A robust trajectory prediction method and system based on decoupled spatiotemporal network

Abstract

This invention discloses a robust trajectory prediction method and system based on a decoupled spatiotemporal network. The method comprises: obtaining and preprocessing a vehicle trajectory dataset; constructing a decoupled spatiotemporal network model; and inputting the preprocessed vehicle trajectory dataset into the decoupled spatiotemporal network model to obtain future vehicle trajectory signals. The method provides reliable prediction results regardless of whether the historical signal is disturbed or not.

Description

Technical Field

[0001] The present invention belongs to the field of spatiotemporal data mining in intelligent transportation systems (ITS), and specifically relates to a robust trajectory prediction method and system based on a decoupling spatiotemporal network (DSTNet). Background Art

[0002] ITS is a comprehensive application that combines advanced control, sensing, communication, information, and computing technologies, dedicated to providing innovative services for traffic management. As a cornerstone of ITS, topological traffic signal prediction aims to leverage historical traffic signals to predict future traffic signals, enabling comprehensive perception and precise control of the traffic network, thereby reducing traffic congestion. Traffic signals refer to data collected by all sensors (nodes) within a time window (such as vehicle trajectories, flow, speed, lane occupancy, etc.). In real-world ITS, traffic signals are inevitably disturbed due to sensor or data management platform failures. Topological traffic signal prediction methods based on deep learning typically learn the spatiotemporal dependencies (i.e., spatiotemporal correlations) of traffic signals to aggregate their spatiotemporal features and achieve traffic prediction. If a node in a traffic signal is disturbed, the disturbance will propagate to the remaining nodes when spatiotemporal features are aggregated, following the spatiotemporal dependencies. As the number of feature aggregations increases, the scope of the disturbance gradually increases, eventually spreading to traffic signals across the entire road network, significantly impacting traffic predictions. However, most current traffic prediction methods assume that historical signals are undisturbed. How to reduce the impact of disturbances in traffic signals and build a robust traffic prediction method is a practical problem that needs to be solved urgently. Summary of the Invention

[0003] To address the issue of historical traffic signals being subject to disturbances, the present invention proposes a robust trajectory prediction method and system based on a decoupled spatiotemporal network, which can provide reliable prediction results regardless of whether the historical signals are disturbed or not. The traffic signals in the present invention are vehicle trajectories.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A robust trajectory prediction method based on a decoupled spatiotemporal network includes the following steps:

[0006] Obtain vehicle trajectory dataset and preprocess it;

[0007] Construct a decoupled spatiotemporal network model;

[0008] The preprocessed vehicle trajectory data set is input into the decoupled spatiotemporal network model to obtain a future signal of the vehicle trajectory.

[0009] Preferably, the method for preprocessing the vehicle trajectory dataset includes:

[0010] The vehicle trajectory dataset is normalized by Z-Score to remove the mean and scale to unit variance. The calculation formula is:

[0011]

[0012] in, Refers to input; Used for averaging; Used to obtain standard deviation.

[0013] Preferably, the decoupled spatiotemporal network model comprises: an embedding layer, a plurality of decoupled spatiotemporal modules and a de-embedding layer;

[0014] The decoupled spatiotemporal module includes: an invariant block, a decoupling block and a variable block;

[0015] The embedding layer is used to randomly sample a set of data from the preprocessed vehicle trajectory dataset in a no-replacement manner, wherein the set of data includes: historical traffic signals and future traffic signals, and map the historical traffic signals into high-dimensional features;

[0016] The invariant block is used to generate invariant components of historical or future signals based on the timestamp, date stamp and node index of the traffic signal;

[0017] The decoupling block is used to generate a variable component by removing the invariant component from the historical signal;

[0018] The variable block is used to map the variable components of the historical signal into the variable components of the future signal;

[0019] The de-embedding layer is used to restore and fuse the invariant components of the future signal and the variable components of the future signal to obtain the future signal of the vehicle trajectory.

[0020] Preferably, the method for generating the invariant component of a historical or future signal based on the timestamp, date stamp, and node index of the traffic signal includes:

[0021] Utilize the identity ID index of historical signals Generate the invariant component of the historical signal ,in, 、 and Represent the timestamp, date stamp and node index of the historical signal respectively;

[0022] Leveraging identity indexing of future signals Generate the invariant components of future signals ,in, 、 and Represent the timestamp, date stamp and node index of the future signal respectively;

[0023] The invariant components generated by the invariant blocks in each layer of the decoupled spatiotemporal module are fused, i.e. ;

[0024] in, and The calculation method is:

[0025]

[0026]

[0027] Where m is the layer index of DSTM; 、 and They represent the timestamp ID library, date stamp ID library and node ID library respectively, which are weight parameters; 、 and are generators for timestamp embedding, date stamp embedding, and node embedding, respectively. Represents the gated linear unit GLU, which is used to perform nonlinear transformation on features.

[0028] Preferably, the method of generating a variable component by removing the invariant component from the historical signal comprises:

[0029] The invariant component of the historical signal From historical signals Remove and obtain the variable components of the historical signal , the formula is:

[0030] .

[0031] Preferably, the method of mapping the variable components of the historical signal into the variable components of the future signal includes:

[0032] The variable components of the historical signal Perform spatial feature aggregation to generate spatial features , where N represents the number of nodes;

[0033] The spatial features Input the backtracking and prediction branches respectively to obtain the variable components of the historical signal and the variable components of the future signal , the formula is:

[0034]

[0035]

[0036] in, and It is the weight matrix in the time dimension; T,1 represents the time dimension of looking back to history; T,2 represents the time dimension of predicting the future;

[0037] The variable signals generated by the variable blocks in each layer of decoupled spatiotemporal module are fused, i.e. .

[0038] Preferably, the method of performing dimensionality reduction and fusion of the invariant component of the future signal and the variable component of the future signal to obtain the future signal of the vehicle trajectory includes:

[0039] Generate constant blocks and variable blocks separately and Input them into the de-embedding layer for dimensionality restoration, and we get and ;

[0040] Fusion of the invariant components of future signals and variable components , get the future signal .

[0041] The present invention also discloses a robust trajectory prediction system based on a decoupled spatiotemporal network, comprising: a preprocessing module, a construction module and a prediction module;

[0042] The preprocessing module is used to obtain the vehicle trajectory data set and perform preprocessing;

[0043] The building module is used to build a decoupled spatiotemporal network model;

[0044] The prediction module is used to input the preprocessed vehicle trajectory data set into the decoupled spatiotemporal network model to obtain a future signal of the vehicle trajectory.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention designs an invariant block, which only requires the timestamp, date stamp and node index of the signal to generate the invariant components of the historical and future signals without inputting the historical signal.

[0047] The present invention designs a decoupling block to generate a variable component of the historical signal by removing the invariant component from the historical signal.

[0048] The present invention designs a variable block to map the variable components of the historical signal into the variable components of the future signal. The robust spatial layer changes the distribution of spatial dependencies to make them more uniform, preventing nodes from over-depending on some nodes that may be disturbed.

[0049] The present invention designs constant loss and decoupling loss, thereby increasing the constant component and reducing the proportion of the variable component in the future signal, thereby improving the stability of the future signal.

[0050] The present invention designs a smoothing loss to alleviate the influence of disturbance on the variable component in the future signal by changing the smoothness of the variable component. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1 Schematic diagram of a decoupled spatiotemporal network model according to an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of a decoupled spatiotemporal module according to an embodiment of the present invention;

[0054] Figure 3 Schematic diagram of a flow chart of a robust trajectory prediction method based on a decoupled spatiotemporal network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1

[0058] like Figure 3 As shown, the present invention provides a robust trajectory prediction method based on a decoupled spatiotemporal network, comprising the following steps:

[0059] Obtain vehicle trajectory dataset and preprocess it;

[0060] Construct a decoupled spatiotemporal network model;

[0061] The preprocessed vehicle trajectory data set is input into the decoupled spatiotemporal network model to obtain a future signal of the vehicle trajectory.

[0062] In this embodiment, the method for preprocessing the vehicle trajectory dataset includes:

[0063] The dataset of the present invention is divided into training set, validation set and test set in a ratio of 6:2:2. The dataset is normalized by Z-Score to remove the mean and scale to unit variance. The calculation formula is as follows:

[0064] (1)

[0065] in, Refers to input; Used for averaging; Used to obtain standard deviation.

[0066] In this embodiment, the decoupled spatiotemporal network model includes: an embedding layer, several decoupled spatiotemporal modules, and a de-embedding layer;

[0067] Among them, the decoupled spatiotemporal module includes: an invariant block, a decoupling block and a variable block;

[0068] The embedding layer is used to randomly sample a set of data from the preprocessed vehicle trajectory dataset in a non-replacement manner, wherein the set of data includes: historical traffic signals and future traffic signals, and map the historical traffic signals into high-dimensional features;

[0069] The invariant block is used to generate the invariant components of historical or future signals based on the timestamp, date stamp and node index of the traffic signal;

[0070] The decoupling block is used to generate variable components by removing the invariant components from the historical signal;

[0071] The variable block is used to map the variable components of the historical signal into the variable components of the future signal;

[0072] The de-embedding layer is used to restore and fuse the invariant components of the future signal and the variable components of the future signal to obtain the future signal of the vehicle trajectory.

[0073] Specifically, the embedding layer randomly samples a set of data from the processed original data without replacement. and , which represent historical traffic signals and future traffic signals respectively. Figure 1As shown in Figure 2, the embedding layer is the input layer of the decoupled spatiotemporal network model and is a fully connected layer, in which the Rectified Linear Unit (ReLU) is used as the activation function. The purpose of the input layer is to transform the historical signal Mapping into high-dimensional features ,in, It is both the output of the embedding layer and the input of the first layer of DSTM. The calculation formula is as follows:

[0074] (2)

[0075] in, and Represents the weight parameters and bias of the fully connected layer; Refers to the input. ReLU is calculated as follows:

[0076] (3)

[0077] in, refers to the input. Then the calculation method of the embedding layer can be expressed as:

[0078] (4)

[0079] Decoupled space-time modules: Figure 1 As shown in Figure 1, M layers of Decoupling Spatiotemporal Modules (DSTMs) are stacked after the embedding layer. Each DSTM contains an invariant block, a decoupling block, and a variable block.

[0080] Invariant block: Figure 2 As shown, the invariant block aims to utilize the identity (ID) index of the historical signal Generate the invariant component of the historical signal ,in, 、 and Represent the timestamp, date stamp and node index of the historical signal respectively; use the identity index of the future signal Generate the invariant components of future signals ,in, 、 and Represents the timestamp, date stamp and node index of the future signal respectively. and is calculated as follows:

[0081] (5)

[0082] (6)

[0083] Where m is the layer index of DSTM; 、 and They represent the timestamp ID library, date stamp ID library and node ID library respectively, which are weight parameters; 、 and Represents the generators of timestamp embedding, date stamp embedding and node embedding respectively. Each generator They are all composed of two layers of fully connected layers, aiming to map IDs into embeddings. The second layer of the generator has an activation function ReLU, calculated as follows:

[0084] (7)

[0085] in, Refers to ID; Represents the Gated Linear Unit (GLU), which is used to perform nonlinear transformation on features. The calculation formula is as follows:

[0086] (8)

[0087] in, and represents the weight parameter; and express deviation; Refers to input; represents the Hadamard product; Represents the activation function, and the calculation formula is as follows:

[0088] (9)

[0089] in, Refers to the input. Finally, the invariant signals generated by the invariant blocks in each layer of DSTM are fused, that is, .

[0090] Decoupling block: Figure 2 As shown, the decoupling block converts the invariant components of the historical signal into From historical signals Remove the variable component to obtain the historical signal (i.e., the decoupling operation in the figure), the formula is as follows:

[0091] (10)

[0092] In the subsequent variable blocks, the variable components of the historical signal Will generate itself through autoregression . and The residual between The part that is considered to be unable to be fitted by the DSTM of the current layer (i.e., the residual operation in the figure) will be input into the DSTM of the next layer for further fitting.

[0093] Variable blocks: such as Figure 2 As shown, the variable block converts the variable components of the historical signal into Perform spatial feature aggregation to generate spatial features , N represents the number of nodes, and the formula is as follows:

[0094] , (11)

[0095] in, Represents the set of associated nodes of node t. In each forward propagation of the model, the associated nodes of node t are reselected. In this way, the nodes that need to perform feature aggregation will not be overly dependent on certain nodes, which may be disturbed. Represents the characteristics of node t in the variable component of the historical signal; Represents the characteristics of node t in the spatial feature; Indicates the number of nodes in the set; represents the node selection rate; Indicates the dependency of node t on node s. The calculation formula is as follows:

[0096] (12)

[0097] in, is the weight parameter. Then, the variable block transforms the spatial features Input the backtracking and prediction branches respectively to obtain the variable components of the historical signal and the variable components of the future signal , the formula is as follows:

[0098] (13)

[0099] (14)

[0100] in, and is the weight matrix in the time dimension; T,1 represents the time dimension of looking back to the past; T,2 represents the time dimension of predicting the future. Finally, the variable signals generated by the variable blocks in each layer of DSTM are fused, that is, .

[0101] Ingredient fusion: Figure 1 As shown, the estimated values of the invariant components of the future signal are generated by the invariant block and the variable block respectively. and an estimate of the variable component of the future signal , input them into the de-embedding layer for dimensionality restoration, and we can get and , where the de-embedding layer is a linear fully connected layer. Fusion of the invariant components of the future signal and variable components , you can get the future signal ,in .

[0102] In this embodiment, the present invention designs the following loss function:

[0103] (15)

[0104] in, and represents the loss weight; Represents the invariant loss, which is used to measure the distance between the invariant component of the future signal and the future signal. The formula is as follows:

[0105] (16)

[0106] in, represents the true value of the future signal; Represents Huber loss, which is used to measure the distance between two variables. The calculation formula is as follows:

[0107] (17)

[0108] in, and Refers to input. Represents the threshold used to switch between the two losses. It represents the decoupling loss, which is used to measure the distance between the invariant component of the future signal and the residual of the future signal and the variable component of the future signal. The formula is as follows:

[0109] (18)

[0110] Considering that the generation of the invariant component does not rely on the historical signal, when the historical signal is disturbed, the invariant component of the future signal will not be affected. However, the generation of the variable component of the future signal requires the variable component of the historical signal, so it will be affected by the disturbance. Therefore, the present invention adjusts Increase the proportion of the constant component in the future signal, while reducing the proportion of the variable component in the future signal, thereby improving the stability of the prediction results. It is used to measure the distance between the variable components of the history graph signal matrix when it is not disturbed and when it is disturbed. It is expressed as follows:

[0111] (19)

[0112] in, It is through The perturbation is obtained from the standard normal distribution. The samples are sampled and discarded according to a certain probability. The discard probability is from a uniform distribution. Unlike data augmentation methods, this method does not increase the number of training samples. and The calculation process is the same, the only difference is that their inputs are and .

[0113] Prediction data restoration: During data preprocessing, in order to improve the training effect of the model, the present invention uses Z-Score to preprocess the original data. Therefore, when the model completes the prediction, it is necessary to restore the predicted data through the inverse operation of Z-Score. The calculation formula is as follows:

[0114] (20)

[0115] in, Refers to data to be processed.

[0116] Model settings: The model settings are as follows

[0117] 1) The model optimizer uses Adaptive Moment (Adam), with a learning rate of 1e-2, 200 training iterations, and a training batch size of 32.

[0118] 2) The number of features in the embedding layer is 1, the number of features in the de-embedding layer is 1, and the number of features in the remaining layers is 64. The number of features is the number of neurons in the neural network.

[0119] 3) The length of the historical signal time window is 12, and the length of the future signal time window is 12.

[0120] 4) The number of sensors (nodes) N, the number of sampling timestamps and the number of date stamps have different values in different datasets.

[0121] 5) Number of DSTM layers .

[0122] 6) During the training phase, the node selection rate ; During the testing phase, the node selection rate .

[0123] 7) Weight loss , .

[0124] Evaluation indicators: In order to verify the effectiveness of the present invention, three evaluation indicators including Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) are adopted. Their calculation formulas are as follows

[0125] (twenty one)

[0126] (twenty two)

[0127] (twenty three)

[0128] in, A collection of signal matrices representing historical and future graphs; represents the number of samples; express 's index; Indicates the A historical / future graph signal matrix.

[0129] Example 2

[0130] The present invention also discloses a robust trajectory prediction system based on a decoupled spatiotemporal network, comprising: a preprocessing module, a construction module and a prediction module;

[0131] The preprocessing module is used to obtain the vehicle trajectory dataset and perform preprocessing;

[0132] The building module is used to construct a decoupled spatiotemporal network model;

[0133] The prediction module is used to input the preprocessed vehicle trajectory data set into the decoupled spatiotemporal network model to obtain the future signal of the vehicle trajectory.

[0134] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A robust trajectory prediction method based on decoupled spatiotemporal networks, characterized by: The following steps are involved: Obtain vehicle trajectory dataset and preprocess it; Construct a decoupled spatiotemporal network model; Inputting the preprocessed vehicle trajectory data set into the decoupled spatiotemporal network model to obtain a future signal of the vehicle trajectory; Methods for preprocessing vehicle trajectory datasets include: The vehicle trajectory dataset is normalized by Z-Score to remove the mean and scale to unit variance. The calculation formula is: in, Refers to input; Used for averaging; Used to obtain standard deviation; The decoupled spatiotemporal network model includes: an embedding layer, several decoupled spatiotemporal modules and a de-embedding layer; The decoupled spatiotemporal module includes: an invariant block, a decoupling block and a variable block; The embedding layer is used to randomly sample a set of data from the preprocessed vehicle trajectory dataset in a no-replacement manner, wherein the set of data includes: historical traffic signals and future traffic signals, and map the historical traffic signals into high-dimensional features; The invariant block is used to generate invariant components of historical or future signals based on the timestamp, date stamp and node index of the traffic signal; The decoupling block is used to generate a variable component by removing the invariant component from the historical signal; The variable block is used to map the variable components of the historical signal into the variable components of the future signal; The de-embedding layer is used to restore and fuse the invariant components of the future signal and the variable components of the future signal to obtain the future signal of the vehicle trajectory.

2. The robust trajectory prediction method based on decoupled spatiotemporal network according to claim 1, characterized in that Methods for generating invariant components of historical or future signals based on the timestamp, date stamp, and node index of traffic signals include: Utilize the identity ID index of historical signals Generate the invariant component of the historical signal ,in, 、 and Represent the timestamp, date stamp and node index of the historical signal respectively; Leveraging identity indexing of future signals Generate the invariant components of future signals ,in, 、 and Represent the timestamp, date stamp and node index of the future signal respectively; The invariant components generated by the invariant blocks in each layer of the decoupled spatiotemporal module are fused, i.e. ; in, and The calculation method is: Where m is the layer index of DSTM; 、 and They represent the timestamp ID library, date stamp ID library and node ID library respectively, which are weight parameters; 、 and Generators representing timestamp embedding, date stamp embedding, and node embedding respectively; Represents the gated linear unit GLU, which is used to perform nonlinear transformation on features.

3. The robust trajectory prediction method based on decoupled spatiotemporal network according to claim 2, characterized in that: Methods for generating variable components by removing the invariant components from the historical signal include: The invariant component of the historical signal From historical signals Remove and obtain the variable components of the historical signal , the formula is: 。 4. The robust trajectory prediction method based on decoupled spatiotemporal network according to claim 3, characterized in that Methods for mapping variable components of historical signals into variable components of future signals include: The variable components of the historical signal Perform spatial feature aggregation to generate spatial features , where N represents the number of nodes; The spatial features Input the backtracking and prediction branches respectively to obtain the variable components of the historical signal and the variable components of the future signal , the formula is: in, and It is the weight matrix in the time dimension; T,1 represents the time dimension of looking back to history; T,2 represents the time dimension of predicting the future; The variable signals generated by the variable blocks in each layer of decoupled spatiotemporal module are fused, i.e. .

5. The robust trajectory prediction method based on decoupled spatiotemporal network according to claim 4, characterized in that: The method of restoring and fusing the constant components of the future signal and the variable components of the future signal to obtain the future signal of the vehicle trajectory includes: Generate constant blocks and variable blocks separately and Input them into the de-embedding layer for dimensionality restoration, and we get and ; Fusion of the invariant components of future signals and variable components , get the future signal .

6. A robust trajectory prediction system based on a decoupled spatiotemporal network, the system being used to implement the robust trajectory prediction method based on a decoupled spatiotemporal network according to any one of claims 1 to 5, characterized in that: include: Preprocessing module, building module and prediction module; The preprocessing module is used to obtain the vehicle trajectory data set and perform preprocessing; The building module is used to build a decoupled spatiotemporal network model; The prediction module is used to input the preprocessed vehicle trajectory data set into the decoupled spatiotemporal network model to obtain a future signal of the vehicle trajectory.

Images