Vehicle track prediction method and device based on road ETC data

Abstract

The invention discloses a vehicle track prediction method and device based on road ETC data, and relates to the technical field of intelligent traffic data analysis. According to the invention, based on the Word2Vec model, the ETC gantry sequence in each vehicle historical trajectory in the ETC data is trained, and the embedded vector is generated, so that the time sequence and space information can be processed at the same time in the construction process of the embedded vector, and the ETC data and the relationship between the ETC data and the vehicle operation are fully mined. Meanwhile, the multi-layer bidirectional LSTM model can capture forward and backward dependencies in a vehicle driving track at the same time by combining a reverse information flow, and the learning ability of a complex dynamic mode in long time sequence data is remarkably enhanced. Finally, the accuracy of the vehicle track prediction method based on the road ETC data in the embodiment of the invention is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent traffic data analysis, and in particular to a vehicle trajectory prediction method and device based on road ETC data. Background Art

[0002] Vehicle trajectory prediction is of great significance for intelligent road management. Using ETC data for vehicle trajectory prediction is also an important direction of vehicle trajectory prediction. The ETC system records vehicle traffic data through the gantry, including vehicle ID, gantry ID, travel time and other information. These data provide a rich source of information for the analysis and prediction of vehicle driving trajectories. However, how to extract useful information from massive ETC data and build an efficient vehicle trajectory prediction model is still a challenging problem.

[0003] The existing technology mainly focuses on modeling the time dimension in time series prediction. For example, China's invention patent CN115565376A discloses a vehicle travel time prediction method and system that integrates graph2vec and double-layer LSTM. Practice shows that the challenge of path prediction lies in the fact that vehicle movement is affected by a variety of complex factors, especially the interaction between internal and external stimulus factors. The vehicle prediction model in the existing technology fails to fully explore the ETC data and the relationship between ETC data and vehicle operation, resulting in inaccurate vehicle trajectory prediction results. Summary of the invention

[0004] The present invention provides a vehicle trajectory prediction method and device based on road ETC data, which are used to improve the accuracy of vehicle trajectory prediction results.

[0005] In order to solve the above technical problems, the first aspect of the present invention discloses a vehicle trajectory prediction method based on road ETC data, the method comprising: Obtain all ETC data corresponding to the road system, wherein the ETC data includes record information of each vehicle passing through each ETC gantry; Based on the Word2Vec model, the ETC gantry sequence in each vehicle historical trajectory in the ETC data is trained to generate an embedding vector, wherein the embedding vector is used to represent the relative position of the ETC gantry and the traffic flow pattern; The embedded vector is trained based on a multi-layer bidirectional LSTM model to obtain a trained multi-layer bidirectional LSTM model, wherein the trained multi-layer bidirectional LSTM model is used to predict the vehicle trajectory; The multi-layer bidirectional LSTM model includes two alternately connected bidirectional LSTM layers to capture the forward and backward dependencies in the vehicle trajectory.

[0006] As an optional implementation, in the first aspect of the present invention, the embedding vector is trained based on the multi-layer bidirectional LSTM model to obtain a trained multi-layer bidirectional LSTM model, including: The embedding vector is trained based on a multi-layer bidirectional LSTM model. Through the attention mechanism, the weight of each time step is dynamically adjusted according to the importance of the input data, so as to match the most relevant vehicle historical trajectory for each prediction and obtain a trained multi-layer bidirectional LSTM model.

[0007] As an optional implementation, in the first aspect of the present invention, the obtaining of all ETC data corresponding to the road system includes: Obtain all original ETC data corresponding to the road system, and for any two ETC gantries with a preset association relationship, determine a reasonable range of the travel time between the two ETC gantries based on the road information between the two ETC gantries; For all the original ETC data, after eliminating the erroneous ETC data, all the ETC data corresponding to the road system are obtained; wherein, the erroneous ETC data is the passing record of vehicles between any two ETC gantries with a preset association relationship, whose travel time is not within the reasonable range.

[0008] As an optional implementation, in the first aspect of the present invention, before removing the erroneous ETC data from all the original ETC data to obtain all the ETC data corresponding to the road system, the method further includes: The gantry information error correction, license plate information error correction and trip information error correction operations are performed on all the original ETC data, wherein: The gantry information error correction operation includes: identifying and correcting gantry information errors by comparing the timestamps and geographic locations of historical data with those of current data; The license plate information error correction operation includes: using data cleaning technology to identify abnormal license plate information through data comparison at multiple time points, and correcting the license plate data in combination with the vehicle owner information database; The trip information error correction operation includes: supplementing discontinuous time periods by prediction and interpolation, and recalibrating the trip information in combination with the vehicle's historical trajectory data.

[0009] As an optional implementation, in the first aspect of the present invention, the Word2Vec model is used to train the ETC gantry sequence in each vehicle historical trajectory in the ETC data to generate an embedding vector, including: Set the model input: Set the input vocabulary to ,in is the ID of the target ETC gantry, confirm arrive For vocabulary The surrounding 2k context words; Set the model goal: Set the model's objective to maximize the vocabulary The corresponding contextual vocabulary conditional probability, thereby adjusting the vocabulary The corresponding embedding vector; where the vocabulary The corresponding context word conditional probability is ; Setting the loss function: Set a central word and its context words , the number of negative samples is k, then the loss function is set to:

[0010] In the above formula, are negative samples randomly drawn from the vocabulary, is a given central word When predicting the context word probability; is a given central word When , the randomly selected negative samples probability.

[0011] As an optional implementation, in the first aspect of the present invention, the training of the embedding vector based on the multi-layer bidirectional LSTM model includes: The embedding vector is trained based on a multi-layer bidirectional LSTM model, wherein the output of the multi-layer bidirectional LSTM model at each time step t is:

[0012] In the above formula, and are the hidden states of the forward and reverse LSTM, respectively, where:

[0013] In the above formula, is the input at time step t, and Respectively represent the calculation process of forward and reverse LSTM, and are the hidden states of the previous and next moments corresponding to time step t respectively.

[0014] As an optional implementation, in the first aspect of the present invention, dynamically adjusting the weight of each time step according to the importance of input data through the attention mechanism includes: The attention coefficient corresponding to time step t is calculated according to the following formula:

[0015] In the above formula, is the hidden state at the current time step t The matching score between the hidden states of all historical trajectories is calculated by dot product or weighted sum.

[0016] As an optional implementation, in the first aspect of the present invention, the obtaining of all ETC data corresponding to the road system includes: Acquire all initial ETC data corresponding to the road system, and determine the historical trajectory of each identified vehicle based on the initial ETC data; For each identified vehicle, all ETC gantries corresponding to the identified vehicle are determined according to the initial ETC data, and predictable parameters corresponding to the identified vehicle are determined according to the number and distribution range of all ETC gantries corresponding to the identified vehicle, wherein the predictable parameters are used to measure the accuracy of the prediction of the trajectory of the corresponding identified vehicle; A vehicle to be predicted is selected from all the identified vehicles, and a predictable parameter corresponding to the vehicle to be predicted is greater than or equal to a preset predictable threshold value, and initial ETC data corresponding to all the vehicles to be predicted are determined to be all ETC data corresponding to the road system.

[0017] The second aspect of the present invention discloses a vehicle trajectory prediction device based on road ETC data, the device comprising: An acquisition module, used to acquire all ETC data corresponding to the road system, wherein the ETC data includes record information of each vehicle passing through each ETC gantry; A Word2Vec module, for training the ETC gantry sequence in each vehicle historical trajectory in the ETC data based on the Word2Vec model to generate an embedding vector, wherein the embedding vector is used to represent the relative position of the ETC gantry and the traffic flow pattern; An LSTM module is used to train the embedding vector based on a multi-layer bidirectional LSTM model to obtain a trained multi-layer bidirectional LSTM model, wherein the trained multi-layer bidirectional LSTM model is used to predict the vehicle trajectory; The multi-layer bidirectional LSTM model includes two alternately connected bidirectional LSTM layers to capture the forward and backward dependencies in the vehicle trajectory.

[0018] As an optional implementation, in the second aspect of the present invention, the LSTM module trains the embedding vector based on the multi-layer bidirectional LSTM model to obtain the specific operation of the trained multi-layer bidirectional LSTM model, including: The embedding vector is trained based on a multi-layer bidirectional LSTM model. Through the attention mechanism, the weight of each time step is dynamically adjusted according to the importance of the input data, so as to match the most relevant vehicle historical trajectory for each prediction and obtain a trained multi-layer bidirectional LSTM model.

[0019] As an optional implementation, in the second aspect of the present invention, the specific operation of the acquisition module acquiring all ETC data corresponding to the road system includes: Obtain all original ETC data corresponding to the road system, and for any two ETC gantries with a preset association relationship, determine a reasonable range of the travel time between the two ETC gantries based on the road information between the two ETC gantries; For all the original ETC data, after eliminating the erroneous ETC data, all the ETC data corresponding to the road system are obtained; wherein, the erroneous ETC data is the passing record of vehicles between any two ETC gantries with a preset association relationship, whose travel time is not within the reasonable range.

[0020] As an optional implementation, in the second aspect of the present invention, the device further includes: The correction module is used to perform gantry information error correction, license plate information error correction and trip information error correction operations on all the original ETC data before the acquisition module removes the erroneous ETC data from all the original ETC data to obtain all the ETC data corresponding to the road system, wherein: The gantry information error correction operation includes: identifying and correcting gantry information errors by comparing the timestamps and geographic locations of historical data with those of current data; The license plate information error correction operation includes: using data cleaning technology to identify abnormal license plate information through data comparison at multiple time points, and correcting the license plate data in combination with the vehicle owner information database; The trip information error correction operation includes: supplementing discontinuous time periods by prediction and interpolation, and recalibrating the trip information in combination with the vehicle's historical trajectory data.

[0021] As an optional implementation, in the second aspect of the present invention, the Word2Vec module trains the ETC gantry sequence in each vehicle historical trajectory in the ETC data based on the Word2Vec model, and generates a specific operation of an embedded vector, including: Set the model input: Set the input vocabulary to ,in is the ID of the target ETC gantry, confirm arrive For vocabulary The surrounding 2k context words; Set the model goal: Set the model's objective to maximize the vocabulary The corresponding contextual vocabulary conditional probability, thereby adjusting the vocabulary The corresponding embedding vector; where the vocabulary The corresponding context word conditional probability is ; Setting the loss function: Set a central word and its context words , the number of negative samples is k, then the loss function is set to:

[0022] In the above formula, are negative samples randomly drawn from the vocabulary, is a given central word When predicting the context word The probability of is a given central word When , the randomly selected negative samples probability.

[0023] As an optional implementation, in the second aspect of the present invention, the specific operation of the LSTM module training the embedding vector based on the multi-layer bidirectional LSTM model includes: The embedding vector is trained based on a multi-layer bidirectional LSTM model, wherein the output of the multi-layer bidirectional LSTM model at each time step t is:

[0024] In the above formula, and are the hidden states of the forward and reverse LSTM, respectively, where:

[0025] In the above formula, is the input at time step t, and Respectively represent the calculation process of forward and reverse LSTM, and are the hidden states of the previous and next moments corresponding to time step t respectively.

[0026] As an optional implementation, in the second aspect of the present invention, the specific operation of the LSTM module dynamically adjusting the weight of each time step according to the importance of the input data through the attention mechanism includes: The attention coefficient corresponding to time step t is calculated according to the following formula:

[0027] In the above formula, is the hidden state at the current time step t The matching score between the hidden states of all historical trajectories is calculated by dot product or weighted sum.

[0028] As an optional implementation, in the second aspect of the present invention, the specific operation of the acquisition module acquiring all ETC data corresponding to the road system includes: Acquire all initial ETC data corresponding to the road system, and determine the historical trajectory of each identified vehicle based on the initial ETC data; For each identified vehicle, all ETC gantries corresponding to the identified vehicle are determined according to the initial ETC data, and predictable parameters corresponding to the identified vehicle are determined according to the number and distribution range of all ETC gantries corresponding to the identified vehicle, wherein the predictable parameters are used to measure the accuracy of the prediction of the trajectory of the corresponding identified vehicle; A vehicle to be predicted is selected from all the identified vehicles, and a predictable parameter corresponding to the vehicle to be predicted is greater than or equal to a preset predictable threshold value, and initial ETC data corresponding to all the vehicles to be predicted are determined to be all ETC data corresponding to the road system.

[0029] The third aspect of the present invention discloses a vehicle trajectory prediction system based on road ETC data, the system comprising: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the vehicle trajectory prediction method based on road ETC data disclosed in the first aspect of the present invention.

[0030] Compared with the prior art, the present invention has the following beneficial effects: Based on the Word2Vec model, the present invention trains the ETC gantry sequence in each vehicle historical trajectory in the ETC data to generate an embedding vector, so that the temporal and spatial information can be processed simultaneously in the process of constructing the embedding vector, and the relationship between the ETC data and the vehicle operation is fully mined. At the same time, the multi-layer bidirectional LSTM model of the present invention, by combining the reverse information flow, can simultaneously capture the forward and backward dependencies in the vehicle's driving trajectory, and significantly enhances the learning ability of complex dynamic patterns in long time series data. Ultimately, the vehicle trajectory prediction method based on road ETC data in the embodiment of the present invention improves the accuracy of vehicle trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of 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 creative work.

[0032] Figure 1 It is a flow chart of a vehicle trajectory prediction method based on road ETC data disclosed in an embodiment of the present invention; Figure 2 It is a flowchart of a multi-layer bidirectional LSTM model in an embodiment of the present invention; Figure 3 It is a framework diagram of a multi-layer bidirectional LSTM model integrating a self-attention mechanism disclosed in an embodiment of the present invention; Figure 4 It is a structural schematic diagram of another vehicle trajectory prediction device based on road ETC data disclosed in an embodiment of the present invention; Figure 5 It is a structural schematic diagram of a vehicle trajectory prediction system based on road ETC data disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0034] The terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, device, product or end including a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or ends.

[0035] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0036] Embodiment 1 See also Figure 1 , Figure 1 : is a flow chart of a vehicle trajectory prediction method based on road ETC data disclosed in an embodiment of the present invention. Figure 1 The vehicle trajectory prediction method based on road ETC data described above can be used in a vehicle trajectory prediction device based on road ETC data, and the vehicle trajectory prediction device based on road ETC data can be integrated in a cloud server or a local server. Figure 1 As shown, the vehicle trajectory prediction method based on road ETC data may include the following operations: Step 101, obtaining all ETC data corresponding to the road system; In the embodiment of the present invention, ETC data includes the record information of each vehicle passing through each ETC gantry; specifically, a complete record of each vehicle passing through each gantry on the highway can be constructed through the vehicle ID and gantry ID, and each record contains key information such as the time, location (gantry ID) and vehicle ID of the vehicle passing through a certain gantry. By summarizing and analyzing the gantry passage records of all vehicles, necessary historical data support is provided for subsequent trajectory prediction tasks.

[0037] Step 102: Based on the Word2Vec model, the ETC gantry sequence in the historical trajectory of each vehicle in the ETC data is trained to generate an embedding vector, which is used to represent the relative position of the ETC gantry and the traffic flow pattern.

[0038] In order for computers to process human text information, text encoding is an indispensable part. Common encoding methods include binary encoding, character encoding, and Unicode encoding, which effectively solve the problem of computer storage and calculation of text information. However, in order to enable computers to understand text information more intelligently, traditional encoding methods are still not enough to meet the needs. For this reason, the field of artificial intelligence has proposed language models and word vectorization methods to improve the computer's ability to understand language.

[0039] Common word vectorization methods include one-hot representation and distributed representation. One-hot representation maps each word to a sparse binary vector, while distributed representation more effectively captures the semantic relationship between words by mapping the words to low-dimensional vectors, and has stronger expressive power. The vehicle trajectory model of the neural network has the disadvantage of a large number of parameters due to its structure, and requires a lot of calculations to obtain the feature vector of the gantry. The embodiment of the present invention proposes a method for constructing a gantry feature vector based on word2vec on the basis of ETC data, which is specifically used to convert gantry information into a feature vector. In an embodiment of the present invention, the ETC gantry sequence is mapped to a high-dimensional embedding space based on the Word2Vec model to capture the relative relationship between the gantries and the semantic features in the traffic flow. Optionally, by training the gantry ID sequence in the historical trajectory of the vehicle, the Word2Vec model can automatically learn the similarity between the gantries and their spatial relationship, thereby generating an embedding vector that can reflect the relative position of the gantries and the traffic flow pattern. This provides rich space-time information for the path prediction task and solves the problem of traditional time series prediction methods ignoring spatial features. By combining the gantry feature vector with the historical trajectory data, the time series and spatial information can be processed simultaneously in the process of constructing the embedding vector, fully mining the ETC data and the relationship between the ETC data and the vehicle operation.

[0040] Step 103: Train the embedding vector based on the multi-layer bidirectional LSTM model to obtain a trained multi-layer bidirectional LSTM model, and use the trained multi-layer bidirectional LSTM model to predict the vehicle trajectory; In an embodiment of the present invention, a multi-layer bidirectional LSTM model may include two alternatingly connected bidirectional LSTM layers to capture forward and backward dependencies in the vehicle trajectory. Specifically, the traditional unidirectional LSTM model can only utilize the forward information flow when processing sequence data, which limits the model's ability to capture bidirectional dependencies in long time series data. In the vehicle path prediction task, considering that the future driving trajectory of the vehicle is affected by the bidirectional driving conditions of the front and rear vehicles and the traffic environment, the prediction effect of the unidirectional LSTM model is relatively poor. To solve this problem, the multi-layer bidirectional LSTM model proposed in an embodiment of the present invention combines the forward and backward information flows by adding a bidirectional structure on the basis of LSTM. Figure 2 As shown in Figure 2, the multi-layer bidirectional LSTM model is connected alternately through two bidirectional LSTM layers, which can simultaneously capture the forward and backward dependencies in the vehicle trajectory, thereby enhancing the ability to learn complex dynamic patterns.

[0041] The embodiment of the present invention constructs an improved multi-layer bidirectional LSTM model for predicting the future driving trajectory of a vehicle. When processing long time series data, the traditional unidirectional LSTM structure often cannot fully exploit the bidirectional dependencies in the sequence data, thereby limiting the prediction accuracy of the model. The multi-layer bidirectional LSTM model of the embodiment of the present invention, by combining the reverse information flow, can simultaneously capture the forward and backward dependencies in the vehicle's driving trajectory, significantly enhancing the ability to learn complex dynamic patterns in long time series data.

[0042] It can be seen that the embodiment of the present invention trains the ETC gantry sequence in each vehicle historical trajectory in the ETC data based on the Word2Vec model to generate an embedding vector, so that the temporal and spatial information can be processed simultaneously in the process of constructing the embedding vector, and the relationship between the ETC data and the vehicle operation is fully mined. At the same time, the multi-layer bidirectional LSTM model of the embodiment of the present invention, by combining the reverse information flow, can simultaneously capture the forward and backward dependencies in the vehicle's driving trajectory, and significantly enhances the learning ability of complex dynamic patterns in long time series data. Ultimately, the vehicle trajectory prediction method based on road ETC data of the embodiment of the present invention improves the accuracy of vehicle trajectory prediction.

[0043] In an optional embodiment, obtaining all ETC data corresponding to the road system may include: Obtain all original ETC data corresponding to the road system, and for any two ETC gantries with a preset association relationship, determine the reasonable range of the travel time between the two ETC gantries based on the road information between the two ETC gantries; For all original ETC data, after eliminating the erroneous ETC data, all ETC data corresponding to the road system are obtained; wherein, the erroneous ETC data is the passing record of vehicles whose travel time is not within a reasonable range between any two ETC gantries with a preset association relationship.

[0044] In this optional embodiment, there are often some errors or inaccurate data in the ETC data, and data screening is required to reduce the inaccuracy of prediction results caused by errors or inaccurate data. This optional embodiment determines the reasonable range of the travel time between the two ETC gantries based on the road information between the two ETC gantries. For example, for two adjacent ETC gantries, the maximum and minimum values ​​of the travel time can be calculated based on the normal travel speed of the vehicle, thereby forming a reasonable range of travel time. For travel conditions that are not within the reasonable range of the travel time, such as too long travel time and too short travel time, they are all considered unreasonable data and are eliminated, thereby improving the rationality of the data.

[0045] In another optional embodiment, the ETC data may contain unreasonable data caused by various other factors, such as license plate recognition errors, information errors, etc. Therefore, in this optional embodiment, before removing the erroneous ETC data from all original ETC data and obtaining all ETC data corresponding to the road system, the method may further include: Perform gantry information error correction, license plate information error correction and trip information error correction operations on all original ETC data, including: The gantry information error correction operation may include: identifying and correcting gantry information errors by comparing the timestamps and geographic locations of historical data with current data; The license plate information error correction operation may include: using data cleaning technology to identify abnormal license plate information through data comparison at multiple time points, and correcting the license plate data in combination with the vehicle owner information database; The trip information error correction operation may include: supplementing discontinuous time periods through prediction and interpolation, and recalibrating the trip information in combination with the vehicle's historical trajectory data.

[0046] This optional embodiment effectively reduces the impact of erroneous data on vehicle trajectory integrity and prediction accuracy through the above-mentioned data preprocessing method, and provides high-quality basic data for subsequent path prediction research.

[0047] In yet another optional embodiment, based on the Word2Vec model, the ETC gantry sequence in each vehicle historical trajectory in the ETC data is trained to generate an embedding vector, which may include: Set the model input: Set the input vocabulary to ,in is the ID of the target ETC gantry, confirm arrive For vocabulary The surrounding 2k context words; Set the model goal: Set the model's objective to maximize the vocabulary The corresponding contextual vocabulary conditional probability, thereby adjusting the vocabulary The corresponding embedding vector; where the vocabulary The corresponding context word conditional probability is ; Conditional probability refers to the probability of In the case of , the probability of the occurrence of its context word, the calculation of conditional probability can be achieved by the softmax function.

[0048] Setting the loss function: Set a central word and its context words , the number of negative samples is k, then the loss function is set to:

[0049] In the above formula, are negative samples randomly drawn from the vocabulary, is a given central word When predicting the context word The probability of is a given central word When , the randomly selected negative samples probability.

[0050] In this optional embodiment, the Word2Vec model continuously updates the word vectors so that the distance between vectors in the word vector space is consistent with the semantic relationship between words. During the training process, the model continuously adjusts the vectors of the gantry ID so that the word vectors in similar contexts are closer, and the distance between unrelated word vectors is farther. Optionally, specific training methods may include techniques such as negative sampling and hierarchical Softmax to accelerate training and reduce computational costs. By maximizing the conditional probability of contextual vocabulary, the Word2Vec model can capture the similarity and dependency between gantry IDs. In the application of highway ETC data, it can effectively improve the accuracy of tasks such as path prediction and traffic flow analysis, and provide data support for subsequent traffic management.

[0051] In yet another optional embodiment, training the embedding vector based on the multi-layer bidirectional LSTM model may include: The embedding vector is trained based on a multi-layer bidirectional LSTM model, where the output of the multi-layer bidirectional LSTM model at each time step t is:

[0052] In the above formula, and are the hidden states of the forward and reverse LSTM, respectively, where:

[0053] In the above formula, is the input at time step t, and Respectively represent the calculation process of forward and reverse LSTM, and are the hidden states of the previous and next moments corresponding to time step t respectively.

[0054] It can be seen that the output of the multi-layer bidirectional LSTM model at each time step t depends not only on the state of the previous moment, but also on the information of the subsequent moments. By alternately connecting two bidirectional LSTM layers, the forward and backward dependencies in the vehicle trajectory can be captured simultaneously, thereby enhancing the learning ability of complex dynamic patterns.

[0055] In yet another optional embodiment, training the embedding vector based on the multi-layer bidirectional LSTM model to obtain the trained multi-layer bidirectional LSTM model may include: The embedding vector is trained based on the multi-layer bidirectional LSTM model. Through the attention mechanism, the weight of each time step is dynamically adjusted according to the importance of the input data, so as to match the most relevant vehicle historical trajectory for each prediction and obtain a trained multi-layer bidirectional LSTM model.

[0056] In this optional embodiment, the existing time series prediction model usually assigns the same weight to the input of each time step, which makes it difficult for the model to effectively distinguish key information from secondary information during the prediction process. In order to solve this problem, the present invention introduces an attention mechanism, and implements the training of a multi-layer bidirectional LSTM model through the attention mechanism. The multi-layer bidirectional LSTM model can dynamically adjust the weight according to the task requirements and automatically focus on the historical trajectory part that is most relevant to the current prediction, thereby improving the prediction accuracy and effectively reducing the computational complexity.

[0057] In an optional embodiment, the weight of each time step is dynamically adjusted according to the importance of the input data through an attention mechanism, which may include: The attention coefficient corresponding to time step t is calculated according to the following formula:

[0058] In the above formula, is the hidden state at the current time step t The matching score between the hidden states of all historical trajectories, It is the matching score between the hidden state of the i-th time step and the hidden state of all historical trajectories. The matching score is calculated by dot product or weighted sum.

[0059] In another optional embodiment, not all vehicle trajectories can be used for model training and trajectory prediction, such as vehicles with short mileage, vehicles with low travel frequency, vehicles with short road time, vehicles that only travel within a small range, etc. Therefore, obtaining all ETC data corresponding to the road system may include: Obtain all initial ETC data corresponding to the road system, and determine the historical trajectory of each identified vehicle based on the initial ETC data; For each identified vehicle, all ETC gantries corresponding to the identified vehicle are determined based on the initial ETC data. The predictable parameters corresponding to the identified vehicle are determined based on the number and distribution range of all ETC gantries corresponding to the identified vehicle. The predictable parameters are used to measure the accuracy of the prediction of the trajectory of the corresponding identified vehicle. Optionally, the fewer the number of ETC gantries and the smaller the distribution range, the smaller the corresponding predictable parameters.

[0060] The vehicle to be predicted is selected from all identified vehicles, and the predictable parameters corresponding to the vehicle to be predicted are greater than or equal to the preset predictable threshold, and the initial ETC data corresponding to all the vehicles to be predicted are determined to be all ETC data corresponding to the road system. Through the above operation, the object used for prediction in this optional embodiment is a reasonable object, which improves the prediction accuracy of the model.

[0061] Embodiment 2 The embodiment of the present invention discloses a specific vehicle trajectory prediction method based on road ETC data, which mainly includes: (1) Preprocessing of ETC data and extraction of trajectory data. This step involves cleaning and integrating the original ETC data to extract the driving trajectory of each vehicle. Specifically, a complete record of each vehicle passing through each gantry on the highway is constructed through the vehicle ID and gantry ID. Each record contains key information such as the time, location (gantry ID) and vehicle ID of the vehicle passing through a gantry. By summarizing and analyzing the gantry passage records of all vehicles, the necessary historical data support is provided for the subsequent trajectory prediction task.

[0062] (2) An innovative method for constructing ETC gantry feature vectors is proposed. Based on the Word2Vec model, the gantry ID is mapped to a high-dimensional embedding space to capture the relative relationship between gantries and the semantic features in traffic flow. By training the gantry ID sequence in the vehicle's historical trajectory, the Word2Vec model can automatically learn the similarity between gantries and their spatial relationship, thereby generating an embedding vector that can reflect the relative position of the gantries and the traffic flow pattern. The model provides rich spatial-temporal information for the path prediction task and solves the problem that traditional time series prediction methods ignore spatial features. By combining the gantry feature vector with historical trajectory data, the model can process both time series and spatial information at the same time, significantly improving the prediction accuracy.

[0063] (3) An improved MBiLSTM-ATT highway path prediction model was constructed to predict the future driving trajectory of vehicles. When processing long time series data, the traditional unidirectional LSTM structure often cannot fully explore the bidirectional dependencies in the sequence data, thereby limiting the prediction accuracy of the model. To solve this problem, the embodiment of the present invention proposes a multi-layer bidirectional LSTM (MBiLSTM) model, which can simultaneously capture the forward and backward dependencies in the vehicle's driving trajectory by combining the reverse information flow, significantly enhancing the learning ability of complex dynamic patterns in long time series data. In addition, the existing time series prediction models usually assign the same weight to the input of each time step, which makes it difficult for the model to effectively distinguish key information from secondary information during the prediction process. To solve this problem, the present invention introduces an attention mechanism and proposes an MBiLSTM-ATT model. The model can dynamically adjust the weight according to the task requirements and automatically focus on the part of the historical trajectory that is most relevant to the current prediction, thereby improving the prediction accuracy and effectively reducing the computational complexity.

[0064] The specific steps of the embodiment of the present invention are described below respectively: 1. ETC data preprocessing The vehicle traffic data recorded by the highway ETC gantry system is called ETC data, but there are often missing or abnormal data. The erroneous data usually include gantry information errors, license plate information errors and trip ID errors, which will affect the integrity of the vehicle trajectory and have an adverse impact on subsequent research.

[0065] In order to solve the problem of missing or abnormal data in ETC data, the embodiment of the present invention adopts a series of data processing methods. First, for gantry information errors, by comparing the timestamps and geographic locations of historical data with current data, gantry ID errors are identified and corrected, and reasonable range restrictions on the distance between gantries and the length of travel time are introduced to eliminate abnormal data that does not conform to the actual situation. For license plate information errors, data cleaning technology is used to identify abnormal license plate information through data comparison at multiple time points, and the license plate data is automatically corrected in combination with license plate recognition technology and the owner information database, thereby improving the accuracy of license plate information. For trip ID errors, the present invention introduces a time series correction method, supplements discontinuous time periods through interpolation and prediction, and recalibrates the trip ID in combination with historical trajectory data to ensure the continuity and accuracy of each trip. These processing methods effectively reduce the impact of erroneous data on vehicle trajectory integrity and prediction accuracy, and provide high-quality basic data for subsequent path prediction research.

[0066] 2. ETC gantry feature vector construction method The vehicle trajectory model of the neural network has the disadvantage of a large number of parameters due to its structure, and requires a lot of calculations to obtain the feature vector of the gantry. The embodiment of the present invention proposes a method for constructing a gantry feature vector based on word2vec based on ETC data, which is specifically used to convert the gantry into a feature vector.

[0067] Wrod2vec is actually an unsupervised learning method, which aims to learn the feature vector representation of words based on the distribution of elements in the sequence. Its evaluation depends more on the effect of its downstream tasks, that is, the subsequent vehicle trajectory prediction effect.

[0068] The word2vec model in the embodiment of the present invention is trained by learning the distributed representation of a given gantry ID in its context, and its goal is to predict its context words through the input center word, so as to obtain the embedded representation of each word. Specifically, it may include: Set the input vocabulary to ,in is a gantry ID in a sequence, and is the 2k words around the vocabulary (context words). The goal of the word2vec model is to maximize the given central word The conditional probability of the corresponding context word is:

[0069] The model adjusts the embedding vector of each word by maximizing the above conditional probability, so that words that appear in similar contexts have similar vector representations. The training process of the word2vec model involves minimizing the error through gradient descent to obtain the vector of each word (gantry ID).

[0070] Furthermore, the loss function of the word2vec model is negative sampling loss, which simplifies the training process by randomly sampling negative samples. and its context words , the number of negative samples is k, then the loss function can be expressed as:

[0071] in, is the target context word, are negative samples randomly drawn from the vocabulary, is a given central word When predicting the context word probability.

[0072] The word2vec model continuously updates word vectors so that the distance between vectors in the word vector space is consistent with the semantic relationship between words. During the training process, the model continuously adjusts the vectors of the gantry ID so that word vectors in similar contexts are closer and the distance between unrelated word vectors is farther. Specific training methods can include techniques such as negative sampling and hierarchical softmax to accelerate training and reduce computational costs.

[0073] By maximizing the conditional probability of contextual words, the model can capture the similarities and dependencies between gantry IDs. This method can effectively improve the accuracy of tasks such as path prediction and traffic flow analysis in the application of highway ETC data, and provide data support for subsequent traffic management.

[0074] 3. Highway Path Prediction Model (1) Building a multi-layer bidirectional LSTM model The traditional unidirectional LSTM model can only use the forward information flow when processing sequence data, which limits the model's ability to capture bidirectional dependencies in long time series data. In the vehicle path prediction task, considering that the future driving trajectory of a vehicle is affected by the driving conditions of the vehicles ahead and behind and the traffic environment, the prediction effect of the unidirectional LSTM model is poor.

[0075] To solve this problem, the multi-layer bidirectional LSTM model proposed in the embodiment of the present invention combines the forward and reverse information flows by adding a bidirectional structure on the basis of LSTM. Figure 2As shown in Figure 2, the multi-layer bidirectional LSTM model is connected alternately through two bidirectional LSTM layers, which can simultaneously capture the forward and backward dependencies in the vehicle trajectory, thereby enhancing the ability to learn complex dynamic patterns.

[0076] Specifically, the output of the multi-layer bidirectional LSTM model at each time step t depends not only on the state of the previous moment, but also on the information of the subsequent moments. is the input at time t, and are the hidden states of the forward and reverse LSTMs respectively, then the bidirectional LSTM output at the tth moment is:

[0077] in, and Computed from the following forward and backward LSTM states:

[0078] and Respectively represent the calculation process of forward and reverse LSTM, and are the hidden states at the previous and next moments respectively.

[0079] (2) Fusion Attention Mechanism In the vehicle path prediction task, the influence of each time step in the historical trajectory data is not the same. Traditional time series prediction models usually give the same weight to the input of each time step, which makes it difficult for the model to effectively distinguish key information from secondary information. To solve this problem, the embodiment of the present invention introduces an attention mechanism, such as Figure 3 As shown in Figure 3, by dynamically adjusting the weights of each time step in the historical trajectory data, we focus on the information most relevant to the current prediction.

[0080] Assume that the hidden state sequence of the model is , where T is the number of time steps, the attention mechanism calculates the weight coefficient of each hidden state To control the contribution of each time step. Specifically, the attention mechanism is based on the hidden state of the current time step t. The weight is calculated by the similarity between the hidden states of all historical trajectories. The calculation method is as follows:

[0081] in, is the current time step t and the historical state The matching score between them is usually calculated in the form of dot product or weighted sum. It indicates the contribution of the tth time step to the final prediction. The final weighted representation It can be obtained by taking the weighted sum of all historical states:

[0082] The multi-layer bidirectional LSTM model of the present invention can effectively utilize the complex dynamic patterns in long time series data when processing vehicle trajectory prediction tasks, overcome the limitations of traditional unidirectional LSTM models, and further improve the prediction accuracy and computational efficiency by introducing the attention mechanism.

[0083] Existing path prediction methods are mainly based on traditional unidirectional LSTM or other time series prediction models. When processing long time series data, these methods can usually only capture the unidirectional dependency of data and cannot effectively mine the relationship between the previous and next time steps, resulting in certain limitations in the learning of complex dynamic patterns. In addition, traditional models usually give the same weight to the input of each time step, making it difficult to accurately distinguish between key information and secondary information, thus affecting the prediction accuracy.

[0084] Compared with the prior art, the multi-layer bidirectional LSTM model proposed in the embodiment of the present invention can capture both forward and backward dependencies by introducing reverse information flow, significantly improving the learning ability of complex dynamic patterns of time series data. Especially when processing long time series data, the multi-layer bidirectional LSTM model can better capture the bidirectional dependencies in the time series data, overcoming the shortcomings of the unidirectional LSTM model.

[0085] In addition, the attention mechanism introduced in this invention enables the multi-layer bidirectional LSTM model to dynamically adjust the weight of each time step according to the importance of the input data, thereby focusing on the historical information most relevant to the current prediction, significantly improving the prediction accuracy and effectively reducing the computational complexity. This innovation in the field of traffic path prediction can solve the problem that traditional models cannot effectively focus on key time steps when processing complex traffic data.

[0086] Therefore, compared with the prior art, the embodiments of the present invention have obvious advantages in feature representation, time series data modeling and prediction accuracy improvement, especially in processing complex long time series data and improving prediction accuracy.

[0087] Embodiment 3 See also Figure 4 , Figure 4 Schematic diagram of the structure of a vehicle trajectory prediction device based on road ETC data disclosed in an embodiment of the present invention. Figure 4 As shown, the vehicle trajectory prediction device based on road ETC data may include: An acquisition module 201 is used to acquire all ETC data corresponding to the road system, and the ETC data may include record information of each vehicle passing through each ETC gantry; The Word2Vec module 202 is used to train the ETC gantry sequence in each vehicle historical trajectory in the ETC data based on the Word2Vec model to generate an embedding vector, where the embedding vector is used to represent the relative position of the ETC gantry and the traffic flow pattern; The LSTM module 203 is used to train the embedding vector based on the multi-layer bidirectional LSTM model to obtain a trained multi-layer bidirectional LSTM model, and the trained multi-layer bidirectional LSTM model is used to predict the vehicle trajectory; Among them, the multi-layer bidirectional LSTM model can include two alternately connected bidirectional LSTM layers to capture the forward and backward dependencies in the vehicle trajectory.

[0088] In an optional embodiment, the LSTM module 203 trains the embedding vector based on the multi-layer bidirectional LSTM model, and the specific operation of obtaining the trained multi-layer bidirectional LSTM model may include: The embedding vector is trained based on the multi-layer bidirectional LSTM model. Through the attention mechanism, the weight of each time step is dynamically adjusted according to the importance of the input data, so as to match the most relevant vehicle historical trajectory for each prediction and obtain a trained multi-layer bidirectional LSTM model.

[0089] In another optional embodiment, the specific operation of the acquisition module 201 acquiring all ETC data corresponding to the road system may include: Obtain all original ETC data corresponding to the road system, and for any two ETC gantries with a preset association relationship, determine the reasonable range of the travel time between the two ETC gantries based on the road information between the two ETC gantries; For all original ETC data, after eliminating the erroneous ETC data, all ETC data corresponding to the road system are obtained; wherein, the erroneous ETC data is the passing record of vehicles whose travel time is not within a reasonable range between any two ETC gantries with a preset association relationship.

[0090] In yet another optional embodiment, the device may further include: The correction module (not shown) is used to perform gantry information error correction, license plate information error correction and trip information error correction operations on all original ETC data before the acquisition module obtains all ETC data corresponding to the road system after removing the erroneous ETC data from all original ETC data, wherein: The gantry information error correction operation may include: identifying and correcting gantry information errors by comparing the timestamps and geographic locations of historical data with current data; The license plate information error correction operation may include: using data cleaning technology to identify abnormal license plate information through data comparison at multiple time points, and correcting the license plate data in combination with the vehicle owner information database; The trip information error correction operation may include: supplementing discontinuous time periods through prediction and interpolation, and recalibrating the trip information in combination with the vehicle's historical trajectory data.

[0091] In another optional embodiment, the Word2Vec module 202 trains the ETC gantry sequence in each vehicle historical trajectory in the ETC data based on the Word2Vec model, and the specific operation of generating the embedding vector may include: Set the model input: Set the input vocabulary to ,in is the ID of the target ETC gantry, confirm arrive For vocabulary The surrounding 2k context words; Set the model goal: Set the model's objective to maximize the vocabulary The corresponding contextual vocabulary conditional probability, thereby adjusting the vocabulary The corresponding embedding vector; where the vocabulary The corresponding context word conditional probability is ; Setting the loss function: Set a central word and its context words , the number of negative samples is k, then the loss function is set to:

[0092] In the above formula, are negative samples randomly drawn from the vocabulary, is a given central word When predicting the context word The probability of is a given central word When , the randomly selected negative samples probability.

[0093] In yet another optional embodiment, the specific operation of the LSTM module 203 training the embedding vector based on the multi-layer bidirectional LSTM model may include: The embedding vector is trained based on a multi-layer bidirectional LSTM model, where the output of the multi-layer bidirectional LSTM model at each time step t is:

[0094] In the above formula, and are the hidden states of the forward and reverse LSTM, respectively, where:

[0095] In the above formula, is the input at time step t, and Respectively represent the calculation process of forward and reverse LSTM, and are the hidden states of the previous and next moments corresponding to time step t respectively.

[0096] In yet another optional embodiment, the specific operation of the LSTM module 203 dynamically adjusting the weight of each time step according to the importance of the input data through the attention mechanism may include: The attention coefficient corresponding to time step t is calculated according to the following formula:

[0097] In the above formula, is the hidden state at the current time step t The matching score between the hidden states of all historical trajectories is calculated by dot product or weighted sum.

[0098] In yet another optional embodiment, the specific operation of the acquisition module 201 acquiring all ETC data corresponding to the road system may include: Obtain all initial ETC data corresponding to the road system, and determine the historical trajectory of each identified vehicle based on the initial ETC data; For each identified vehicle, all ETC gantries corresponding to the identified vehicle are determined based on the initial ETC data, and the predictable parameters corresponding to the identified vehicle are determined based on the number and distribution range of all ETC gantries corresponding to the identified vehicle. The predictable parameters are used to measure the accuracy of the prediction of the trajectory of the corresponding identified vehicle; The vehicles to be predicted are selected from all identified vehicles, the predictable parameters corresponding to the vehicles to be predicted are greater than or equal to a preset predictable threshold, and the initial ETC data corresponding to all the vehicles to be predicted are determined to be all ETC data corresponding to the road system.

[0099] Embodiment 4 See also Figure 5 , Figure 5Schematic diagram of the structure of a vehicle trajectory prediction system based on road ETC data disclosed in an embodiment of the present invention. Figure 5 As shown, the vehicle trajectory prediction system based on road ETC data may include: A memory 301 storing executable program codes; a processor 302 coupled to the memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the steps in the vehicle trajectory prediction method based on road ETC data described in the first or second embodiment of the present invention.

[0100] The device embodiments described above are only illustrative, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, i.e., they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0101] Through the specific description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on such an understanding, the above technical solution can be essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, and the storage medium includes a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0102] Finally, it should be noted that the vehicle trajectory prediction method and device based on road ETC data disclosed in the embodiment of the present invention discloses only the preferred embodiments of the present invention, which are only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A vehicle trajectory prediction method based on road ETC data, characterized in that: The method comprises: Obtain all ETC data corresponding to the road system, wherein the ETC data includes record information of each vehicle passing through each ETC gantry; Based on the Word2Vec model, the ETC gantry sequence in each vehicle historical trajectory in the ETC data is trained to generate an embedding vector, wherein the embedding vector is used to represent the relative position of the ETC gantry and the traffic flow pattern; The embedded vector is trained based on a multi-layer bidirectional LSTM model to obtain a trained multi-layer bidirectional LSTM model, wherein the trained multi-layer bidirectional LSTM model is used to predict the vehicle trajectory; The multi-layer bidirectional LSTM model includes two alternately connected bidirectional LSTM layers to capture the forward and backward dependencies in the vehicle trajectory.

2. The vehicle trajectory prediction method based on road ETC data according to claim 1, characterized in that: The embedding vector is trained based on the multi-layer bidirectional LSTM model to obtain a trained multi-layer bidirectional LSTM model, including: The embedding vector is trained based on a multi-layer bidirectional LSTM model. Through the attention mechanism, the weight of each time step is dynamically adjusted according to the importance of the input data, so as to match the most relevant vehicle historical trajectory for each prediction and obtain a trained multi-layer bidirectional LSTM model.

3. The vehicle trajectory prediction method based on road ETC data according to claim 1, characterized in that: The obtaining of all ETC data corresponding to the road system includes: Obtain all original ETC data corresponding to the road system, and for any two ETC gantries with a preset association relationship, determine a reasonable range of the travel time between the two ETC gantries based on the road information between the two ETC gantries; For all the original ETC data, after eliminating the erroneous ETC data, all the ETC data corresponding to the road system are obtained; wherein, the erroneous ETC data is the passing record of vehicles between any two ETC gantries with a preset association relationship, whose travel time is not within the reasonable range.

4. The vehicle trajectory prediction method based on road ETC data according to claim 3 is characterized in that: Before obtaining all ETC data corresponding to the road system after eliminating erroneous ETC data from all the original ETC data, the method further includes: The gantry information error correction, license plate information error correction and trip information error correction operations are performed on all the original ETC data, wherein: The gantry information error correction operation includes: identifying and correcting gantry information errors by comparing the timestamps and geographic locations of historical data with those of current data; The license plate information error correction operation includes: using data cleaning technology to identify abnormal license plate information through data comparison at multiple time points, and correcting the license plate data in combination with the vehicle owner information database; The trip information error correction operation includes: supplementing discontinuous time periods by prediction and interpolation, and recalibrating the trip information in combination with the vehicle's historical trajectory data.

5. The vehicle trajectory prediction method based on road ETC data according to claim 1, characterized in that: The Word2Vec model is used to train the ETC gantry sequence in each vehicle historical trajectory in the ETC data to generate an embedding vector, including: Set the model input: Set the input vocabulary to ,in is the ID of the target ETC gantry, confirm arrive For vocabulary The surrounding 2k context words; Set the model goal: Set the model's objective to maximize the vocabulary The corresponding contextual vocabulary conditional probability, thereby adjusting the vocabulary The corresponding embedding vector; where the vocabulary The corresponding context word conditional probability is ; Setting the loss function: Set a central word and its context words , the number of negative samples is k, then the loss function is set to: In the above formula, are negative samples randomly drawn from the vocabulary, is a given central word When predicting the context word probability; is a given central word When , the randomly selected negative samples probability.

6. The vehicle trajectory prediction method based on road ETC data according to claim 2, characterized in that: The training of the embedding vector based on the multi-layer bidirectional LSTM model includes: The embedding vector is trained based on a multi-layer bidirectional LSTM model, wherein the output of the multi-layer bidirectional LSTM model at each time step t is: In the above formula, and are the hidden states of the forward and reverse LSTM, respectively, where: In the above formula, is the input at time step t, and Respectively represent the calculation process of forward and reverse LSTM, and are the hidden states of the previous and next moments corresponding to time step t respectively.

7. The vehicle trajectory prediction method based on road ETC data according to claim 6 is characterized in that: The attention mechanism dynamically adjusts the weight of each time step according to the importance of the input data, including: The attention coefficient corresponding to time step t is calculated according to the following formula : In the above formula, is the hidden state at the current time step t The matching score between the hidden states of all historical trajectories is calculated by dot product or weighted sum.

8. The vehicle trajectory prediction method based on road ETC data according to claim 1, characterized in that: The obtaining of all ETC data corresponding to the road system includes: Acquire all initial ETC data corresponding to the road system, and determine the historical trajectory of each identified vehicle based on the initial ETC data; For each identified vehicle, all ETC gantries corresponding to the identified vehicle are determined according to the initial ETC data, and predictable parameters corresponding to the identified vehicle are determined according to the number and distribution range of all ETC gantries corresponding to the identified vehicle, wherein the predictable parameters are used to measure the accuracy of the prediction of the trajectory of the corresponding identified vehicle; A vehicle to be predicted is selected from all the identified vehicles, and a predictable parameter corresponding to the vehicle to be predicted is greater than or equal to a preset predictable threshold value, and initial ETC data corresponding to all the vehicles to be predicted are determined to be all ETC data corresponding to the road system.

9. A vehicle trajectory prediction device based on road ETC data, characterized in that: The device comprises: An acquisition module, used to acquire all ETC data corresponding to the road system, wherein the ETC data includes record information of each vehicle passing through each ETC gantry; A Word2Vec module, for training the ETC gantry sequence in each vehicle historical trajectory in the ETC data based on the Word2Vec model to generate an embedding vector, wherein the embedding vector is used to represent the relative position of the ETC gantry and the traffic flow pattern; An LSTM module is used to train the embedding vector based on a multi-layer bidirectional LSTM model to obtain a trained multi-layer bidirectional LSTM model, wherein the trained multi-layer bidirectional LSTM model is used to predict the vehicle trajectory; The multi-layer bidirectional LSTM model includes two alternately connected bidirectional LSTM layers to capture the forward and backward dependencies in the vehicle trajectory.

10. A vehicle trajectory prediction system based on road ETC data, characterized in that: The system includes: a memory storing executable program code; a processor coupled to the memory; the processor calls the executable program code stored in the memory to execute the vehicle trajectory prediction method based on road ETC data as described in any one of claims 1-7.

Images