A future interaction modeling trajectory prediction method for complex traffic scenarios

Abstract

This invention relates to a future interaction modeling trajectory prediction method for complex traffic scenarios. The method encodes the historical motion states of each traffic participant based on the historical trajectories of the target vehicle and its surrounding participants, as well as a scene map. It then projects future patterns and unfolds the historical features over time to generate potential future features. Subsequently, based on the degree of correlation between the target vehicle and its surrounding participants in the future feature space, it constructs a future interaction association representation and selects a set of key participants with potential future interaction relationships with the target vehicle. The future features of the key participants set are then interactively fused with the future features of the target vehicle to obtain enhanced features with future interaction perception capabilities. Based on these enhanced features, the method outputs the predicted trajectory results of the target traffic participants in future time periods, thereby improving the real-time performance, prediction accuracy, and scene adaptability of trajectory prediction in complex traffic scenarios.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and autonomous driving technology, specifically relating to a future interaction modeling trajectory prediction method for complex traffic scenarios. Background Technology

[0002] With the development of autonomous driving technology, motion prediction has become one of the key technologies in autonomous driving systems. Trajectory prediction is mainly used to estimate the motion trend of traffic participants over a period of time based on their historical motion information, surrounding scene information, and road topology constraints, thereby providing forward-looking environmental information support for subsequent decision-making, planning, and risk avoidance. Most existing trajectory prediction methods are based on historical observation information to model the interaction relationships between traffic participants, typically employing a process of historical trajectory encoding, subject interaction modeling, and future trajectory decoding. While these methods can reflect the behavioral correlations between subjects to some extent, their interaction modeling mainly relies on observed historical information and lacks effective representation of potential interactions that may occur in the future. Therefore, in complex traffic scenarios, problems such as trajectory overlap, inaccurate interaction judgment, and unreasonable prediction results are prone to occur.

[0003] To address the aforementioned shortcomings, some methods further introduce conditional prediction mechanisms, utilizing the future states, future target points, or future trajectory estimates of other traffic participants to assist in the trajectory prediction of the current participant. For example, CN119068711A discloses "A system and method for predicting vehicle lane-changing trajectories in merging areas considering future spatial interactions," and CN119068711A discloses "A vehicle trajectory prediction method based on lane point future trajectory offset auxiliary supervision." Although such methods can explicitly consider future interactions, they typically rely on the accuracy of future state predictions, which can easily lead to error propagation problems. They also suffer from complex processing procedures and limited real-time performance. Therefore, there is an urgent need to provide a multi-traffic participant trajectory prediction method for complex traffic scenarios, enabling the modeling of potential future interaction relationships without relying on explicit future trajectory priors, thereby improving the accuracy, rationality, and real-time performance of trajectory prediction. Summary of the Invention

[0004] Given that existing trajectory prediction methods primarily rely on historical observation information for interaction modeling, making it difficult to effectively characterize potential future interaction relationships, and that conditional prediction methods based on explicit future states suffer from problems such as large error propagation, complex processing procedures, and insufficient real-time performance, this invention proposes a future interaction modeling trajectory prediction method for complex traffic scenarios. This method, tailored to complex traffic scenarios, constructs an integrated processing framework encompassing historical motion encoding, potential future feature generation, future interaction object identification, and trajectory prediction output. By modeling potential future interaction relationships at the intermediate feature level, it achieves collaborative prediction of the future movement trends of multiple traffic participants. This method can directly mine potential future interaction relationships from historical observation information without relying on explicit future trajectory priors, improving the accuracy and rationality of trajectory prediction for multiple traffic participants in complex traffic scenarios. Compared to existing processing methods based on historical interactions or conditional prediction, this invention effectively reduces reliance on future state prediction results, reduces error accumulation and propagation problems, and improves the real-time performance, prediction accuracy, and scenario adaptability of trajectory prediction in complex traffic scenarios by screening key interaction objects to reduce invalid interaction calculations.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for predicting future interaction trajectories in complex traffic scenarios, comprising the following steps:

[0007] Step 1. Obtain the motion state of traffic participants. Using the position and heading angle of the target vehicle in the last frame as a reference, process the position and heading angle of each traffic participant, and backfill the processed data to obtain the rotated historical state information; at the same time, search for high-precision map information.

[0008] Step 2. Process the data obtained in Step 1 to obtain the interaction features between historical trajectory and map information. ;

[0009] Step 3. Based on the data of each traffic participant in Step 2, process it using an expert network to obtain expert features; use a gating network to assign expert weights to the expert network under different future patterns to obtain the pattern features of traffic participants under future patterns, then encode them, and organize all traffic participants and all future patterns to obtain a set of potential future features.

[0010] Step 4. Project the potential future features of each traffic participant onto the target vehicle's reference coordinate system, then combine them with the target vehicle's potential future features to construct future interaction relationship connections. After normalization, organize them into a future interaction association set according to future patterns. Filter the top-ranked traffic participants, obtain the potential future features associated with them, and concatenate them with the target vehicle's potential future features. Process this with a multi-head self-attention mechanism to obtain the future interaction features under a certain future pattern. After concatenating all data, obtain the future interaction features. ;

[0011] Step 5. After concatenating the features from Step 2 and Step 4, process them through a multi-head self-attention mechanism and a multilayer perceptron to obtain the trajectory of the target vehicle.

[0012] As a preferred embodiment of the present invention, in step 1, the global position of the traffic participant at any historical moment is first translated, and then the heading angle of the target vehicle in the last frame of the historical time step is used as the reference direction of the local coordinate system. The translated coordinates are rotated to obtain the relative position of the traffic participant in the local coordinate system of the target vehicle. Then, the position information obtained after rotation and the relative heading of the traffic participant relative to the target vehicle are backfilled into the historical state information to obtain the rotated historical state information.

[0013] In a preferred embodiment of the present invention, step 2 performs high-dimensional mapping on the rotated historical state information using a multilayer perceptron, and then sequentially passes it through a bidirectional LSTM encoder and a Transformer encoder to obtain the historical information interaction features. The high-precision map information is first mapped to a high dimension using a multilayer perceptron, and then encoded using a polyline encoder to obtain polyline map features. Finally, the features of the polyline map will be displayed. Interaction features with historical information The data is stitched together, and a multi-head self-attention mechanism is used to capture the interaction between the target vehicle and the map, thus obtaining the interaction features between historical trajectory and map information. .

[0014] As a preferred embodiment of the present invention, the gating network in step 3 consists of a linear mapping layer, a ReLU activation function, and a Softmax function; the expert network consists of a multilayer perceptron, two linear mapping layers, and a ReLU activation function; and a recurrent neural network is used to analyze the pattern features of the i-th traffic participant in the m-th future pattern. Encode the potential future features of the i-th traffic participant in the m-th future pattern. .

[0015] As a preferred embodiment of the present invention, step 4 specifically includes the following steps:

[0016] Step 4.1. Project the potential future characteristics of traffic participant i onto the target vehicle using MLP. Under the reference coordinate system, the future interaction relationship characteristics of traffic participant i relative to the target vehicle are obtained;

[0017] Step 4.2. Based on the potential future features of the target vehicle in the m-th future pattern. Features of future interaction relationships between traffic participant i and the target vehicle in the m-th mode Building Future Interaction Relationships ;

[0018] Step 4.3. Normalize the future interaction relationships between all traffic participants and the target vehicle in the scenario to obtain the final future interaction relationship representation. ;

[0019] Step 4.4. Organize the set of future interaction associations between the target vehicle and all traffic participants in the m-th pattern. , Indicates the number of traffic participants in the surrounding area;

[0020] Step 4.5. Based on the future interaction association set obtained in Step 4.4, select the top 10 traffic participants to form a key set. ;

[0021] Step 4.6. Based on the set of key participants Filter out the potential future features associated with the top 10 traffic participants from the potential future feature set;

[0022] Step 4.7. After concatenating the latent future features obtained in Step 4.6 with the latent future features of the target vehicle, a multi-head self-attention mechanism is used to perform future interactions between the target vehicle and the most relevant traffic participants, resulting in the future interaction features under the m-th pattern. ;

[0023] Step 4.8. Combine the future interactions of the M patterns to obtain the future interaction features.

[0024] As a preferred embodiment of the present invention, the future modes include straight-ahead mode, parking mode, left-turn mode, right-turn mode, U-turn mode, and lane-change mode.

[0025] As a further preferred embodiment of the present invention, the multilayer perceptron consists of two linear layers and one ReLU activation function; the multi-head self-attention mechanism has eight attention heads and the number of attention layers is set to 1.

[0026] As a further preferred embodiment of the present invention, the pattern characteristics of the i-th traffic participant in step 3 in the m-th future pattern are... ;in, This represents the interaction features between the historical trajectory and map information of the i-th traffic participant. The obtained k-th expert feature, This represents the gating weight assigned to the k-th expert network by the i-th traffic participant in the m-th future pattern. This indicates the number of expert networks.

[0027] As a further preferred embodiment of the present invention, the future interaction relationship features of traffic participant i relative to the target vehicle in step 4 in the m-th mode The expression is:

[0028] ;

[0029] In the formula, This represents the relative position of traffic participant i in the local coordinate system of the target vehicle. The MLP represents the relative heading angle between traffic participant i and the target vehicle. It consists of two linear layers and a ReLU activation function. Let be the potential future features of the i-th traffic participant in the m-th future pattern.

[0030] As a further preferred embodiment of the present invention, the future interaction relationship degree in step 4 The expression is:

[0031] ;

[0032] In the formula, and These represent the query mapping matrix and the key mapping matrix, respectively. For feature dimension, Represents the transpose symbol.

[0033] Advantages and beneficial effects of the present invention:

[0034] (1) This invention addresses the problem that existing trajectory prediction methods mainly rely on historical observation information to model interaction relationships and are difficult to effectively characterize potential future interactions. It proposes a trajectory prediction method for future interaction modeling in complex traffic scenarios. This method models potential future interaction relationships at the intermediate feature level, and can explore potential future influence relationships between traffic participants without relying on explicit future trajectory priors, thereby improving the accuracy and rationality of trajectory prediction in complex traffic scenarios.

[0035] (2) This invention constructs a processing mechanism that combines historical motion feature encoding with future potential feature generation, which can map the historical behavior information of traffic participants into potential representations with future semantics. Compared with methods that rely solely on historical trajectories to directly regress future positions, this invention can pre-characterize future behavior patterns before prediction, thereby improving the model's ability to distinguish multiple potential motion trends and enhancing the expressive power and prediction accuracy of trajectory prediction in complex traffic scenarios.

[0036] (3) In the potential future feature generation stage, this invention introduces a gating and expert hybrid network. The gating network adaptively weights the outputs of multiple expert networks, enabling different traffic participants to select a more suitable future pattern generation path based on their historical behavior. This mechanism enhances the discriminative power and scene adaptability of future pattern representation, which is conducive to more accurately representing various potential future behavior patterns such as going straight, turning, changing lanes, deceleration and avoidance, and interactive game, thereby improving the performance of multimodal trajectory prediction.

[0037] (4) This invention quantifies the potential future interaction degree between the target traffic participant and surrounding traffic participants by constructing a future interaction association representation, and further filters out the key interaction objects that have the greatest impact on the future movement of the target traffic participant. Compared with neighbor selection based on current distance or using fully connected interaction modeling, this invention can more accurately focus on the most relevant traffic participants in the future, reduce interference from invalid interaction information, and thus improve the effectiveness of future interaction modeling and the credibility of prediction results.

[0038] (5) This invention integrates historical motion features, map constraint information, and future interaction features into a unified framework for joint modeling, forming an integrated prediction process from historical observation, scene understanding, future pattern generation to future interaction enhancement and trajectory decoding. Compared with existing condition-based prediction methods, this invention does not require explicit estimation of the future trajectories or target points of other traffic participants in advance, which can effectively reduce error accumulation and propagation problems, while reducing the complexity of the processing flow and improving the real-time performance and engineering applicability of the model.

[0039] (6) Under the combined effect of risk field feedback and joint optimization mechanism, this invention can not only improve the reliability of target detection, but also enhance the ability of trajectory prediction to characterize complex interactive behaviors, dangerous evolution trends and abnormal traffic actions. Especially in high-risk scenarios such as sudden lane changes, emergency deceleration, occlusion and crossing, and dense game, this invention can demonstrate stronger risk perception ability, detection robustness and trajectory prediction stability.

[0040] (7) This invention is applicable to various complex traffic scenarios such as urban roads, highways, intersections, merging and merging, and dense multi-vehicle games. It can be deployed in the prediction module of existing autonomous driving systems. It mainly relies on conventional input data such as the historical trajectories of multiple traffic participants, high-precision maps, and vehicle status information. It does not require additional high-cost perception equipment and has good engineering feasibility, system compatibility, and promotion and application value. Attached Figure Description

[0041] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0042] Figure 1 The flowchart of a future interaction modeling trajectory prediction method for complex traffic scenarios provided by the present invention is shown. Detailed Implementation

[0043] To enable those skilled in the art to better understand the technical solutions and advantages of the present invention, the present application will be described in detail below with reference to the accompanying drawings, but this is not intended to limit the scope of protection of the present invention.

[0044] like Figure 1 As shown in the figure, this embodiment provides a future interaction modeling trajectory prediction method for complex traffic scenarios. The method first encodes the historical motion state of each traffic participant based on historical trajectory information, relative positional relationships, and scene map information of the target traffic participant and its surrounding participants, obtaining historical motion features. Based on this, future pattern projection and time expansion are performed on the historical features to generate potential future features representing different future behavior patterns and future time intervals. Subsequently, based on the degree of correlation between the target traffic participant and its surrounding participants in the future feature space, a future interaction association representation is constructed, and a set of key participants with potential future interaction relationships with the target traffic participant is selected. Further, the future features of the key participant set are interactively fused with the future features of the target traffic participant to obtain enhanced features with future interaction perception capabilities. Based on these enhanced features, the predicted trajectory results of the target traffic participant in the future time period are output, thereby improving the real-time performance, prediction accuracy, and scene adaptability of trajectory prediction in complex traffic scenarios.

[0045] Specifically, the future interaction modeling trajectory prediction method for complex traffic scenarios described in this embodiment includes the following steps:

[0046] Step 1. Data Acquisition and Preprocessing:

[0047] Step 1.1. Multi-source environmental sensing data acquisition:

[0048] Based on the multi-source sensing devices on autonomous vehicles, environmental perception of the current traffic scene is performed, and the motion state information of each traffic participant in the scene is acquired over continuous historical time periods. The collected data is then used to construct historical state information. The motion state information includes at least the position coordinates, speed information, and heading angle of the traffic participants; the historical state information... N represents the number of traffic participants in the surrounding area. Indicates the step of the i-th traffic participant in historical time. The motion state information, .

[0049] Step 1.2. To better model the interaction relationships in the following steps, we will focus on the target vehicle. The last frame of the historical time step Location and heading angle For reference, the historical trajectory information of each traffic participant in the scene is processed by coordinate alignment to construct a local coordinate system centered on the target vehicle, which can be represented as:

[0050] Let the global position of neighboring traffic participant i at any historical time t be . First, perform a translation transformation:

[0051]

[0052] In the formula, This represents the translated coordinates obtained at time t, with the position of the target vehicle at the last moment of the historical observation sequence as the origin.

[0053] With the target vehicle Heading angle at the last moment of the historical observation sequence Using the local coordinate system reference direction, a rotation transformation is performed on the translated coordinates to obtain the position of traffic participant i in the target vehicle. Relative position in local coordinate system :

[0054]

[0055] In the formula, Let be a rotation matrix, and its expression is:

[0056]

[0057] Accordingly, the relative heading of the neighboring traffic participant i with respect to the target vehicle It can be represented as:

[0058]

[0059] Then, the position information obtained after rotation relative heading Backfill to To obtain the historical state information after rotation .

[0060] Step 1.3. Based on the last frame of the target vehicle in the historical time step. Location Perform matching and search for high-precision map information. The high-precision map information includes lane coordinates, lane type, speed limit, traffic light layer information, etc.

[0061] Step 2. Encoding historical motion features:

[0062] Step 2.1. Obtain the rotated historical state information obtained in Step 1.2 Using a multilayer perceptron (MLP) for high-dimensional mapping, we obtain a high-dimensional historical state information feature vector, which can be represented as:

[0063]

[0064] In the formula, As a high-dimensional historical state information feature vector, the MLP consists of two linear layers and a ReLU activation function.

[0065] Step 2.2. Convert the high-dimensional historical state information feature vector obtained in Step 2.1 into a single vector. A bidirectional LSTM encoder is used to perform time-series modeling to obtain the time-series features of historical state information. , can be represented as:

[0066]

[0067] In the formula, This indicates a bidirectional LSTM encoder, which consists of a three-layer bidirectional LSTM network.

[0068] Step 2.3. Analyze the time-series characteristics of historical state information. The system uses a standard Transformer encoder to capture interactions between traffic participants, obtaining historical interaction features. , can be represented as:

[0069]

[0070] In the formula, The Transformer encoder is mainly composed of a multi-head self-attention mechanism, a feedforward neural network, residual connections, and a layer normalization network.

[0071] Step 2.4. Use a simple MLP to perform high-dimensional feature mapping on the high-precision map information M obtained in Step 1.3 to obtain a high-dimensional map. , can be represented as:

[0072]

[0073] The MLP consists of two linear layers and a ReLU activation function.

[0074] Step 2.5. Based on the high-dimensional map obtained in Step 2.4 A polyline encoder is used to encode high-dimensional map features to obtain polyline map features. , can be represented as:

[0075]

[0076] In the formula, This represents a piecewise linear encoder.

[0077] Step 2.6. Utilize the polyline map features obtained in Step 2.5. Interaction features with historical information obtained in step 2.3 By stitching together the data, we can obtain historical trajectory and map information features. Subsequently, the multi-head self-attention mechanism (MHA) was used to capture the interaction between the target vehicle and the map, obtaining the interaction features between historical trajectory and map information. , can be represented as:

[0078]

[0079]

[0080] In the formula, This indicates a splicing operation. This indicates a multi-head self-attention mechanism, with 8 attention heads and 1 attention layer.

[0081] Step 3. Generation of potential future features:

[0082] Step 3.1. Interact with the historical trajectory and map information obtained in Step 2.6. Interaction features between the historical trajectory and map information of each traffic participant Gated networks are used to assign expert weights to each expert network. , can be represented as:

[0083]

[0084] In the formula, This represents a gated network, which consists of a linear mapping layer, a ReLU activation function, and a Softmax function. This indicates the number of expert networks, typically set to 6. This represents the gating weight assigned to the k-th expert network by the i-th traffic participant in the m-th future pattern. .

[0085] Step 3.2. Interact with the historical trajectory obtained in Step 2.6 and the map information. Interaction features between the historical trajectory and map information of each traffic participant The k-th expert feature is obtained by using an expert network. , can be represented as:

[0086]

[0087] In the formula, Let represent the k-th expert network. The expert network consists of a multilayer perceptron (MLP), two linear mapping layers, and a ReLU activation function. The MLP consists of two linear layers and a ReLU activation function.

[0088] Step 3.3. Based on the expert weights obtained in Step 3.1 and the k-th expert feature obtained in step 3.2 Obtain the pattern features of the i-th traffic participant in the m-th future pattern. , can be represented as:

[0089]

[0090] In the formula, This represents the m-th future pattern. This represents the mode characteristics of the i-th traffic participant in the m-th future mode. Future modes include straight-ahead mode, stop mode, left-turn mode, right-turn mode, U-turn mode, lane-change mode, etc.

[0091] Step 3.4. Use a GRU network to analyze the pattern features of the i-th traffic participant in the m-th future pattern. Encode the potential future features of the i-th traffic participant in the m-th future pattern. , can be represented as:

[0092]

[0093] In the formula, This represents a recurrent neural network (GRU).

[0094] Step 3.5. Organize all traffic participants and all future patterns to obtain the set of potential future features: .

[0095] Step 4. Future Interactive Object Recognition and Interaction Fusion:

[0096] Step 4.1. Project the potential future characteristics of traffic participant i onto the target vehicle using MLP. Under the reference coordinate system, the future interaction relationship characteristics of traffic participant i relative to the target vehicle (target traffic participant) are obtained. , can be represented as:

[0097]

[0098] In the formula, Indicates that traffic participant i is in the target vehicle Relative position in local coordinate system The relative heading angle between traffic participant i and the target vehicle is represented by , and the MLP consists of two linear layers and a ReLU activation function.

[0099] Step 4.2. Based on the potential future features of the target vehicle in the m-th future pattern. Features of future interaction relationships between traffic participant i and the target vehicle in the m-th mode Building Future Interaction Relationships , can be represented as:

[0100]

[0101] In the formula, and These represent the query mapping matrix and the key mapping matrix, respectively. For feature dimension, Represents the transpose symbol.

[0102] Step 4.3. Determine the future interaction relationships between all traffic participants and the target vehicle in the scenario. Normalization is performed to obtain the final representation of future interaction relationships. , can be represented as:

[0103]

[0104] In the formula, This represents the normalized future interaction association weight, used to reflect the importance of traffic participant i to the target vehicle's potential interaction in the corresponding future mode and future time interval.

[0105] Step 4.4. Organize the set of future interaction associations between the target vehicle and all traffic participants in the m-th pattern. .

[0106] Step 4.5. Based on the set of future interaction associations obtained in Step 4.4 Filter the top 10 Traffic participants form a key set .

[0107] Step 4.6. Based on the set of key participants Filter out the set of potential future features Potential future characteristics associated with the top 10 traffic participants , This represents the potential future characteristics of the traffic participants who rank first and second in the m-th mode.

[0108] Step 4.7. Associate the potential future features obtained in Step 4.6 with the top 10 traffic participants. Potential future features of the target vehicle After concatenation, a multi-head self-attention mechanism is used to analyze the future interactions between the target vehicle and the most relevant traffic participants, resulting in the future interaction features under the m-th pattern. , can be represented as:

[0109]

[0110] In the formula, This indicates a splicing operation. This indicates a multi-head self-attention mechanism, with 8 attention heads and 1 attention layer.

[0111] Step 4.8. Concatenate the future interactions of the M patterns to obtain the future interaction features. , can be represented as:

[0112]

[0113] Step 5. Future Feature Fusion and Trajectory Decoding:

[0114] Step 5.1. Interact with the historical trajectory and map information from Step 2.6. Interaction features with the future The data is concatenated and interacted with using a multi-head self-attention mechanism to ultimately decode the features. , can be represented as:

[0115]

[0116] In the formula, This indicates a splicing operation. This indicates a multi-head self-attention mechanism, with 8 attention heads and 1 attention layer.

[0117] Step 5.2. Decode the features obtained in Step 5.1 Decoding is performed using a simple MLP to obtain the target vehicle in the future. The trajectory at each time step , can be represented as:

[0118]

[0119] In the formula, Indicates the target vehicle in the future The trajectory at each time step.

[0120] The present invention also provides an electronic device, comprising: one or more processors and a memory; wherein the memory is used to store one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for predicting future interactive models in complex traffic scenarios.

[0121] The present invention also provides a computer-readable medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned method for predicting future interactive models in complex traffic scenarios.

[0122] Those skilled in the art will understand that all or part of the functions of the various methods / modules in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the above functions can be implemented by executing the program with a computer. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be implemented.

[0123] In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the programs can also be stored in storage media such as servers, other computers, disks, optical discs, flash drives, or portable hard drives. They can be downloaded or copied to the memory of the local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be implemented.

[0124] The above-described specific examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for predicting future interaction trajectories in complex traffic scenarios, characterized in that, The method includes the following steps: Step 1. Obtain the motion state of traffic participants. Using the position and heading angle of the target vehicle in the last frame as a reference, process the position and heading angle of each traffic participant, and backfill the processed data to obtain the rotated historical state information; at the same time, search for high-precision map information. Step 2. Process the data obtained in Step 1 to obtain the interaction features between historical trajectory and map information. ; Step 3. Based on the data of each traffic participant in Step 2, process it using an expert network to obtain expert features; use a gating network to assign expert weights to the expert network under different future patterns to obtain the pattern features of traffic participants under future patterns, then encode them, and organize all traffic participants and all future patterns to obtain a set of potential future features. Step 4. Project the potential future features of each traffic participant onto the target vehicle's reference coordinate system, then combine them with the target vehicle's potential future features to construct future interaction relationship connections. After normalization, organize them into a future interaction association set according to future patterns. Filter the top-ranked traffic participants, obtain the potential future features associated with them, and concatenate them with the target vehicle's potential future features. Process this with a multi-head self-attention mechanism to obtain the future interaction features under a certain future pattern. After concatenating all data, obtain the future interaction features. ; Step 4 specifically includes the following steps: Step 4.

1. Project the potential future features of traffic participant i onto the reference coordinate system of the target vehicle using MLP to obtain the future interaction relationship features of traffic participant i relative to the target vehicle; Step 4.

2. Based on the potential future features of the target vehicle in the m-th future pattern. Features of future interaction relationships between traffic participant i and the target vehicle in the m-th mode Building Future Interaction Relationships ; Future Interaction Relationship Characteristics The expression is: ； In the formula, This represents the relative position of traffic participant i in the local coordinate system of the target vehicle. The MLP represents the relative heading angle between traffic participant i and the target vehicle. It consists of two linear layers and a ReLU activation function. Let be the potential future characteristics of the i-th traffic participant in the m-th future pattern; Future Interaction Relationship Connectivity The expression is: ； In the formula, and These represent the query mapping matrix and the key mapping matrix, respectively. For feature dimension, Represents the transpose symbol; Step 4.

3. Normalize the future interaction relationships between all traffic participants and the target vehicle in the scenario to obtain the final future interaction relationship representation. ; Step 4.

4. Organize the set of future interaction associations between the target vehicle and all traffic participants in the m-th pattern. , Indicates the number of traffic participants in the surrounding area; Step 4.

5. Based on the future interaction association set obtained in Step 4.4, select the top 10 traffic participants to form a key set. ; Step 4.

6. Based on the set of key participants Filter out the potential future features associated with the top 10 traffic participants from the potential future feature set; Step 4.

7. After concatenating the latent future features obtained in Step 4.6 with the latent future features of the target vehicle, a multi-head self-attention mechanism is used to perform future interactions between the target vehicle and the most relevant traffic participants, resulting in the future interaction features under the m-th pattern. ; Step 4.

8. Concatenate the future interactions of the M patterns to obtain the future interaction features; Step 5. After concatenating the features from Step 2 and Step 4, process them through a multi-head self-attention mechanism and a multilayer perceptron to obtain the trajectory of the target vehicle.

2. The future interaction modeling trajectory prediction method for complex traffic scenarios according to claim 1, characterized in that, In step 1, the global position of the traffic participants at any historical moment is first translated. Then, the heading angle of the target vehicle in the last frame of the historical time step is used as the reference direction of the local coordinate system. The translated coordinates are rotated to obtain the relative position of the traffic participants in the local coordinate system of the target vehicle. Then, the position information obtained after rotation and the relative heading of the traffic participants relative to the target vehicle are backfilled into the historical state information to obtain the rotated historical state information.

3. The future interaction modeling trajectory prediction method for complex traffic scenarios according to claim 1, characterized in that, Step 2 uses a multilayer perceptron to perform high-dimensional mapping on the rotated historical state information, and then sequentially passes it through a bidirectional LSTM encoder and a Transformer encoder to obtain the historical information interaction features. The high-precision map information is first mapped to a high dimension using a multilayer perceptron, and then encoded using a polyline-based encoder to obtain polyline map features. Finally, the features of the polyline map will be displayed. Interaction features with historical information The data is stitched together, and a multi-head self-attention mechanism is used to capture the interaction between the target vehicle and the map, thus obtaining the interaction features between historical trajectory and map information. .

4. The future interaction modeling trajectory prediction method for complex traffic scenarios according to claim 1, characterized in that, The gating network described in step 3 consists of a linear mapping layer, a ReLU activation function, and a Softmax function; the expert network consists of a multilayer perceptron, two linear mapping layers, and a ReLU activation function; a recurrent neural network is used to analyze the pattern features of the i-th traffic participant in the m-th future pattern. Encode the potential future features of the i-th traffic participant in the m-th future pattern. .

5. The future interaction modeling trajectory prediction method for complex traffic scenarios according to claim 1, characterized in that, The future modes include straight-ahead mode, parking mode, left-turn mode, right-turn mode, U-turn mode, and lane-change mode.

6. The future interaction modeling trajectory prediction method for complex traffic scenarios according to claim 3, characterized in that, The multilayer perceptron consists of two linear layers and one ReLU activation function; the multi-head self-attention mechanism has eight attention heads and the number of attention layers is set to 1.

7. The future interaction modeling trajectory prediction method for complex traffic scenarios according to claim 4, characterized in that, The pattern characteristics of the i-th traffic participant in step 3 in the m-th future pattern ;in, This represents the interaction features between the historical trajectory and map information of the i-th traffic participant. The obtained k-th expert feature, This represents the gating weight assigned to the k-th expert network by the i-th traffic participant in the m-th future pattern. This indicates the number of expert networks.

Images