean and dispersion information of the future location of the target traffic subject.