A Multi-Scale Interactive Sensing Method for Highway Vehicle Trajectory Prediction

Abstract

This invention belongs to the field of vehicle trajectory prediction technology and discloses a multi-scale interactive perception method for predicting vehicle trajectories on highways. It jointly models short-term sudden features and long-term evolution trends using convolutional neural networks and bidirectional gated recurrent units, and introduces a temporal attention mechanism to enhance the perception of key time segments. Combined with a dynamic graph attention mechanism incorporating physical edge features such as relative position, relative speed, and relative acceleration, it updates vehicle interaction relationships in real time to improve spatial modeling accuracy and interpretability. In the decoding stage, it integrates guidance vectors and lane semantic information, ensuring that the predicted trajectory conforms to the road geometry spatially and remains smooth and continuous temporally. This method can improve the robustness and adaptability of the model in sparse neighboring vehicle environments on highways while maintaining prediction accuracy, providing more stable and reliable trajectory prediction results for intelligent transportation systems and offering strong technical support for highway traffic safety management and operation scheduling.

Description

Technical Field

[0001] This invention belongs to the field of vehicle trajectory prediction technology, specifically relating to a multi-scale interactive perception method for predicting highway vehicle trajectories. Background Technology

[0002] With the continuous development of Intelligent Transportation Systems (ITS), vehicle trajectory prediction technology plays an increasingly important role in improving road traffic efficiency, ensuring traffic safety, and supporting autonomous driving decisions. Especially in highway scenarios, where vehicles travel at high speeds and the number of lanes is limited, traffic accidents often trigger chain reactions, causing severe traffic congestion, vehicle collisions, and other incidents, and making accident rescue extremely difficult. Therefore, vehicle trajectory prediction technology is crucial for improving highway driving safety, reducing accident risks, and ensuring stable traffic flow. The goal of vehicle trajectory prediction is to predict the future position of a target vehicle based on its historical trajectory information and that of neighboring vehicles. Accurate vehicle trajectory prediction is essential for applications such as traffic management and accident prevention.

[0003] However, vehicle behavior in real-world highway traffic scenarios is highly dynamic and uncertain. Trajectory evolution is influenced not only by the target vehicle's own historical state but also by the interference of neighboring vehicles. Therefore, constructing a high-accuracy trajectory prediction model requires a comprehensive consideration of vehicle temporal variation characteristics and dynamic spatial interactions. Many studies have attempted to model vehicle trajectories using deep learning models such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Graph Neural Networks (GNNs), achieving some success. However, existing methods still face many challenges in the highly dynamic highway traffic environment, requiring further solutions.

[0004] First, existing methods have limitations in multi-scale temporal dependency modeling. Accurate trajectory prediction requires modeling both short-term temporal features (e.g., sudden braking) and long-term temporal dependencies (e.g., lane changes). Furthermore, in real-world trajectories, not all historical frames have the same reference value for future behavior; the truly decisive "critical moments" in trajectory development often concentrate on a small number of frames, such as slight lateral drift before a lane change. If the model fails to focus on these critical time segments, it will miss important decision signals, leading to decreased prediction accuracy. However, many existing models cannot simultaneously capture both short-term and long-term temporal dependencies and frequently ignore critical time segments in the trajectory.

[0005] Secondly, existing methods are insufficient in modeling spatial dynamic interactions. Most methods use static graphs or simple adjacency updates, which are difficult to adapt to the continuous changes in the states of adjacent vehicles and cannot effectively model the dynamic nonlinear characteristics of spatial interactions in complex scenarios.

[0006] Finally, the interactive modeling process lacks interpretability. Some studies use attention mechanisms or graph neural networks to capture the degree of influence of neighboring vehicles on the target vehicle, ignoring physical side characteristics such as relative speed and relative acceleration between vehicles, making it difficult to clearly explain why or how neighboring vehicles affect the predicted trajectory of the target vehicle. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a multi-scale interactive perception method for predicting highway vehicle trajectories. This invention aims to solve the problems of insufficient multi-scale temporal dependency modeling, insufficient spatial dynamic interaction modeling, and insufficient interpretability of existing methods.

[0008] This invention provides a multi-scale interactive sensing method for predicting highway vehicle trajectories, comprising the following steps:

[0009] S1. Collect the historical movement trajectory of the target vehicle on the highway and the interaction information of nearby vehicles;

[0010] S2. Construct the MIST model;

[0011] The MIST model includes an input feature embedding module, a multi-scale temporal coding module, a dynamic spatial interaction modeling module, and a trajectory decoding and prediction module.

[0012] Input feature embedding module: maps the trajectory state vectors of the target vehicle and neighboring vehicles to a high-dimensional embedding space, and fuses temporal and positional encoding with lane semantics;

[0013] Multi-scale temporal coding module: used to simultaneously capture short-term variation features and long-term dependency features in the trajectories of the target vehicle and neighboring vehicles, and enhance the model's ability to perceive key moments;

[0014] Dynamic spatial interaction modeling module: used to depict the dynamic spatial interaction relationship between the target vehicle and neighboring vehicles;

[0015] Trajectory decoding and prediction module: used to generate the future trajectory of the target vehicle;

[0016] S3. Input the data collected in step S1 into the MIST model constructed in step S2, and finally output the future trajectory of the target vehicle.

[0017] Furthermore, the design of the input feature embedding module in step S2 is as follows:

[0018] The input feature embedding module uses the original state vector of each vehicle at time t. For input, The six dimensions represent the position coordinates (x, y), velocity components (v), and velocity components, respectively. x , v y ) and acceleration components (a x , a y );

[0019] First, through the linear mapping matrix Project the original state vector onto The initial feature embedding is obtained by using a dimensional embedding space, stabilizing the feature distribution using layer normalization, and then enhancing the nonlinear expressive power through the LeakyReLU activation function. ;

[0020] exist Superimposed learnable temporal location codes , From the learnable matrix Provided, among which Maximum number of historical frames;

[0021] In addition, to incorporate road geometric semantic information, the lane number lᵢ of the target vehicle is embedded through a lane embedding matrix. Convert to lane embedding vector and with Add them together to obtain the final single-frame embedding vector for each frame. Given a historical time window length of T, stacking the embedding vectors along the time dimension yields the target vehicle embedding feature sequence. Where ego represents the target vehicle currently being predicted; for each neighboring vehicle Similarly, the neighboring vehicle embedding feature sequence is obtained. ,in This represents the k-th neighboring vehicle within the perceived range of the historical time window.

[0022] Furthermore, the design of the multi-scale temporal coding module in step S2 is as follows:

[0023] The multi-scale temporal coding module embeds the target vehicle feature sequence. Embedded feature sequences with neighboring vehicles For input;

[0024] The multi-scale temporal coding module includes a short-term feature extraction module, a long-term dependency modeling module, and a critical moment attention mechanism.

[0025] Short-term feature extraction module: The inputs of the multi-scale temporal coding module are arranged according to dimension as (B, (T), extracting change patterns within a short time window through one-dimensional convolution;

[0026]

[0027]

[0028] In the formula, x is the convolution input, and the dimensions of x are arranged as (B, , T ), where B represents the batch size, It is the embedding dimension, where T is the time step; This is the intermediate output of the convolution; It is a short-term time series characteristic; This represents a one-dimensional convolution operation; This indicates one-dimensional batch normalization; This represents the modified linear unit, activation function; express operate; Representation layer normalization;

[0029] Long-term dependency modeling module: The input bidirectional gated loop unit performs state updates simultaneously on both the forward and reverse time series, capturing historical and future contextual information.

[0030]

[0031]

[0032]

[0033] In the formula, This indicates a positive GRU update gate; This represents the hidden state in the previous time step in a forward GRU; This represents the candidate hidden state at the current time step of the forward GRU; This represents the hidden state of the forward GRU at the current moment; This represents the update gate of the reverse GRU; This represents the hidden state in the previous time step in the reverse GRU; This represents the candidate hidden state of the reverse GRU at the current moment; This represents the hidden state of the reverse GRU at the current moment; This represents the final encoded hidden state of the BiGRU (Bidirectional Gated Recurrent Unit).

[0034] Finally, after LayerNorm... Perform normalization to obtain local temporal features. ;

[0035] Attention mechanisms at critical moments: Based on this, a multi-head self-attention mechanism is introduced to calculate the correlation between different time points;

[0036]

[0037] In the formula, Q, K, and V represent the query, key, and value, respectively. The result is obtained through linear mapping and is used to calculate attention weights and weighted summation. This is an attention mask matrix used to mask invalid time steps; Indicates the dimensions of Q and K. Divide by Avoidable Excessively large values ​​can cause the softmax gradient to vanish.

[0038] Finally, the attention weights are summed with V to obtain the output of the attention mechanism, which is... , It can adaptively aggregate features of important moments, enabling the model to focus on the time segments that have the greatest impact on future trajectory prediction.

[0039] Furthermore, the design of the dynamic spatial interaction modeling module in step S2 is as follows:

[0040] I. By introducing physical edge features, a directed graph is constructed that is dynamically updated frame by frame;

[0041] At each frame, the nodes of the graph represent the coded features of the vehicle. ,side This indicates the physical relationship characteristics between vehicles;

[0042]

[0043] In the formula, Represents a relative position vector; , Let ||| represent the relative velocity and relative acceleration vectors, respectively; ||·|| represents the vector magnitude.

[0044] II. Modeling interactive information using a graph attention mechanism that integrates edge features;

[0045] Features of each node Projected onto H attention heads, node features That is, the vehicle's encoded features, and the features of each attention head are calculated as follows:

[0046]

[0047]

[0048]

[0049] In the formula, Indicates the first Features of node i under each attention head; Indicates the first The query weight matrix corresponding to each attention head; Indicates the first Features of node j under each attention head; Indicates the first The key weight matrix corresponding to each attention head; Indicates the first The physical relationship characteristics of the edge between node i and node j under each attention head; Indicates the first The edge feature weight matrix corresponding to each attention head;

[0050] Among them, attention score The calculation formula is as follows:

[0051]

[0052] In the formula, || represents vector concatenation, and λ is the distance penalty coefficient, which is used to reduce the attention score of distant neighboring vehicles; Here is the weight matrix in the attention mechanism of layer h; This is an activation function used to perform a nonlinear transformation on the calculation result within the parentheses;

[0053] The neighbor vehicle attention weights are normalized using softmax, and the calculation formula is as follows:

[0054]

[0055] In the formula, It is the attention weight of the subject vehicle i to its neighbor vehicle j after Softmax normalization. The value is between 0 and 1, and the sum of the attention weights of all neighboring cars is 1; Let represent the attention score of the subject vehicle i in the h-th layer to its neighbor vehicle j; This represents the set of neighboring vehicles of vehicle i at time t;

[0056] The output of each attention head is obtained by weighted aggregation of neighbor node features, and the calculation formula is as follows:

[0057]

[0058] In the formula, It is the projection matrix of the h-th attention head, with dimension . , used to project neighbor node features; This represents the output feature of the h-th attention head for node i;

[0059] After concatenating all attention head outputs, the final node representation is obtained through residual connection and normalization, as follows:

[0060]

[0061] In the formula, This represents the final feature representation of node i after processing by the graph attention network. Represents the learnable projection matrix; This means concatenating the output features of H attention heads; and combining all nodes. By stacking the nodes, we can obtain the feature matrix of all nodes in the current frame. , as subsequent input;

[0062] III. Finally, the output is processed by a gated linear unit (GLU) to enhance its nonlinear expressive power and training stability.

[0063]

[0064] In the formula, This represents the output of the dynamic spatial interaction modeling module; Represents the weight matrix; Indicates that all nodes The node feature matrix is ​​formed by stacking nodes; GLU represents the gated linear unit, and its calculation formula is: Where x1 and x2 are The two-part segmentation result, This is the activation function.

[0065] Furthermore, the trajectory decoding and prediction module in step S2 is designed as follows:

[0066] The trajectory decoding and prediction module is based on a multi-layer Transformer decoder and, with the assistance of guiding vectors and lane semantics, generates a sequence of future trajectory points for the target vehicle through autoregression.

[0067] The trajectory decoding and prediction module uses the output of the dynamic space interaction modeling module. As input;

[0068] First, to generate predictions for future frames T′, a set of learnable query vectors is defined. The query vector for each time step t is obtained by adding four parts, and the calculation formula is as follows:

[0069]

[0070] In the formula, This represents the query vector at time step t. This indicates a learnable query embedding; Indicates time location encoding; The guiding vector, generated from the final hidden state of the target vehicle encoder through the projection layer and the gated linear unit, is used to guide the decoder to focus on the historical features of the target vehicle. Lane embedding is used to provide road geometric semantic information; , indicating the predicted time step range, from 1 to T′;

[0071] The constructed query sequence Q is input to the Transformer decoder, along with the encoder's memory features. Interact with the system to generate trajectory-level feature representations. The calculation formula is as follows:

[0072]

[0073] In the formula, Y is the output of the Transformer decoder, i.e., the trajectory-level feature representation; The query vector sequence has a dimension of , The number of time steps for the future trajectory. The dimension of the encoder output features; sequence Query vectors at each time step This constitutes a guiding sequence for the decoder to generate the trajectory;

[0074] Applying causal masks in decoder self-attention computation To ensure autoregressive constraints, the masking rule is that when j ≤ i, When j > i, This ensures that the prediction at each time step depends only on the current and past prediction results; the decoder outputs features for each time step. The predicted two-dimensional position is obtained through linear mapping, and the calculation formula is as follows:

[0075]

[0076] In the formula, This represents the two-dimensional position predicted at time step t; This is a linear mapping function used to convert feature vectors into two-dimensional positions; This represents the feature vector output by the decoder at time step t.

[0077] Beneficial effects:

[0078] This invention addresses the highly dynamic nature of vehicle interactions in highway scenarios by proposing a multi-scale interactive perception-based method for predicting highway vehicle trajectories. It employs a convolutional neural network and a bidirectional gated recurrent unit to jointly model short-term burst features and long-term evolutionary trends, and introduces a temporal attention mechanism to enhance the perception of key time segments. A dynamic graph attention mechanism incorporating physical edge features such as relative position, relative speed, and relative acceleration is used to update vehicle interaction relationships in real time, improving spatial modeling accuracy and interpretability. During the decoding stage, guidance vectors and lane semantic information are integrated, ensuring that the predicted trajectory conforms to the road geometry spatially and remains smooth and continuous temporally. This method improves the robustness and adaptability of the model in sparse neighboring vehicle environments on highways while maintaining prediction accuracy, providing more stable and reliable trajectory prediction results for intelligent transportation systems and offering strong technical support for highway traffic safety management and operation scheduling.

[0079] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0080] Figure 1 This is a diagram of the overall architecture of the MIST model;

[0081] Figure 2 This is an architecture diagram of the dynamic graph attention module;

[0082] Figure 3 Visualize the RMSE of trajectory predictions on the NGSIM dataset;

[0083] Figure 4 Visualize the RMSE of trajectory predictions on the HighD dataset;

[0084] Figure 5 The prediction results are based on conditions of smooth traffic flow.

[0085] Figure 6 The prediction results are for moderate traffic conditions;

[0086] Figure 7 The prediction results are for conditions of relatively congested traffic.

[0087] Figure 8 Set up visualization for the grid space of the selected samples;

[0088] Figure 9 GAT attention heatmap for nearby vehicles. Detailed Implementation

[0089] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of this application.

[0090] In a dynamic traffic scenario on a highway, assume there are N vehicles. The goal of vehicle trajectory prediction is to predict the future position sequence of a target vehicle (aᵢ∈A) within a fixed prediction time window based on the historical movement trajectory of the target vehicle and the interaction information of its neighboring vehicles.

[0091] For each vehicle aᵢ, its historical trajectory over the past T frames can be represented as a state sequence of length T;

[0092]

[0093] in, This represents the vehicle's state vector at time t, which typically includes features such as position, velocity, and acceleration.

[0094] The goal is to predict the vehicle's position and trajectory within the next T′ frame. The calculation expression is as follows:

[0095]

[0096] in, Let aᵢ represent the predicted two-dimensional position coordinates of vehicle aᵢ at time t+k.

[0097] Since vehicle movement on highways is influenced not only by its own historical behavior but also by the dynamic interference of neighboring vehicles, the prediction function should simultaneously consider the historical sequence of the target vehicle and the spatial interaction context information provided by neighboring vehicles.

[0098]

[0099] in, lᵢ represents the set of neighboring vehicles that have an interaction relationship with vehicle aᵢ at time t; lᵢ represents the lane number (e.g., 0~6) where the target vehicle is located. It is a learnable spatiotemporal prediction function with parameter θ. It is worth noting the neighbor vehicle set. It changes dynamically over time, updating in real time based on information such as the spatial distance and relative speed between the target vehicle and surrounding vehicles, thereby realizing a dynamic interactive graph structure built frame by frame, and adaptively focusing on the most relevant neighboring vehicles to improve the accuracy of prediction.

[0100] This invention provides a multi-scale interactive sensing method for predicting highway vehicle trajectories, comprising the following steps:

[0101] S1. Collect the historical movement trajectory of the target vehicle on the highway and the interaction information of nearby vehicles;

[0102] S2. Construct the MIST model;

[0103] The overall architecture of the MIST model consists of four modules: (1) Input feature embedding module: mapping the trajectory state vectors of the target vehicle and neighboring vehicles to a high-dimensional embedding space, and fusing temporal position encoding and lane semantics; (2) Multi-scale temporal encoding module: using CNN to extract short-term features, BiGRU to model long-term dependencies, and highlighting key moments through temporal attention; (3) Dynamic spatial interaction modeling module: constructing a dynamic graph attention network that fuses physical edge features frame by frame to extract the interaction relationship between the target vehicle and neighboring vehicles; (4) Trajectory decoding and prediction module: based on the Transformer decoder combined with the guide vector and lane semantics, autoregressively generating future trajectories. The overall architecture diagram of the MIST model is shown below. Figure 1 As shown, green vehicles represent target vehicles, dark blue vehicles represent neighboring vehicles, the lines behind the vehicles represent their historical trajectories, and the thicker lines in front of the vehicles represent the predicted future trajectories of the target vehicles.

[0104] Input Feature Embedding Module

[0105] The input feature embedding module takes the original state vector of each vehicle at time t. For input, The six dimensions represent the position coordinates (x, y), velocity components (v), and velocity components, respectively. x , v y ) and acceleration components (a x , a y );

[0106] First, through the linear mapping matrix Project the original state vector onto The initial feature embedding is obtained by using a dimensional embedding space, stabilizing the feature distribution using layer normalization (LayerNorm), and then enhancing the nonlinear expressive power through the LeakyReLU activation function. ;

[0107] In order to incorporate temporal location information, Superimposed learnable temporal location codes , From the learnable matrix Provided, of which T ax The maximum number of historical frames is used; furthermore, to incorporate road geometric semantic information, the lane number lᵢ of the target vehicle is embedded through a lane embedding matrix. Convert to lane embedding vector and with Add them together to obtain the final single-frame embedding vector for each frame. Given a historical time window length of T, stacking the embedding vectors along the time dimension yields the target vehicle embedding feature sequence. , where ego represents the target vehicle currently being predicted. For each neighboring vehicle Similarly, the neighboring vehicle embedding feature sequence is obtained. ,in This represents the k-th neighboring vehicle within the perceived range of the historical time window.

[0108] By using the input feature embedding module, low-dimensional physical quantities such as original position, velocity, and acceleration are uniformly mapped into high-dimensional feature representations that integrate temporal sequence and lane semantics, providing a unified input for subsequent multi-scale temporal modeling and dynamic interaction modeling.

[0109] Multi-scale timing coding module

[0110] This module is used to simultaneously capture short-term variation features and long-term dependency features in the trajectories of the target vehicle and neighboring vehicles, and enhances the model's ability to perceive key moments. The multi-scale temporal coding module embeds the target vehicle's feature sequence. Embedded feature sequences with neighboring vehicles For input.

[0111] The multi-scale temporal coding module includes a short-term feature extraction module, a long-term dependency modeling module, and a key-moment attention mechanism;

[0112] (1) Short-term feature extraction module: The inputs of the multi-scale temporal coding module are arranged according to the dimension as (B, (, T), extracting change patterns within a short time window through one-dimensional convolution;

[0113]

[0114]

[0115] In the formula, x is the convolution input, and the dimensions of x are arranged as (B, , T ), where B represents the batch size, It is the embedding dimension, where T is the time step; This is the intermediate output of the convolution; It is a short-term time series characteristic; This represents a one-dimensional convolution operation; This indicates one-dimensional batch normalization; This represents the modified linear unit, activation function; express operate; Representation layer normalization;

[0116] (2) Long-term dependency modeling module: The input bidirectional gated loop unit performs state updates simultaneously on both the forward and reverse time series, capturing historical and future contextual information.

[0117]

[0118]

[0119]

[0120] In the formula, This indicates a positive GRU update gate; This represents the hidden state in the previous time step in a forward GRU; This represents the candidate hidden state at the current time step of the forward GRU; This represents the hidden state of the forward GRU at the current moment; This represents the update gate of the reverse GRU; This represents the hidden state in the previous time step in the reverse GRU; This represents the candidate hidden state of the reverse GRU at the current moment; This represents the hidden state of the reverse GRU at the current moment; This represents the final encoded hidden state of the BiGRU (Bidirectional Gated Recurrent Unit).

[0121] The formulas for calculating the update gate z and the reset gate r are as follows:

[0122]

[0123]

[0124] In the formula, Indicates the reset gate for a positive GRU; This represents the candidate hidden state of a forward GRU; Refers to the sigmoid activation function; , Update the gate weight parameters for the forward GRU; , Reset the weight parameters of the gate for the positive GRU; , The weight parameters are those of the positive GRU candidate hidden states; This represents the input to the current time step of the GRU. This indicates the hidden state at the previous moment; The hyperbolic tangent activation function; It represents the Hadamardi (or Hadama) stack;

[0125] The calculation method for the reverse GRU is similar to that for the forward GRU. Finally, after LayerNorm... Perform normalization to obtain local temporal features. .

[0126] By combining forward and backward propagation, BiGRU can capture bidirectional dependencies in the input sequence. This design enables the encoder to capture past and future contextual information, which is crucial for modeling vehicle motion patterns and trajectory changes.

[0127] (3) Attention mechanism at critical moments: Based on this, a multi-head self-attention mechanism is introduced to calculate the correlation between different time points;

[0128]

[0129] In the formula, Q, K, and V represent the query, key, and value, respectively. The result is obtained through linear mapping and is used to calculate attention weights and weighted summation. This is an attention mask matrix used to mask invalid time steps; Indicates the dimensions of Q and K. Divide by Avoidable Excessively large values ​​can cause the softmax gradient to vanish.

[0130] Finally, the attention weights are summed with V to obtain the output of the attention mechanism, which is... , It can adaptively aggregate features of important moments, enabling the model to focus on the time segments that have the greatest impact on future trajectory prediction.

[0131] This multi-scale temporal coding module uses a combination of CNN, BiGRU and multi-head temporal attention to extract short-term and long-term temporal features and enhances the focus on key temporal segments, providing multi-scale and context-rich temporal feature representations for subsequent dynamic interaction modeling.

[0132] Dynamic Spatial Interaction Modeling Module

[0133] This module is used to characterize the dynamic spatial interaction relationship between the target vehicle and its neighboring vehicles. By introducing physical edge features, it constructs a directed graph that is dynamically updated frame by frame, and uses the graph attention mechanism (GAT) that fuses edge features to model the interaction information, thereby obtaining a more accurate representation of spatial relationships.

[0134] The design of the dynamic spatial interaction modeling module is as follows:

[0135] I. By introducing physical edge features, a directed graph is constructed that is dynamically updated frame by frame;

[0136] At each frame, the nodes of the graph represent the coded features of the vehicle. ,side This indicates the physical relationship characteristics between vehicles;

[0137]

[0138] In the formula, Represents a relative position vector; , These represent the relative velocity and relative acceleration vectors, respectively; ||·|| represents the vector magnitude. This design incorporates both directional and scalar information, facilitating the interpretation of its physical meaning.

[0139] II. Modeling interactive information using a graph attention mechanism that integrates edge features;

[0140] Features of each node Projected onto H attention heads, node features That is, the vehicle's encoded features, and the features of each attention head are calculated as follows:

[0141]

[0142]

[0143]

[0144] In the formula, Indicates the first Features of node i under each attention head; Indicates the first The query weight matrix corresponding to each attention head; Indicates the first Features of node j under each attention head; Indicates the first The key weight matrix corresponding to each attention head; Indicates the first The physical relationship characteristics of the edge between node i and node j under each attention head; Indicates the first The edge feature weight matrix corresponding to each attention head;

[0145] Among them, attention score The calculation formula is as follows:

[0146]

[0147] In the formula, || represents vector concatenation, and λ is the distance penalty coefficient, which is used to reduce the attention score of distant neighboring vehicles; Here is the weight matrix in the attention mechanism of layer h; This is an activation function used to perform a nonlinear transformation on the calculation result within the parentheses;

[0148] The neighbor vehicle attention weights are normalized using softmax, and the calculation formula is as follows:

[0149]

[0150] In the formula, It is the attention weight of the subject vehicle i to its neighbor vehicle j after Softmax normalization. The value is between 0 and 1, and the sum of the attention weights of all neighboring cars is 1; Let represent the attention score of the subject vehicle i in the h-th layer to its neighbor vehicle j; This represents the set of neighboring vehicles of vehicle i at time t;

[0151] The output of each attention head is obtained by weighted aggregation of neighbor node features, and the calculation formula is as follows:

[0152]

[0153] In the formula, It is the projection matrix of the h-th attention head, with dimension . , used to project neighbor node features; This represents the output feature of the h-th attention head for node i;

[0154] After concatenating all attention head outputs, the final node representation is obtained through residual connection and normalization, as follows:

[0155]

[0156] In the formula, This represents the final feature representation of node i after processing by the graph attention network. Represents the learnable projection matrix; This means concatenating the output features of H attention heads; and combining all nodes. By stacking the nodes, we can obtain the feature matrix of all nodes in the current frame. , as subsequent input;

[0157] III. Finally, the output is processed by a gated linear unit (GLU) to enhance its nonlinear expressive power and training stability.

[0158]

[0159] In the formula, This represents the output of the dynamic spatial interaction modeling module; Represents the weight matrix; Indicates that all nodes The node feature matrix is ​​formed by stacking nodes; GLU represents the gated linear unit, and its calculation formula is: Where x1 and x2 are The two-part segmentation result, This is the activation function.

[0160] The output of this module As input features for subsequent decoders, dynamic spatial interaction representations that integrate multi-scale temporal information and enhanced physical semantics are used. For example... Figure 2 As shown, this diagram simplifies how the dynamic graph attention mechanism models the interaction between a target vehicle and its neighboring vehicles. The left image shows a traffic scene with a target vehicle (green) and surrounding vehicles (blue). The middle image shows the constructed dynamic graph, and the right image visualizes the learned attention weights: the brighter the edge color, the higher the attention intensity, indicating a greater influence from neighboring vehicles. As shown by the rightmost color bar, the gradient from top (bright) to bottom (dark) reflects the decreasing attention weights.

[0161] Trajectory Decoding and Prediction Module

[0162] The trajectory decoding and prediction module is based on a multi-layer Transformer decoder. With the assistance of guide vectors and lane semantics, it generates a sequence of future trajectory points for the target vehicle through autoregression.

[0163] The trajectory decoding and prediction module outputs the dynamic spatial interaction modeling module. As input, it includes multi-scale spatiotemporal interaction information between the target vehicle and neighboring vehicles.

[0164] First, to generate predictions for future frames T′, a set of learnable query vectors is defined. The query vector for each time step t is obtained by adding four parts, and the calculation formula is as follows:

[0165]

[0166] In the formula, This represents the query vector at time step t. This indicates a learnable query embedding; Indicates time location encoding; The guiding vector, generated from the final hidden state of the target vehicle encoder via a projection layer and a gated linear unit (GLU), is used to guide the decoder to focus on the historical features of the target vehicle. Lane embedding is used to provide road geometric semantic information; , indicating the predicted time step range, from 1 to T′;

[0167] The constructed query sequence Q is input to the Transformer decoder, along with the encoder's memory features. Interact with the system to generate trajectory-level feature representations. The calculation formula is as follows:

[0168]

[0169] In the formula, Y is the output of the Transformer decoder, i.e., the trajectory-level feature representation; The query vector sequence has a dimension of , The number of time steps for the future trajectory. The dimension of the encoder output features; sequence Query vectors at each time step This constitutes a guiding sequence for the decoder to generate the trajectory;

[0170] Applying causal masks in decoder self-attention computation To ensure autoregressive constraints, the masking rule is that when j ≤ i, When j > i, This ensures that the prediction at each time step depends only on the current and past prediction results; the decoder outputs features for each time step. The predicted two-dimensional position is obtained through linear mapping, and the calculation formula is as follows:

[0171]

[0172] In the formula, This represents the two-dimensional position predicted at time step t; This is a linear mapping function used to convert feature vectors into two-dimensional positions; This represents the feature vector output by the decoder at time step t.

[0173] By combining query guidance vectors and lane perception context, the trajectory decoding and prediction module enables the decoder to generate more accurate and coherent future trajectories, especially in complex and interactive driving scenarios.

[0174] S3. Input the data collected in step S1 into the MIST model constructed in step S2, and finally output the future trajectory of the target vehicle.

[0175] Experimental verification

[0176] I. Experimental Setup

[0177] Datasets and Dataset Settings

[0178] This invention primarily evaluates the proposed MIST model through four experiments. The first type of experiment is a comparative experiment, used to verify the improvement of the constructed framework compared to other state-of-the-art methods; the second type of experiment is a prediction computation speed experiment, used to analyze the model trajectory prediction efficiency; the third and fourth types of experiments are ablation and visualization experiments, respectively, used to demonstrate the effectiveness of each module of the framework.

[0179] This invention selected two real-world highway public datasets, NGSIM and HighD, for experiments. Each dataset includes real vehicle trajectory driving data, and key statistical data are shown in Table 1.

[0180] Table 1. Comparison of key statistics between NGSIM and HIGHD datasets

[0181]

[0182] The NGSIM dataset contains detailed vehicle trajectory data collected from two U.S. highway segments (US 101 and I-80). Each segment contains 45 minutes of vehicle trajectories sampled at a 10Hz sampling rate, and each 45-minute dataset contains three 15-minute subsets recorded at different time intervals. This dataset was collected by the U.S. Department of Transportation in 2015. Data preprocessing and formatting followed the same procedures as CS-LSTM to ensure fair and direct comparisons. Specifically, the sampling rate for each vehicle trajectory was reduced by a factor of 2, resulting in a time step of 0.2 seconds. A 3×13 spatial grid centered on the target vehicle was constructed based on its relative position in the scene, with each grid cell 15 feet (approximately 4.6 meters) long and the same width as the lane. The dataset was divided into training, validation, and test sets in a ratio of 0.7:0.1:0.2.

[0183] The HighD dataset was collected in 2017 and 2018 using camera-equipped drones at a frame rate of 25 Hz. These recordings are taken from an aerial perspective, covering six different German highways. Compared to NGSIM, the HighD dataset contains significantly more data, covering over 110,000 vehicles, approximately 12 times the amount in NGSIM. The total driving distance covered in this dataset is 45,000 kilometers. Similar to the NGSIM dataset, we observe 3 seconds of historical trajectories and predict 5 seconds of future trajectories.

[0184] Contrast Model

[0185] To evaluate the effectiveness of the MIST model, this invention selected CS-LSTM, PiP, MHA-LSTM, STA-LSTM, STDAN, DeepTrack, WSiP, FHIF, HTP, and VT-Former baseline methods for comparative experiments. The baseline methods are described below:

[0186] CS-LSTM: This model is an encoder-decoder based model that captures the social interaction features of vehicles by stacking convolutional layers and pooling layers;

[0187] PiP: This model combines vehicle planning and spatiotemporal interaction to predict future trajectories and uses an attention mechanism to capture the effects of historical motion and surrounding agents.

[0188] MHA-LSTM: This model is an encoder-decoder based model that uses a multi-head attention mechanism to capture the importance of surrounding vehicles and combines additional features such as speed, acceleration and category to generate multimodal trajectories for the target vehicle.

[0189] STA-LSTM: This model combines an attention mechanism and uses spatiotemporal information to explain the influence of historical trajectories and neighboring vehicles on the vehicle itself;

[0190] STDAN: This model uses attention mechanisms to extract features across different dimensions, including self-attention and attention mechanisms, to extract temporal and spatial interaction features.

[0191] Deeptrack: This model uses a time-convolutional network to achieve a lightweight design, which significantly reduces computational and storage overhead while maintaining high prediction accuracy;

[0192] WSiP: This model combines a wave superposition-inspired pooling method to capture dynamic and high-order interactions between vehicles.

[0193] FHIF: This model extracts the historical interaction features between the target vehicle and neighboring vehicles through a multi-head attention mechanism, and fuses behavioral intent and temporal trajectory features through a multimodal fusion module;

[0194] HTP: This model is a hybrid trajectory prediction framework based on LSTM encoding, which introduces a reactive social convolutional structure and a spatiotemporal attention mechanism;

[0195] VT-Former: This model is based on the Transformer decoder and combines graph attention mechanism to capture the social relationships of vehicles.

[0196] Model evaluation metrics

[0197] This experiment uses the root mean square error (RMSE) as the evaluation metric. Let... This represents the actual position of the i-th vehicle at time step t. Indicates its predicted location, with a prediction time range of [missing information]. Let N represent the number of predicted vehicle trajectories. Then, the RMSE calculation formula at time step t is as follows:

[0198]

[0199] Where N represents the number of vehicles effectively predicted at time step t. The average RMSE over the prediction time range is shown in the following formula:

[0200]

[0201] II. Comparative Experimental Results Analysis

[0202] This invention reports experimental results of MIST and other contrasting models. The quantitative trajectory prediction performance of all contrasting models on the NGSIM and HighD datasets is summarized in Tables 2 and 3. Figure 3 and Figure 4 The visualization was performed.

[0203] Table 2 Performance comparison of MIST and baseline methods on the NGSIM dataset

[0204]

[0205] Table 3 Performance comparison of MIST and baseline methods on the HighD dataset.

[0206]

[0207] The experimental results demonstrate that our proposed MIST model achieves highly competitive performance on the NGSIM dataset. While it slightly lags behind baseline models in short-term predictions (1–2 seconds), it surpasses all baseline models in long-term predictions (3–5 seconds), indicating its superior ability to model long-term temporal dependencies. On the HighD dataset, the MIST model also achieves state-of-the-art results at most prediction time points. Furthermore, on both the NGSIM and HighD datasets, MIST exhibits the lowest average RMSE across all prediction time steps, demonstrating its strong stability across diverse driving scenarios.

[0208] III. Model Calculation Speed ​​Analysis

[0209] Table 4 compares the inference complexity of several models on the NGSIM dataset in terms of multiplication-accumulation floating-point operations (MFLOPs) and average inference time per trajectory (ms). Since inference cost depends on the programming method and computing device, we selected CS-LSTM, PiP, STDAN, and FHIF models for comparison; the code for these models is publicly available or their relevant metrics are clearly published. The results show that although MIST has higher computational complexity than lightweight baselines such as CS-LSTM and FHIF, its inference time remains relatively efficient, at approximately 0.115 ms per trajectory, which is within the acceptable range for real-time trajectory prediction. This indicates that MIST achieves a good balance between modeling capability and practical inference efficiency.

[0210] Table 4. Prediction speed of the MIST model on the NGSIM dataset

[0211]

[0212] IV. Effectiveness Analysis of Method Components

[0213] To test the effectiveness of each component of the method of this invention, we conducted ablation experiments on the NGSIM dataset, removing key modules and components from the complete model. The following variants were evaluated:

[0214] (1) w / o MTEM: Remove the multi-scale temporal encoding module, which uses a CNN-BiGRU-Attention structure to model short-term temporal features and long-term temporal dependencies.

[0215] (2) w / o DGAM: Remove the Dynamic Graph Attention Space Modeling module, which is used to model the spatial dependencies of the interaction perception between agents.

[0216] (3) w / o CNN: Remove the convolutional component in MTEM responsible for extracting short-term temporal features.

[0217] (4) w / o BiGRU: Remove BiGRU from MTEM, which is used to capture continuous long-term time dependencies.

[0218] (5) w / o edge features: Remove physical edge features in DGAM that are used to guide attention weights during interaction modeling.

[0219] Table 5 summarizes the performance of each ablation variant. The results show that both MTEM and DGAM are crucial for accurate trajectory prediction, and removing either leads to a significant performance degradation, highlighting the importance of capturing multi-scale temporal features and dynamic spatial interactions between agents. In MTEM, removing the CNN significantly reduces short-term prediction accuracy, indicating that short-term temporal features are crucial for early motion estimation. Similarly, removing the BiGRU results in a consistent performance degradation across all prediction ranges, confirming its role in long-term temporal dependency modeling. In DGAM, removing physical edge features also leads to performance degradation, especially in long-term predictions, highlighting the effectiveness of incorporating physical priors into interaction modeling. Overall, the consistent performance degradation across all ablation settings validates the effectiveness of the proposed module and its internal components.

[0220] Table 5 Ablation experimental results of the MIST model on the NGSIM dataset

[0221]

[0222] V. Visualization Experiment

[0223] To visually evaluate the performance of MIST under different traffic conditions, we visualize the prediction results under different maneuvers and traffic densities. Inspired by STDAN, we categorize scenarios into three types based on traffic density: light, moderate, and heavy. For each traffic density, we present representative samples of left lane change, lane keeping, and right lane change behaviors from the NGSIM dataset. Figure 5 –7 shows the prediction results. In the figure, (a) is the prediction result for a left lane change, (b) is the prediction result for maintaining the lane, and (c) is the prediction result for a right lane change. The historical trajectory of the target vehicle is represented by a red dashed line; the historical trajectories of adjacent vehicles are represented by blue solid lines; the actual future trajectory is represented by a yellow dotted line; and the future trajectory predicted by the MIST model is represented by a green solid line.

[0224] The visualization results demonstrate that our proposed model can generate reasonable and reliable predicted trajectories under various traffic conditions and different behaviors. This shows that the MIST model can adapt to complex and changing traffic conditions and different vehicle behaviors, effectively capturing vehicle motion patterns in different scenarios.

[0225] To further illustrate how the proposed model captures dynamic spatial interactions between agents, we randomly select a sample from the test set and analyze the attention weights assigned to surrounding vehicles by the Graph Attention Network (GAT). Figure 8 As shown in the scene, green vehicles represent the target vehicle, while surrounding blue vehicles represent valid neighboring vehicles in the current frame, with the overall driving direction from left to right. The scene's spatial area is divided into a 3×13 grid structure, covering a 90-foot radius around the target vehicle along the vertical axis, including the current lane and the adjacent left and right lanes, totaling 39 grid cells. Each grid cell spans 15 feet vertically and approximately 12 feet horizontally, corresponding to a standard lane width. The grid indexes start from the left front of the target vehicle: indices 1-13 correspond to the left lane, 14-26 to the middle lane, and 27-39 to the right lane.

[0226] Table 6 shows the edge information between the target vehicle and its six valid neighbors in this scenario, including relative positions. (in feet), Euclidean distance (feet), relative speed Relative acceleration And the attention weight W assigned by the GAT module, which reflects the importance of interaction with each neighboring vehicle. Figure 9 The spatial visualization of these attention weights is visually displayed using a 13×3 grid structure, compared with... Figure 8 The grid positions correspond to the target vehicle, which is located in the center and indicated by a black dashed box.

[0227] Table 6 Neighbor Features and GAT Attention Weights (Example Case)

[0228]

[0229] The heatmap reveals that MIST assigns significantly higher attention weights to vehicles in the area to the right front of the target vehicle. Notably, even when a vehicle is not spatially closest, MIST tends to allocate more attention to vehicles exhibiting active interaction patterns, such as those rapidly approaching or accelerating. For example, although vehicle 2 is relatively close to the target vehicle, its relatively stable motion results in a lower attention weight. In contrast, while vehicle 6 is spatially farther from the target vehicle, its obvious dynamic changes lead to higher attention. This indicates that MIST focuses on temporal dynamics rather than simple positional relationships.

[0230] This attention allocation behavior reflects an interpretable logic consistent with real-world driving behavior. In practice, drivers tend to pay more attention to surrounding vehicles that may pose potential risks, such as those approaching rapidly, decelerating suddenly, or exhibiting unpredictable behavior. In contrast, nearby vehicles with stable motion typically receive less attention. Therefore, the attention distribution learned by MIST not only enhances the model's ability to dynamically perceive interactive situations but also demonstrates decision-making patterns at the behavioral level that align with human driving intuition.

[0231] It is hereby declared that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting highway vehicle trajectories using multi-scale interactive sensing, characterized in that, Includes the following steps: S1. Collect the historical movement trajectory of the target vehicle on the highway and the interaction information of nearby vehicles; S2. Construct the MIST model; The MIST model includes an input feature embedding module, a multi-scale temporal coding module, a dynamic spatial interaction modeling module, and a trajectory decoding and prediction module. Input Feature Embedding Module: Taking the original state vector of each vehicle at time t as input, it maps the trajectory state vectors of the target vehicle and its neighboring vehicles to a high-dimensional embedding space, and fuses temporal and positional encoding with lane semantics, finally outputting the target vehicle's embedded feature sequence. Embedded feature sequences with neighboring vehicles For input; Multi-scale temporal coding module: used to simultaneously capture short-term variation features and long-term dependency features in the trajectories of the target vehicle and neighboring vehicles, and enhance the model's ability to perceive key moments; the multi-scale temporal coding module uses the target vehicle embedded feature sequence. Embedded feature sequences with neighboring vehicles As input, sequence features are extracted through a Conv1D layer, and the ReLU activation function is used to enhance nonlinear expressive power. A bidirectional gated recurrent unit is used to capture the bidirectional temporal information of the sequence, and LayerNorm is used to improve training stability. A multi-head self-attention mechanism is introduced to calculate the correlation between different time steps, and finally the encoded features of the vehicle are output. ; Dynamic spatial interaction modeling module: used to depict the dynamic spatial interaction relationship between the target vehicle and neighboring vehicles; The dynamic spatial interaction modeling module uses the vehicle's coded features As input, a graph attention mechanism that fuses edge features is used to model interactive information, and a gated linear unit (GLU) is used to enhance nonlinear expressive power and training stability. The final output is... ; Trajectory Decoding and Prediction Module: Used to generate the future trajectory of the target vehicle; the trajectory decoding and prediction module uses the output of the dynamic spatial interaction modeling module. As input, based on a multi-layer Transformer decoder, and with the assistance of guide vectors and lane semantics, an autoregressive sequence of future trajectory points for the target vehicle is generated. S3. Input the data collected in step S1 into the MIST model constructed in step S2, and finally output the future trajectory of the target vehicle.

2. The method for predicting highway vehicle trajectories using multi-scale interactive sensing according to claim 1, characterized in that, The design of the input feature embedding module in step S2 is as follows: The input feature embedding module uses the original state vector of each vehicle at time t. For input, The six dimensions represent the position coordinates (x, y), velocity components (v), and velocity components, respectively. X , v y ) and acceleration components (a X , a y ); First, through the linear mapping matrix Project the original state vector to The initial feature embedding is obtained by using a dimensional embedding space, stabilizing the feature distribution using layer normalization, and then enhancing the nonlinear expressive power through the LeakyReLU activation function. ; exist Superimposed learnable temporal location codes , From the learnable matrix Provided, of which T max Maximum number of historical frames; In addition, to incorporate road geometric semantic information, the lane number lᵢ of the target vehicle is embedded through a lane embedding matrix. Convert to lane embedding vector and with Add them together to obtain the final single-frame embedding vector for each frame. ; within a historical time window of length Under these conditions, stacking the embedding vectors along the time dimension yields the target vehicle embedding feature sequence. Where ego represents the target vehicle currently being predicted; for each neighboring vehicle Similarly, the neighboring vehicle embedding feature sequence is obtained. ,in This represents the k-th neighboring vehicle within the perceived range of the historical time window.

3. The method for predicting highway vehicle trajectories using multi-scale interactive sensing according to claim 2, characterized in that, The design of the multi-scale temporal coding module in step S2 is as follows: The multi-scale temporal coding module embeds the target vehicle feature sequence. Embedded feature sequences with neighboring vehicles For input; The multi-scale temporal coding module includes a short-term feature extraction module, a long-term dependency modeling module, and a critical moment attention mechanism. Short-term feature extraction module: The inputs of the multi-scale temporal coding module are arranged according to dimension as (B, (, T), extracting change patterns within a short time window through one-dimensional convolution; In the formula, x is the convolution input, and the dimensions of x are arranged as (B, , T ), where B represents the batch size, Here, T represents the embedding dimension, T is the time step, and h̃ is the intermediate output of the convolution. It is a short-term time series characteristic; This represents a one-dimensional convolution operation; This indicates one-dimensional batch normalization; This represents the modified linear unit, activation function; express operate; Representation layer normalization; Long-term dependency modeling module: The input bidirectional gated loop unit performs state updates simultaneously on both the forward and reverse time series, capturing historical and future contextual information. In the formula, This indicates a positive GRU update gate; This represents the hidden state in the previous time step in a forward GRU; This represents the candidate hidden state at the current time step of the forward GRU; This represents the hidden state of the forward GRU at the current moment; This represents the update gate of the reverse GRU; This represents the hidden state in the previous time step in the reverse GRU; This represents the candidate hidden state of the reverse GRU at the current moment; This represents the hidden state of the reverse GRU at the current moment; This represents the final encoded hidden state of the BiGRU (Bidirectional Gated Recurrent Unit). Finally, after LayerNorm... Perform normalization to obtain local temporal features. ; Attention mechanisms at critical moments: Based on this, a multi-head self-attention mechanism is introduced to calculate the correlation between different time points; In the formula, Q, K, and V represent the query, key, and value, respectively. The result is obtained through linear mapping and is used to calculate attention weights and weighted summation. This is an attention mask matrix used to mask invalid time steps; Indicates the dimensions of Q and K; Finally, the attention weights are summed with V to obtain the output of the attention mechanism, which is... , It can adaptively aggregate features of important moments, enabling the model to focus on the time segments that have the greatest impact on future trajectory prediction.

4. The method for predicting highway vehicle trajectories using multi-scale interactive sensing according to claim 3, characterized in that, The design of the dynamic spatial interaction modeling module in step S2 is as follows: I. By introducing physical edge features, a directed graph is constructed that is dynamically updated frame by frame; At each frame, the nodes of the graph represent the coded features of the vehicle. ,side This indicates the physical relationship characteristics between vehicles; In the formula, Represents a relative position vector; , Let ||| represent the relative velocity and relative acceleration vectors, respectively; ||·|| represents the vector magnitude. II. Modeling interactive information using a graph attention mechanism that integrates edge features; Features of each node Projected onto H attention heads, node features That is, the vehicle's encoded features, and the features of each attention head are calculated as follows: In the formula, Indicates the first Features of node i under each attention head; Indicates the first The query weight matrix corresponding to each attention head; Indicates the first Features of node j under each attention head; Indicates the first The key weight matrix corresponding to each attention head; Indicates the first The physical relationship characteristics of the edge between node i and node j under each attention head; Indicates the first The edge feature weight matrix corresponding to each attention head; Among them, attention score The calculation formula is as follows: In the formula, || represents vector concatenation, and λ is the distance penalty coefficient, which is used to reduce the attention score of distant neighboring vehicles; Here is the weight matrix in the attention mechanism of layer h; This is an activation function used to perform a nonlinear transformation on the calculation result within the parentheses; The neighbor vehicle attention weights are normalized using softmax, and the calculation formula is as follows: In the formula, It is the attention weight of the subject vehicle i to its neighbor vehicle j after Softmax normalization. The value is between 0 and 1, and the sum of the attention weights of all neighboring cars is 1; Let represent the attention score of the subject vehicle i in the h-th layer to its neighbor vehicle j; This represents the set of neighboring vehicles of vehicle i at time t; The output of each attention head is obtained by weighted aggregation of neighbor node features, and the calculation formula is as follows: In the formula, It is the projection matrix of the h-th attention head, with dimension . , used to project neighbor node features; This represents the output feature of the h-th attention head for node i; After concatenating all attention head outputs, the final node representation is obtained through residual connection and normalization, as follows: In the formula, This represents the final feature representation of node i after processing by the graph attention network. Represents the learnable projection matrix; This means concatenating the output features of H attention heads; and combining all nodes. By stacking the nodes, we can obtain the feature matrix of all nodes in the current frame. , as subsequent input; III. Finally, the output is processed by a gated linear unit (GLU) to enhance its nonlinear expressive power and training stability. In the formula, This represents the output of the dynamic spatial interaction modeling module; Represents the weight matrix; Indicates that all nodes The node feature matrix is ​​formed by stacking nodes; GLU represents the gated linear unit, and its calculation formula is: Where x1 and x2 are The two-part segmentation result, This is the activation function.

5. The method for predicting highway vehicle trajectories using multi-scale interactive sensing according to claim 4, characterized in that, The trajectory decoding and prediction module in step S2 is designed as follows: The trajectory decoding and prediction module is based on a multi-layer Transformer decoder and, with the assistance of guiding vectors and lane semantics, generates a sequence of future trajectory points for the target vehicle through autoregression. The trajectory decoding and prediction module uses the output of the dynamic space interaction modeling module. As input; First, to generate predictions for future frames T′, a set of learnable query vectors is defined. The query vector for each time step t is obtained by adding four parts, and the calculation formula is as follows: In the formula, This represents the query vector at time step t; This indicates a learnable query embedding; Indicates time location encoding; The guiding vector, generated from the final hidden state of the target vehicle encoder through the projection layer and the gated linear unit, is used to guide the decoder to focus on the historical features of the target vehicle. Lane embedding is used to provide road geometric semantic information; , indicating the predicted time step range, from 1 to T′; The constructed query sequence Q is input to the Transformer decoder, along with the encoder's memory features. Interact with the system to generate trajectory-level feature representations. The calculation formula is as follows: In the formula, Y is the output of the Transformer decoder, i.e., the trajectory-level feature representation; The query vector sequence has a dimension of , The number of time steps for the future trajectory. The dimension of the encoder output features; sequence Query vectors at each time step This constitutes a guiding sequence for the decoder to generate the trajectory; Applying causal masks in decoder self-attention computation To ensure autoregressive constraints, the masking rule is that when j ≤ i, When j > i, This ensures that the prediction at each time step depends only on the current and past prediction results; the decoder outputs features for each time step. The predicted two-dimensional position is obtained through linear mapping, and the calculation formula is as follows: In the formula, This represents the two-dimensional position predicted at time step t; This is a linear mapping function used to convert feature vectors into two-dimensional positions; This represents the feature vector output by the decoder at time step t.

Images