Neural network-based evolutionary time perception trajectory prediction method and device

Abstract

The invention relates to a trajectory prediction method and device based on evolutionary time perception of a neural network, and aims at a trajectory prediction task in a traffic scene, and predicts a driving trajectory of a road participant in a future period of time by using historical state information of the road participant and road information of the traffic scene where the road participant is located. According to the method, firstly, a learnable constant time code is defined, then, a method for capturing interaction information generated by a plurality of road participants in a specific timestamp is designed, meanwhile, a cross-timestamp dynamic interaction capturing method is provided, and finally, a global interaction aggregation method is provided. The final global interaction information is used for track prediction at future moments. According to the method, the evolution mode of static interaction occurring on a specific timestamp and dynamic interaction occurring across timestamps is comprehensively considered, and more accurate trajectory prediction can be achieved.

Description

Technical Field

[0001] The present invention belongs to the fields of information technology and autonomous driving technology, and specifically relates to a trajectory prediction method and device based on evolutionary time perception of a neural network. Background Art

[0002] Safely operating in complex traffic scenarios is crucial for intelligent systems such as autonomous driving platforms and robotic systems. To achieve safe and stable operation, participants in traffic scenarios must perceive changes in their surroundings in real time while driving. This surrounding environment includes road information such as the movements of other road users and the structure of the road section. Due to its importance to intelligent systems, motion prediction has become a hot topic in the field of autonomous driving. The goal is to predict the future trajectory of each road user based on their historical state, where each user's state is related to their spatiotemporal position and the road they are on.

[0003] However, the travel intentions of each participant are unknown, and there are complex interactions between multiple participants that evolve over time, posing a challenge to accurately predicting their future trajectories. To achieve more accurate trajectory prediction, most existing methods that simulate interactions between participants using graph structures fail to fully exploit the inherent temporal characteristics of interaction sequences when modeling interaction information. Furthermore, these graph neural network-based methods lack comprehensive consideration of both static interactions occurring at specific timestamps and the dynamic interaction evolution patterns that occur across timestamps during interaction modeling. Summary of the Invention

[0004] Considering that in real complex traffic scenarios, the spatiotemporal states of participants will continuously evolve according to changes in their surrounding environment (including the states of surrounding participants and the road structure at different times), the present invention aims to propose an evolutionary time-aware trajectory prediction method and device based on neural networks. By simultaneously modeling the static interactions occurring at specific timestamps and the dynamic interaction evolution patterns occurring across timestamps, the learned interaction information is finally aggregated to obtain a global interaction representation, which is used to predict the driving trajectories of participants at future moments.

[0005] The technical solution adopted in the present invention is as follows:

[0006] A neural network-based evolutionary time-aware trajectory prediction method is characterized by comprising the following steps: S1, acquiring trajectory data of a traffic scene; S2, extracting static interaction features of a specific timestamp and evolutionary dynamic interaction features across timestamps from the trajectory data of the traffic scene using a neural network; S3, aggregating the extracted static interaction features of the specific timestamp and the evolutionary dynamic interaction features across timestamps to obtain a global interaction representation; S4, using the global interaction representation to predict the driving trajectory of participants at future moments.

[0007] Furthermore, the acquisition of the trajectory data of the traffic scene in S1 specifically includes:

[0008] S11. Trajectory sequence formalization: Define the scope of the surrounding environment of the autonomous vehicle. Assume that there are N participants within the scope of the autonomous vehicle, where the state of the i-th participant at a specific time stamp t in the historical time period includes its spatial position Yaw angle and speed t∈[1:T]; the road segment ξ is represented in vector form as in Represents the starting coordinate point of the road segment. Represents the end coordinate point of the road section; given the state information and road section information of the road participants in the historical observation window of T time steps, the historical trajectory of the i-th participant is expressed as in Indicates spatial location The x-axis value of the point, Indicates spatial location The y-coordinate value of the point.

[0009] S12. Dataset division: Divide the trajectory sequence data after formal processing into three subsets: training set, validation set, and test set.

[0010] Furthermore, extracting the static interaction feature of the specific timestamp specifically includes:

[0011] S21. Use constant time encoding for a specific timestamp t:

[0012] TE(t)=[cos(ω1t+φ1),…,cos(ω d t+φ d )]

[0013] where ω * and φ * represents the learnable parameter that controls the temporal encoding, and d represents the dimension of the temporal encoding embedding TE(t);

[0014] S22. In the encoding stage, the i-th participant is taken as the central node, the surrounding participants are taken as neighbor nodes, and j represents any neighbor node. Multiple MLP blocks are used to encode the state information of the central node i, the state information of the neighbor node j, and the relative positions between the nodes. The encoding of the relative position preserves the spatial relationship between the central node and the neighbor nodes, and the encoding is obtained as:

[0015]

[0016] in, Represents the encoding of the state information of the central node i, Represents the encoding of the state information of neighbor node j, The code representing the relative position between nodes i and j;

[0017] S23. Assign different weights to each neighbor node using the following method:

[0018]

[0019] Where W i ,W j ,W ij Represents the transformation matrix of the central node, neighbor nodes and their relative positions, N i represents the set of neighbor nodes around the central node i, || represents the vector splicing operation, Represents the calculated neighbor node weight at any specific time stamp t;

[0020] S24, according to the obtained neighbor node weight at the specific time stamp t Aggregate all neighbor node information to obtain the static interaction features at a specific time stamp t within the historical time period:

[0021] Furthermore, extracting the evolutionary dynamic interaction features across timestamps specifically includes:

[0022] S25. Use a global time window to focus on all the encoding history states before the current specific timestamp t∈[1:T]. Assume that the long-term dependency feature captured at the previous moment t-1 is The long-term dependency features captured at the current moment are:

[0023]

[0024] in represents a time series model;

[0025] S26, using a local time window to focus on the historical state of two adjacent timestamps, and comprehensively considering the long-term dependency features obtained at the previous moment, assuming that the short-term dependency features captured at the previous moment t-1 are Long-term dependence is characterized by A custom gating mechanism is used to control the weights of features at different moments in the local time window. The short-term dependency features at the current moment are:

[0026]

[0027] Where ⊙ represents the element-by-element multiplication operation, W g A learnable matrix representing the linear mapping between hidden layers, Represents the time-aware gating matrix, which is designed to implement a custom gating mechanism and is calculated from the short-term dependency features of the previous moment:

[0028]

[0029] Where W1 and b1 are linear mapping learnable parameters in the gating mechanism execution process, and σ(·) is the sigmoid function.

[0030] Furthermore, in S3, the extracted static interaction features of a specific timestamp and the evolving dynamic interaction features across timestamps are aggregated to obtain a global interaction representation, which specifically includes:

[0031] All static interaction features and evolving dynamic interaction features are aggregated into a fully connected graph, and initialized learnable vectors are provided for participants that are ignored within a bounded range. Dot product attention is then performed on the resulting fully connected graph, where the features of the central node are converted into a query vector, and the features of the neighboring nodes are converted into keys and values:

[0032]

[0033] Where W q ,W k ,W v Represents the linear mapping matrix of the three components respectively; when constructing the fully connected graph, the features of the last timestamp T are used, and then the three vectors obtained are used as input to calculate the global interaction representation:

[0034]

[0035] The final output vector S is the global interaction representation.

[0036] Furthermore, the use of the global interactive representation to predict the participant's driving trajectory at a future moment includes:

[0037] Assume that the state information of the next T′ time steps is predicted, that is, the timestamp t∈[T:T+T′] of the prediction stage. Given a future timestamp t, the prediction representation at this timestamp is expressed as follows:

[0038] S′=[S||TE(t)]

[0039] Input the predicted representation S′ and pass it through the prediction function to obtain the possible future trajectory of the participant:

[0040] Y=predict({S′ i t} i∈N,t∈[T:T+T′] )

[0041] Where Y represents the predicted trajectory of the output, predict(·) represents the prediction function, and S′ i t represents the predicted representation at the timestamp t.

[0042] Furthermore, predicting the future trajectory of road participants is considered as a regression problem, and selecting the best trajectory from the generated K trajectories is considered as a classification problem; for regression loss The negative log-likelihood function is used to calculate the error between the predicted trajectory and the true trajectory:

[0043]

[0044] where P(·|·) is the probability density function of the Laplace distribution, which is determined by the position of the best predicted trajectory of participant i and the uncertainty parameter Jointly determined; In addition, for the classification loss Cross entropy is used to optimize the mixing coefficient M; the final loss function is the sum of regression loss and classification loss, and both are equally weighted:

[0045]

[0046] In another embodiment, the present invention further provides an evolutionary time-aware trajectory prediction device based on a neural network, comprising:

[0047] The trajectory data processing module is used to cut the raw trajectory data collected from the traffic scene according to the timestamp and standardize it into formal input, and then divide the formal trajectory sequence into different data sets;

[0048] A static interaction feature extraction module is used to extract static interaction features of specific timestamps from the trajectory data of traffic scenes using a neural network;

[0049] An evolutionary dynamic interaction feature extraction module, which uses a neural network to extract cross-timestamp evolutionary dynamic interaction features from the trajectory data of traffic scenes;

[0050] The interaction feature aggregation module is used to aggregate the extracted static interaction features of a specific timestamp and the evolving dynamic interaction features across timestamps to obtain a global interaction representation;

[0051] The trajectory prediction module is used to predict the driving trajectory of participants in the future by using the global interaction representation.

[0052] In another embodiment, the present invention further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing the above method.

[0053] In another embodiment, the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the above method is implemented.

[0054] The beneficial effects of the present invention are as follows:

[0055] The present invention is aimed at trajectory prediction tasks in traffic scenarios. It uses the historical status information of road participants (vehicles, pedestrians) and the road information of the traffic scenario to predict the driving trajectories of road participants in the future, and proposes an evolutionary time-aware trajectory prediction scheme based on neural networks. The present invention first defines a learnable constant time code, then designs a method for capturing the interaction information generated by multiple road participants within a specific timestamp, and proposes a dynamic interaction capture method across timestamps. Finally, a global interaction aggregation method is proposed to obtain the final global interaction information for trajectory prediction at future moments. The present invention comprehensively considers the static interactions that occur at specific timestamps and the dynamic interaction evolution patterns that occur across timestamps, and can achieve more accurate trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of the trajectory prediction method based on neural network evolutionary time perception of the present invention. DETAILED DESCRIPTION

[0057] The present invention is further described in detail below through specific embodiments and drawings.

[0058] This paper aims at trajectory prediction tasks, and uses the status information of road participants (such as spatiotemporal position, speed, direction, etc.) and the road information of the road section in a historical period to predict the possible movement trajectory of participants in the future. It proposes an evolutionary time-aware trajectory prediction method based on neural networks. The entire execution process of this method is as follows: Figure 1 As shown in the figure, it mainly includes: trajectory data processing module, static interaction feature extraction module, evolutionary dynamic interaction feature extraction module, interaction feature aggregation module and the final trajectory prediction module.

[0059] Among them, the trajectory data processing module is mainly responsible for cutting the raw trajectory data collected from the traffic scene according to timestamps and standardizing it into formal input acceptable to the model. Subsequently, the formalized trajectory sequence is divided into different data sets; the static interaction feature extraction module is mainly responsible for using neural networks to obtain the interaction information existing at specific timestamps in the trajectory sequence; and the evolutionary dynamic interaction feature extraction module is responsible for obtaining the interaction patterns that evolve over time in continuous time periods in the trajectory sequence; the interaction feature aggregation module is responsible for aggregating the interaction information in the entire traffic scene, including some interaction information ignored in previous steps; finally, the trajectory prediction module is responsible for outputting multiple reasonable future trajectories to complete the trajectory prediction task.

[0060] S1. Obtain the trajectory data of the traffic scene and divide it into data sets. Specifically, it includes:

[0061] S11. Trajectory sequence formalization.

[0062] In an autonomous driving system, the perception module provides the prediction module with a stream of road participant states at regular intervals, where each state is associated with its spatiotemporal location and the road segment it is on. Due to the wide range of traffic scenarios, computational resources and time are limited, and trajectory prediction tasks require extremely short timeframes for future predictions.

[0063] Therefore, the present invention limits the scope of the surrounding environment of the autonomous vehicle when processing data. In this invention, only other road participants within a radius of 50m are considered. Assuming that there are N participants surrounding the autonomous vehicle within the limited range, the state of the i-th participant at a specific time stamp t in the historical time period includes its spatial position Yaw angle and speed For the road segment ξ, it is expressed in vector form as: in Represents the starting coordinate point of the road segment. Represents the end coordinate point of the road section. This vector representation transforms the original road section point set into a vector set, ensuring the translation invariance of the road information.

[0064] In this paper, given the state information and road section information of a road participant within a historical observation window of T time steps, the prediction module is expected to output K possible driving trajectories of the participant in the next T′ time steps and assign a probability score to each possible trajectory. The participant's historical trajectory is represented as: t∈[1:T].

[0065] The present invention uses the relationship between nodes and edges in graph neural networks to fit the interactive relationship between multiple road participants in traffic scenes, where the autonomous driving vehicle is regarded as the central node, the surrounding vehicles are regarded as neighbor nodes, and the relationship between nodes can be represented by edges.

[0066] It is worth noting that in order to effectively analyze the interaction characteristics between participants, the present invention uniformly converts the last historical trajectory segment of the central node c into As a reference vector, and according to the yaw angle of the reference vector Rotates all input vectors in the scene.

[0067] S12. Dataset division.

[0068] The present invention adopts a supervised learning method to model the temporal characteristics and interaction patterns of trajectory data in traffic scenarios, and divides the trajectory sequence data after formal processing in step 1 into three subsets in a ratio of 8:1:1: training set, validation set, and test set.

[0069] In terms of the input and output of the model, a historical subgraph of length T (such as timestamp t∈[1:T]) is constructed as the input of the model based on the historical state information. The predicted trajectory of the future moment (such as timestamp t∈[T:T+T′]) and the confidence of each trajectory are used as output. The trajectory of the future moment (such as timestamp t∈[T:T+T′]) collected in the real scene is used as supervision information.

[0070] S2. A neural network is used to extract static interaction features of specific timestamps and evolving dynamic interaction features across timestamps from the trajectory data of traffic scenes.

[0071] Based on the data processing results from step S1, the trajectory prediction model designed by this invention primarily mines the various interactive information present in the trajectory sequence. By combining temporal features with interactive information, it comprehensively considers both static interactive features at specific timestamps and dynamic interactive features that evolve along the timeline.

[0072] Furthermore, considering that selecting only a limited range of neighboring participants during data processing may overlook some distant but critical interaction information, this paper proposes an interaction feature aggregation module to address this issue. The following describes the static interaction feature extraction, dynamic interaction feature extraction, and interaction feature aggregation methods, respectively.

[0073] Time-aware static interaction feature extraction, which specifically includes:

[0074] Time-aware static interaction information refers to the interaction information present at each specific timestamp. The static interaction information at different timestamps is different. To better understand the diversity of interaction information at each timestamp, in step S21, a constant time encoding is applied to the specific timestamp t:

[0075] TE(t)=[cos(ω1t+φ1),…,cos(ω d t+φ d )]

[0076] where ω * and φ * represents the learnable parameter that controls the temporal encoding, and d represents the dimension of the temporal encoding embedding TE(t).

[0077] S22. In the encoding stage, the i-th participant is taken as the central node, the surrounding participants are taken as neighbor nodes, and j represents any neighbor node. Multiple MLP blocks are used to encode the state information of the central node i, the state information of the neighbor node j, and the relative positions between the nodes. The encoding of the relative position preserves the spatial relationship between the central node and the neighbor nodes, and the encoding is obtained as:

[0078]

[0079] in, Represents the encoding of the state information of the central node i, Represents the encoding of the state information of neighbor node j, The code representing the relative position between nodes i and j.

[0080] Considering that in real traffic scenarios, in order to avoid collisions, the central node needs to allocate different attention levels to neighboring nodes. For example, higher attention needs to be paid to nodes that are closer to itself.

[0081] S23. The present invention assigns different weights to each neighbor node by the following method:

[0082]

[0083] in Represents the transformation matrix of the central node, neighbor nodes and their relative positions, N i represents the set of neighbor nodes around the central node i, || represents the vector splicing operation, Represents the calculated weight of the neighbor node at any specific time stamp t.

[0084] S24, according to the obtained neighbor node weight at the specific time stamp t Aggregate all neighbor node information to obtain the static interaction features at a specific time stamp t within the historical time period:

[0085] This feature is closely related to the current timestamp. The static interaction features obtained at different timestamps are different, so it is called time-aware static interaction feature.

[0086] Evolutionary dynamic interactive feature extraction, which specifically includes:

[0087] Although the static interaction features at each specific time stamp are different, the interaction features in consecutive time periods evolve gradually, i.e. is Evolved.

[0088] Therefore, the present invention models the dynamic interaction evolution process occurring at consecutive timestamps, aiming to mine the geometric relationship between link frames when messaging across timestamps.

[0089] In order to deeply explore this relationship, the present invention adopts two time windows of different sizes to analyze the interaction evolution process within the entire historical time period [1:T]: a global time window focusing on long-term dependencies and a local time window focusing on short-term dependencies.

[0090] S25. Using a global time window can focus on all the encoding history states before any current timestamp t∈[1:T]. Assuming that the long-term dependency feature captured at the previous moment t-1 is Then the long-term dependency features captured at the current moment are:

[0091]

[0092] in Represents time series models, such as simple LSTM, RNN, and Transformer.

[0093] S26, the local time window mainly focuses on the historical status of two adjacent timestamps, and comprehensively considers the long-term dependency features obtained at the previous moment. Assume that the short-term dependency features captured at the previous moment t-1 are Long-term dependence is characterized by The present invention adopts a custom gating mechanism to control the weights of features at different moments in the local time window, so the short-term dependency feature at the current moment is:

[0094]

[0095] Where ⊙ represents the element-by-element multiplication operation, A learnable matrix representing the linear mapping between hidden layers, Represents the time-aware gating matrix, which is designed to implement a custom gating mechanism and is calculated from the short-term dependency features of the previous moment:

[0096]

[0097] in and are all linear mapping learnable parameters in the execution of the gating mechanism, and σ(·) is the sigmoid function.

[0098] S3. Aggregate the extracted static interaction features of a specific timestamp and the evolving dynamic interaction features across timestamps to obtain a global interaction representation, which specifically includes:

[0099] Because the data processing limits the range of neighboring participants to be considered around the central participant, some distant participants are overlooked. However, in real traffic scenarios, this simplistic approach of selecting neighbor nodes based on distance not only leads to overlooking some key participants but also makes it difficult for the model to infer new interaction patterns between participants without prior knowledge.

[0100] To alleviate this problem, the present invention proposes a scope-independent interaction feature aggregation module, which aims to enhance the model's understanding of complex interactions between participants and has dynamic adaptability and autonomous learning capabilities.

[0101] Specifically, the present invention creates a fully connected graph structure to aggregate all interactions obtained in the previous step and provide initialized learnable vectors for those participants that are ignored within a bounded range. Subsequently, dot product attention calculation is performed on the obtained fully connected graph, enabling a deeper understanding of the complex interactions between participants in a dynamic environment. The features of the central node are converted into a query vector, and the features of the neighboring nodes are converted into keys and values:

[0102]

[0103] in Represent the linear mapping matrices of the three components respectively.

[0104] It is worth noting that the interaction feature aggregation module aims to aggregate all interaction information in the historical time period [1:T], so the features at the last timestamp T are used when constructing the fully connected graph. Subsequently, the three vectors obtained are used as input to calculate the global interaction representation:

[0105]

[0106] The final output interaction representation vector S is the global interaction representation, which aggregates the potential influences and dependencies between all participants in the scene.

[0107] S4. Use the global interactive representation to predict the participant's driving trajectory at future moments, that is, output multiple possible predicted trajectories.

[0108] In the prediction phase, to facilitate modeling of the time-varying components of the features, the present invention integrates the previously introduced time encoding TE(t). However, since the prediction is for the state information T′ time steps into the future, the timestamp t∈[T:T+T′] in the prediction phase is t∈[T:T+T′]. Given a future timestamp t, the predicted representation at that timestamp can be expressed as:

[0109]

[0110] To account for multiple possible trends in future motion, this paper uses a mixture model consisting of K components to represent the distribution of potential trajectories. Each component k∈K represents a different motion pattern, and all patterns follow a Laplace distribution. The input to the prediction phase is the predicted representation S′ of time perception obtained in this step. The prediction function is used to obtain the possible future trajectories of the participant and the probability of each trajectory:

[0111] Y=predict({S′ i t} i∈N,t∈[T:T+T′] )

[0112] Output predicted trajectory predict(·) represents a prediction function, which can be a simple multi-layer perceptron. In addition, the present invention also uses another multi-layer perceptron with different parameters to generate the mixing coefficient of its mixed model for each road participant. Simply put, the mixing coefficient represents the likelihood of each trajectory.

[0113] Calculate the loss of training parameters

[0114] Predicting the future trajectory of road participants can be regarded as a regression problem, while selecting the best trajectory from the generated K trajectories can be regarded as a classification problem. The present invention adopts the widely used negative log-likelihood function to calculate the error between the predicted trajectory and the true trajectory, as shown below:

[0115]

[0116] where P(·|·) is the probability density function of the Laplace distribution, which is determined by the position of the best predicted trajectory of participant i and the uncertainty row parameter Jointly determined.

[0117] In addition, selecting the best trajectory from the generated K trajectories can be regarded as a classification problem, where the classification loss is The present invention uses cross entropy to optimize the mixing coefficient M.

[0118] Based on the above analysis, the final loss function is the sum of regression loss and classification loss, and the two have equal weights:

[0119]

[0120] Another embodiment of the present invention provides a computer device (computer, server, smart phone, etc.), which includes a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing each step in the method of the present invention.

[0121] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, magnetic disk, optical disk), wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the steps of the method of the present invention are implemented.

[0122] The specific embodiments of the present invention disclosed above are intended to facilitate understanding and implementation of the present invention. Those skilled in the art will appreciate that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the embodiments disclosed in this specification; the scope of protection of the present invention shall be determined by the scope defined in the claims.

Claims

1. An evolutionary time-aware trajectory prediction method based on a neural network, characterized in that: The following steps are involved: S1. Obtain the trajectory data of the traffic scene and divide it into data sets; S2, using neural networks to extract static interaction features at specific timestamps and evolving dynamic interaction features across timestamps from trajectory data of traffic scenes; S3. Aggregate the extracted static interaction features of a specific timestamp and the evolving dynamic interaction features across timestamps to obtain a global interaction representation. S4. Use the global interaction representation to predict the participants’ driving trajectories in the future.

2. The method according to claim 1, characterized in that The acquisition of the trajectory data of the traffic scene in S1 specifically includes: S11. Trajectory sequence formalization: Define the scope of the surrounding environment of the autonomous vehicle. Assume that there are N participants within the scope of the autonomous vehicle, where the state of the i-th participant at a specific time stamp t in the historical time period includes its spatial position Yaw angle and speed t∈[1:T]; the road segment ξ is represented in vector form as in Represents the starting coordinate point of the road segment. Represents the end coordinate point of the road section; given the state information and road section information of the road participants in the historical observation window of T time steps, the historical trajectory of the i-th participant is expressed as in Indicates spatial location The x-axis value of the point, Indicates spatial location The y-axis value of the point; S12. Dataset division: Divide the trajectory sequence data after formal processing into three subsets: training set, validation set, and test set.

3. The method according to claim 2, characterized in that The extracting of the static interaction feature of the specific timestamp specifically includes: S21. Use constant time encoding for a specific timestamp t: TE(t)=[cos(ω1t+φ1),…,cos(ω d t+φ d )] where ω * and φ * represents the learnable parameter that controls the temporal encoding, and d represents the dimension of the temporal encoding embedding TE(t); S22. In the encoding stage, the i-th participant is taken as the central node, the surrounding participants are taken as neighbor nodes, and j represents any neighbor node. Multiple MLP blocks are used to encode the state information of the central node i, the state information of the neighbor node j, and the relative positions between the nodes. The encoding of the relative position preserves the spatial relationship between the central node and the neighbor nodes, and the encoding is obtained as: in, Represents the encoding of the state information of the central node i, Represents the encoding of the state information of neighbor node j, The code representing the relative position between nodes i and j; S23. Assign different weights to each neighbor node using the following method: Where W i ,W j ,W ij Represents the transformation matrix of the central node, neighbor nodes and their relative positions, N i represents the set of neighbor nodes around the central node i, || represents the vector splicing operation, Represents the calculated weight of the neighbor node at any specific time stamp t; S24, according to the obtained neighbor node weight at the specific time stamp t Aggregate all neighbor node information to obtain the static interaction features at a specific time stamp t within the historical time period:

4. The method according to claim 3, characterized in that Extracting the cross-timestamp evolutionary dynamic interaction features specifically includes: S25. Use a global time window to focus on all the encoding history states before the current specific timestamp t∈[1:T]. Assume that the long-term dependency feature captured at the previous moment t-1 is The long-term dependency features captured at the current moment are: in represents a time series model; S26, using a local time window to focus on the historical state of two adjacent timestamps, and comprehensively considering the long-term dependency features obtained at the previous moment, assuming that the short-term dependency features captured at the previous moment t-1 are Long-term dependence is characterized by A custom gating mechanism is used to control the weights of features at different moments in the local time window. The short-term dependency features at the current moment are: Where ⊙ represents the element-by-element multiplication operation, W g A learnable matrix representing the linear mapping between hidden layers, Represents the time-aware gating matrix, which is designed to implement a custom gating mechanism and is calculated from the short-term dependency features of the previous moment: Where W1 and b1 are linear mapping learnable parameters in the gating mechanism execution process, and σ(·) is the sigmoid function.

5. The method according to claim 4, characterized in that In S3, the extracted static interaction features of a specific timestamp and the evolving dynamic interaction features across timestamps are aggregated to obtain a global interaction representation, which specifically includes: All static interaction features and evolving dynamic interaction features are aggregated into a fully connected graph, and initialized learnable vectors are provided for participants that are ignored within a bounded range. Dot product attention is then performed on the resulting fully connected graph, where the features of the central node are converted into a query vector, and the features of the neighboring nodes are converted into keys and values: Where W q ,W k ,W v Represents the linear mapping matrix of the three components respectively; when constructing the fully connected graph, the features of the last timestamp T are used, and then the three vectors obtained are used as input to calculate the global interaction representation: The final output vector S is the global interaction representation.

6. The method according to claim 5, characterized in that The method of using the global interactive representation to predict the participant's driving trajectory at a future moment includes: Assume that the state information of the next T′ time steps is predicted, that is, the timestamp t∈[T:T+T′] of the prediction stage. Given a future timestamp t, the prediction representation at this timestamp is expressed as follows: S′=[S||TE(t)] Input the predicted representation S′ and pass it through the prediction function to obtain the possible future trajectory of the participant: Y=predict({S′ i t } i∈N , t∈[T:T+T′] ) Where Y represents the predicted trajectory of the output, predict(·) represents the prediction function, and S′ i t represents the predicted representation at the timestamp t.

7. The method according to claim 6, characterized in that Predicting the future trajectory of road participants is considered a regression problem, and selecting the best trajectory from the generated K trajectories is considered a classification problem. The negative log-likelihood function is used to calculate the error between the predicted trajectory and the true trajectory: where P(·|·) is the probability density function of the Laplace distribution, which is determined by the position of the best predicted trajectory of participant i and the uncertainty parameter Jointly determined; In addition, for the classification loss Cross entropy is used to optimize the mixing coefficient M; the final loss function is the sum of regression loss and classification loss, and both are equally weighted:

8. An evolutionary time-aware trajectory prediction device based on a neural network, characterized in that: include: The trajectory data processing module is used to cut the raw trajectory data collected from the traffic scene according to the timestamp and standardize it into formal input, and then divide the formal trajectory sequence into different data sets; A static interaction feature extraction module is used to extract static interaction features of specific timestamps from the trajectory data of traffic scenes using a neural network; An evolutionary dynamic interaction feature extraction module, which uses a neural network to extract cross-timestamp evolutionary dynamic interaction features from the trajectory data of traffic scenes; The interaction feature aggregation module is used to aggregate the extracted static interaction features of a specific timestamp and the evolving dynamic interaction features across timestamps to obtain a global interaction representation; The trajectory prediction module is used to predict the driving trajectory of participants in the future by using the global interaction representation.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a computer, the method according to any one of claims 1 to 7 is implemented.

Images