Vehicle-mounted track prediction and correction method based on encoder-decoder neural network
Through the vehicle-mounted trajectory prediction method based on the encoder-decoder neural network, the multi-parameter time series data processing problem in the prior art is solved, high-precision long-term prediction and multi-scene adaptability are achieved, and it is suitable for intelligent transportation systems.
Patent Information
- Application Number
- CN202510478199.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-16
- Publication Date
- 2025-08-01
AI Technical Summary
The existing vehicle-mounted trajectory prediction methods are difficult to process multi-parameter time series data, cannot make long-term predictions, and are poor in adaptability, especially in different road scenarios, the prediction accuracy needs to be improved.
The vehicle-mounted trajectory prediction method based on the encoder-decoder neural network is adopted, and the vehicle-mounted trajectory prediction and correction model is constructed through data preprocessing, the deep encoder-decoder network structure is expanded and the attention mechanism is introduced, and the trajectory correction is combined with the deep neural network to construct the vehicle-mounted trajectory prediction and correction model.
It realizes high-precision long-term trajectory prediction, adapts to a variety of road scenarios, has good generalization capabilities, can correct prediction errors in real time, and meets practical application needs.
Smart Images
Figure CN120408078A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of intelligent transportation systems and relates to a vehicle trajectory prediction and correction method based on an encoder-decoder neural network. Background Art
[0002] With the rapid development of intelligent connected vehicles and vehicle networking technologies, advanced driver assistance systems play an important role in improving traffic efficiency and reducing traffic accidents. As one of the core algorithms of advanced driver assistance systems, vehicle trajectory prediction can predict the driving trajectory of a vehicle within a certain period in the future, provide safe driving suggestions for drivers, and assist the autonomous driving system in making decisions.
[0003] Existing vehicle trajectory prediction methods are mainly divided into two categories:
[0004] Traditional mathematical methods: such as hidden Markov models, Gaussian process models, rapidly-exploring random tree models, etc. These methods have certain advantages in short-term prediction, but it is difficult to process multi-parameter time series data and long-term prediction cannot be carried out.
[0005] Deep learning methods: such as fully-connected neural networks, recurrent neural networks, convolutional neural networks, etc. These methods can process complex time series data and perform long-term prediction, but existing research mainly focuses on straight road scenarios, has poor adaptability to other scenarios, and the prediction accuracy needs to be improved.
[0006] Therefore, a vehicle trajectory prediction and correction method that can process multi-parameter time series data, adapt to various environments, and effectively improve prediction accuracy is needed to solve the above technical problems. Summary of the Invention
[0007] The present invention aims to provide a vehicle trajectory prediction and correction method based on a deep encoder-decoder neural network, which can effectively improve prediction accuracy, achieve long-term prediction, and adapt to different road scenarios.
[0008] The technical solution adopted by the present invention to solve the technical problems is: a vehicle trajectory prediction and correction method based on an encoder-decoder neural network, including the following steps:
[0009] Step 1: Data preprocessing, using the head and tail sampling algorithm to process the original trajectory data, converting trajectory data of different lengths into data of equal length, and gradually reducing the number of input data coordinates and increasing the number of output data coordinates to extend the prediction time;
[0010] Step 2: Construction of a deep encoder-decoder network, expanding the deep encoder-decoder network structure, introducing an attention mechanism for semantic recognition, and determining the optimal network structure through multi-parameter comparative experiments to predict the future trajectory of the vehicle;
[0011] Step 3: Trajectory correction. Use a deep neural network to correct the prediction results of the deep encoder-decoder network structure. The input of the deep neural network includes vehicle motion state data, and the output of the deep neural network is a trajectory correction matrix;
[0012] Step 4: Model cascade verification. Cascade the deep encoder-decoder network and the deep neural network to construct an in-vehicle trajectory prediction and correction model. Verify the overall performance of the model through cascade experiments and compare it with other prediction models.
[0013] Preferably, in the said Step 1, the specific steps of converting trajectory data of different lengths into data of equal length include:
[0014] Represent the original trajectory data as a vector x = {x1, x2, ……, x n} of a fixed dimension, and calculate the sampling points sp1, sp2 …… sp t as:
[0015]
[0016] and there is
[0017] n ∈ Z
[0018] t ∈ Z (2)
[0019] In formulas (1) and (2), x represents a vector of a fixed dimension, n represents the number of trajectory coordinates, sp1, sp2, …, sp t represents the sampling points, t represents the number of sampling times, and Z is the set of positive integers. Preferably, in Steps 1-3, the generated vehicle different motion state data does not include the vehicle trajectory.
[0020] More preferably, in the said Step 2, expanding the deep encoder-decoder network structure specifically includes:
[0021] Input the input trajectory data x = {x1, x2, ···, x n} into the encoder, and calculate the hidden state ht at time t as:
[0022] h t = f(x t , h t-1 ) (3)
[0023] Use the attention mechanism to convert the hidden state ht into a semantic vector c of a fixed dimension,
[0024] c = q(h1, h2, ……, h t ) (4)
[0025] Decode the decoder using the semantic vector C = {C1, C2, ···, C n} to calculate the prediction result at time t, where S t represents the state of the decoder hidden layer at time t:
[0026]
[0027] P(y t |y1, y2, ···, y t-1 , c) = g(y t-1 , s t , c) (6)
[0028] In equations (5) and (6), P(y) represents the probability distribution of the future trajectory y, and P(y t |y1, y2, ···, y t-1 , c) represents the probability of predicting y t at time t, and y t-1 represents the prediction result at time t - 1.
[0029] More preferably, in step 2, introducing the attention mechanism for semantic recognition specifically includes: adding attention to the i-th layer, and the conditional probability of the future trajectory y at time t is:
[0030] P(y t |y1, y2, ···, y t-1 , x) = g(y t-1 , s t , c t ) (7)
[0031] s t = f(s t-1 y t-1 , c t ) (8)
[0032]
[0033] e tj = score(s t-1 , h j ) = Vσ(W s-1 + Uh j ) (11)
[0034] In equations (7) to (11), x represents the input data, c t represents the attention vector, a tj represents the weight at time t, h j represents the state of the j-th step hidden layer, e tjRepresents the relevance score of the j-th part of the input data x at time t. score represents the scoring function, and s t-1 Represents the decoder hidden state at time t-1. V, W, and U respectively represent the weight matrices to be trained, and σ represents the activation function.
[0035] Preferably, in step 2, determining the optimal network structure specifically includes:
[0036] Increase the number of label data coordinates in different training steps; before training, normalize the data through the MinMaxScaler function to accelerate the model convergence speed and eliminate the differences between data:
[0037]
[0038] In Equation (12), xscaled represents the normalized value, and axis = 0 represents row normalization; select the mean squared error MSE as the loss function, and the mean squared error MSE is:
[0039]
[0040] In Equation (13), y i real Represents the true result, and y i pre Represents the predicted result, and m represents the number of samples in the dataset.
[0041] Preferably, step 3 specifically includes:
[0042] Step 3-1: Form a vehicle state matrix according to the motion state of the vehicle in different sections as the input of the deep neural network;
[0043] Step 3-2: Calculate the trajectory prediction error between the actual trajectory and the predicted trajectory of the vehicle to form a trajectory error matrix as the label of the deep neural network;
[0044] Step 3-3: After the deep neural network is trained, calculate the correction matrix using the new vehicle state matrix, and use the correction matrix to correct the predicted trajectory.
[0045] Preferably, in step 3-3, a structure layer of 3 to 6 layers of deep neural network is used for training.
[0046] Preferably, step 4 specifically includes:
[0047] Step 4-1: Train and test the deep encoder-decoder network and the deep neural network respectively;
[0048] Step 4-2: Use the output of the depth encoder-decoder network as the input of the deep neural network, and use the output of the deep neural network as the final prediction result;
[0049] Step 4-3: Use the test data set to verify the overall performance of the model and compare it with other prediction models.
[0050] The beneficial effects of the present invention are as follows:
[0051] 1. By introducing the attention mechanism and the depth encoder-decoder network structure, the present invention can effectively extract the features in the trajectory data and perform accurate long-term prediction. The prediction accuracy is significantly better than the existing methods. Therefore, the present invention has high prediction accuracy.
[0052] 2. By gradually reducing the number of input data coordinates and increasing the number of output data coordinates, the present invention can achieve a longer prediction time and meet the actual application requirements. Therefore, the present invention has strong long-term prediction ability.
[0053] 3. The present invention uses various road scene data for training, can adapt to different road scenes such as highways, intersections, and roundabouts, and has good generalization ability. Therefore, the present invention has strong multi-scene adaptability.
[0054] 4. By using DNN to correct the prediction results, the present invention can effectively eliminate the prediction error and inverse normalization error, making the predicted trajectory closer to the real trajectory. Therefore, the present invention has good correction effect.
[0055] 5. The deep neural network structure used in the present invention has a fast calculation speed and can meet the requirements of real-time prediction. Therefore, the present invention has strong real-time performance. Brief Description of the Drawings
[0056] Figure 1 is the model architecture diagram of the vehicle trajectory prediction and correction method based on the encoder-decoder neural network of the present invention;
[0057] Figure 2 is the input and label data sampling method diagram of the present invention;
[0058] Figure 3 is the LSTM encoder-decoder diagram of the present invention;
[0059] Figure 4 is the deep encoder-decoder network diagram based on the attention mechanism of the present invention;
[0060] Figure 5 is the schematic diagram of the optimized architecture of the present invention;
[0061] Figure 6 is the trajectory prediction result of the cascade model of the present invention on the highway;
[0062] Figure 7 Trajectory prediction results of the cascade model of the present invention at intersections;
[0063] Figure 8 Trajectory prediction results of the cascade model of the present invention at roundabouts;
[0064] Figure 9 Structural diagram of the cascade prediction model of the present invention. Detailed implementation manners
[0065] Next, the related technologies in the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.
[0066] The technical solutions adopted by the present invention include the following steps:
[0067] Refer to Figures 1 to 9 as shown, wherein, Figure 1 shows the overall architecture of the vehicle trajectory prediction model based on the deep encoder-decoder and deep neural network of the present invention, including input data, a prediction module, and a correction module, reflecting each module and system of the present invention.
[0068] Figure 2 shows the H2T sampling algorithm of the present invention, illustrates how to sample the input data, and ensures the consistency of the lengths of the input and output data, providing a data basis for the training and testing of the subsequent prediction module.
[0069] Figure 3 shows the scalable deep encoder-decoder structure of the present invention, highlighting the improvement in the model structure.
[0070] Figure 4 shows the deep encoder-decoder network structure after introducing the attention mechanism in the present invention, explains the role of the attention mechanism in the model, and how to use the attention mechanism to improve the prediction accuracy.
[0071] Figure 5 shows the model structure after cascading the deep encoder-decoder and DNN in the present invention, and explains how the present invention integrates the prediction and correction functions into one model to achieve more accurate trajectory prediction.
[0072] Figures 6 - 8Shows the comparison between the results of the cascade model and the actual trajectory in the present invention. It can be seen that for all three types of road sections, the lateral and longitudinal trajectories predicted by the cascade model are highly consistent with the actual trajectory. Therefore, the cascading of the prediction models improves the model performance.
[0073] The specific implementation method includes the following steps:
[0074] Step 1, data preprocessing: Use the head-to-tail sampling (H2T-Sampling) algorithm to process the original trajectory data, ensure that the lengths of the input and output data are the same, and gradually reduce the number of input data coordinates and increase the number of output data coordinates to extend the prediction time.
[0075] Step 2, construction of a deep encoder-decoder network: Expand the deep encoder-decoder network structure and introduce an attention mechanism for semantic recognition to improve the prediction accuracy. Through multi-parameter comparison experiments, determine the optimal network structure so that it can adapt to different road scenarios such as highways, intersections, and roundabouts.
[0076] Step 3, trajectory correction: Use a deep neural network (DNN) to correct the prediction results of the deep encoder-decoder network to eliminate prediction errors and inverse normalization errors. The input of the DNN includes vehicle motion state data, and the output is a trajectory correction matrix for correcting the predicted trajectory.
[0077] Step 4, model cascade verification: Cascade the deep encoder-decoder network and the DNN to construct a vehicle trajectory prediction and correction model. Verify the overall performance of the model through cascade experiments and compare it with other advanced prediction models. The results show that the model proposed in the present invention has higher prediction accuracy and longer prediction time.
[0078] The specific steps are as follows:
[0079] The first step: Data preprocessing, convert trajectory data of different lengths into data of equal length for easy model training. The specific steps are as follows:
[0080] Represent the original trajectory data as a vector x = {x1, x2, ……, x n}, where n is the number of trajectory coordinates. Calculate the sampling points sp1, sp2 …… sp t .
[0081]
[0082] And there is
[0083] n ∈ Z
[0084] t ∈ Z (2)
[0085] where x is a vector of fixed dimension, n is the number of trajectory coordinates, sp1, sp2, …, sp t are sampling points, t is the number of sampling times, and Z is the set of positive integers. The H2T-Sampling algorithm is used to preprocess the data, which can ensure that the lengths of the input and output data are consistent, facilitating model training and prediction.
[0086] Step 2: Construct a deep encoder-decoder network to predict the future trajectory of the vehicle. The specific steps are as follows:
[0087] Input the input trajectory data x = {x1, x2, ···, x n} into the encoder to calculate the hidden state at time t.
[0088] h t = f(x t , h t-1 ) (3)
[0089] Use the attention mechanism to convert the hidden state ht into a semantic vector c of fixed dimension.
[0090] c = q(h1, h2, ……, h t ) (4)
[0091] Use the semantic vector C = {C1, C2, ···, C n} to decode the decoder and calculate the prediction result at time t, where S t is the state of the decoder hidden layer at time t.
[0092]
[0093] P(y t |y1, y2,..., y t-1 , c) = g(y t-1 , s t , c) (6)
[0094] These two formulas describe how the decoder predicts the future trajectory based on the semantic vector c and the previous prediction results y1, y2, …, y t-1 . P(y) is the probability distribution of the future trajectory y. P(y t |y1, y2,..., y t-1 , c) represents the probability of predicting y t at time t, which is affected by the previous prediction results and the semantic vector c. g represents the decoding function, which calculates the probability of predicting y t based on the current hidden state s t-1 , the previous prediction result y t and the semantic vector c. s trepresents the hidden state of the decoder at time t. y t-1 : The prediction result at time t-1. c: The semantic vector.
[0095] After the decoder models the conditional probability distribution of the prediction result, the attention mechanism uses the input data as a reference output from the decoder or other layers. If attention is added to the i-th layer, the conditional probability of y at time t is expressed as:
[0096] P(y t |y1, y2..., y t-1 , x) = g(y t-1 , s t , c t ) (7)
[0097] s t = f(s t-1 y t-1 c t ) (8)
[0098] These two formulas describe how the attention mechanism affects the prediction process of the decoder. P(yt|y1,y2...,yt-1,x): The probability of predicting yt at time t considering the attention mechanism. x represents the input data, i.e., the historical trajectory data.
[0099]
[0100] Different from the classical encoder-decoder, in the proposed framework, after using the attention mechanism in the encoder-decoder, the semantic vector is dynamically changing rather than fixed. Formally, Equation (9) describes the calculation method of the attention vector c t , T represents the current length of the hidden state, h j represents the state of the j-th hidden layer. Furthermore, a tj is the weight at time t, expressed as
[0101]
[0102] where e tj is an important variable that measures the correlation between s t and h j . If the correlation is high, the attention is concentrated on the elements near the j position, and this part has a high weight. The value of e tj is calculated by the following formula:
[0103] e tj = score(s t-1 , h j ) = Vσ(W s-1 + Uh j) (11)
[0104] The meanings of each parameter are as follows:
[0105] e tj : The relevance score of the j-th part of the input data x at time t.
[0106] score: The scoring function, usually using the dot product or inner product to calculate the relevance.
[0107] s t-1 : The decoder hidden state at time t-1.
[0108] h j : The j-th hidden state.
[0109] V, W, U: Weight matrices to be trained.
[0110] σ: The activation function, usually using the soft max function.
[0111] To improve the long-term prediction performance of the network, the number of input data coordinates is reduced, while the number of label data coordinates is increased in different training steps. The calculation core is one of the key parameters of the deep encoder-decoder. Before training, the data is normalized by the MinMaxScaler function to accelerate the model convergence speed and eliminate the differences between data:
[0112]
[0113] Here, xscaled represents the normalized value. Further, axis = 0 represents row normalization. In this study, the trajectory data is arranged by rows, which means that one row represents a vehicle trajectory. Further, the mean squared error (MSE) is selected as the loss function. MSE is calculated as:
[0114]
[0115] Equation 13 describes the calculation method of the mean squared error (MSE) for measuring the difference between the prediction result and the true result. The meanings of the parameters are as follows:
[0116] MSE: Mean squared error.
[0117] y i real : The true result.
[0118] y i pre : The prediction result.
[0119] m: The number of samples in the dataset.
[0120] From the calculation method of the mean square error, the difference between the prediction result and the true result can be known. It can be concluded that the deep structure can better process complex trajectory data, and the attention mechanism can also focus on key information to improve the prediction accuracy.
[0121] Step 3: Trajectory correction:
[0122] Correct the prediction result of the deep encoder-decoder network to improve the prediction accuracy.
[0123] First, form a vehicle state matrix according to the motion state of the vehicle on different road sections as the input of the DNN; then, calculate the trajectory prediction error between the actual trajectory and the predicted trajectory of the vehicle to form a trajectory error matrix as the DNN label. After the DNN training is completed, calculate the correction matrix using the new vehicle state matrix, and use the correction matrix to correct the predicted trajectory. To obtain the optimal structure of the DNN, we used the structure layers of 3-6 layers of DNN for experiments. The number of layers of the DNN is up to 6 layers because the DNN with more than 6 layers is prone to gradient dispersion, which causes the DNN to fail to converge. After determining the optimal structure of the DNN, use the new data to test the three-layer DNN and calculate the trajectory correction matrix for each trajectory.
[0124] Step 4: Model cascade verification:
[0125] Conduct cascade experiments on the deep encoder-decoder and DNN to evaluate the overall performance after cascading the deep encoder-decoder network and DNN. First, train and test the deep encoder-decoder network and DNN separately. Then, use the output of the deep encoder-decoder network as the input of the DNN, and use the output of the DNN as the final prediction result. Secondly, use the test data set to verify the overall performance of the model and compare it with other advanced prediction models.
[0126] The training and testing of the prediction model use the same data and processing methods as the previously reported experiments. The model performance metrics are also the same (i.e., the convergent mean square error, the calculation time of a single data, the number of convergent steps, and the prediction error). The performance comparison of different models on cyclic data is shown in Table 1:
[0127] Table 1
[0128]
[0129] Table 1 gives the performance results of the cascade prediction model under three different road section types. It can be seen that the cascade prediction model can well predict future trajectories, and there is a slight performance improvement compared with a single neural network. Figures 6 - 8Shows the comparison between the results of the cascaded model and the actual trajectory. It can be seen that for all three types of road sections, the lateral and longitudinal trajectories predicted by the cascaded model are highly consistent with the actual trajectory. Therefore, the cascading of the prediction model improves the model performance. To further verify the model, the errors of the cascaded model are presented and analyzed in different forms. Figures 6 - 8 Ten trajectories were extracted from the test dataset, and the fitting degree of some trajectories is shown in the figure. The error of the predicted trajectory is small in the first stage and gradually increases as the trajectory extends. However, all errors are within an acceptable range, especially the longitudinal estimation, which is controlled at the lane level. These results indicate that the cascaded prediction model can predict the trajectories of all three types of road sections with high accuracy.
[0130] In summary, the present invention can effectively extract features in trajectory data and make accurate long-term predictions by introducing an attention mechanism and a deep encoder-decoder network structure; it can achieve a longer prediction time and meet the actual application requirements by gradually reducing the number of input data coordinates and increasing the number of output data coordinates; it can adapt to different road scenarios such as highways, intersections, and roundabouts and has good generalization ability by using various road scenario data for training; it can effectively eliminate prediction errors and inverse normalization errors and make the predicted trajectory closer to the real trajectory by correcting the prediction results with DNN; therefore, the present invention has the advantages of high prediction accuracy, strong long-term prediction ability, strong multi-scenario adaptability, good correction effect, and strong real-time performance.
[0131] It should be emphasized that the above are only the preferred embodiments of the present invention, and there is no any form of limitation to the present invention. Any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.
Claims
1. An on-vehicle trajectory prediction and correction method based on an encoder-decoder neural network, characterized in that, It includes the following steps: Step 1: Data preprocessing. Use the head and tail sampling algorithm to process the original trajectory data, convert trajectory data of different lengths into data of equal length, and gradually reduce the number of input data coordinates and increase the number of output data coordinates to extend the prediction time. Step 2: Construction of a deep encoder-decoder network. Expand the deep encoder-decoder network structure, introduce an attention mechanism for semantic recognition, and determine the optimal network structure through multi-parameter comparison experiments to predict the future trajectory of the vehicle. Step 3: Trajectory correction. Use a deep neural network to correct the prediction results of the deep encoder-decoder network structure. The input of the deep neural network includes vehicle motion state data, and the output of the deep neural network is a trajectory correction matrix. Step 4: Model cascade verification. Cascade the deep encoder-decoder network and the deep neural network to construct an in-vehicle trajectory prediction and correction model, verify the overall performance of the model through cascade experiments, and compare it with other prediction models.
2. The vehicle trajectory prediction and correction method based on an encoder-decoder neural network according to claim 1, wherein In the said Step 1, the specific steps of converting trajectory data of different lengths into data of equal length include: Represent the original trajectory data as a vector x of fixed dimension = {x1, x2, ……, x n}, and calculate the sampling points sp1, sp2 …… sp t as follows: And there is n ∈ Z t ∈ Z (2) In Formula (1) and Formula (2), x represents a vector of a fixed dimension, n represents the number of trajectory coordinates, sp1, sp2, ..., sp t represent sampling points, t represents the number of sampling times, and Z is the set of positive integers.
3. The on-vehicle trajectory prediction and correction method based on an encoder-decoder neural network according to claim 1, wherein In the said Steps 1-3, the generated different vehicle motion state data do not include the vehicle trajectory.
4. The on-vehicle trajectory prediction and correction method based on an encoder-decoder neural network according to claim 2, wherein In the said Step 2, the expansion of the deep encoder-decoder network structure specifically includes: The input trajectory data x = {x1, x2, …, x n} is input into the encoder, and the hidden state ht at time t is calculated as follows: h t = f(x t , h t-1 ) (3) Use the attention mechanism to convert the hidden state ht into a semantic vector c of a fixed dimension. c = q(h1, h2, ……, h t ) (4) Use the semantic vector C = {C1, C2,..., C n} to decode the decoder and calculate the prediction result at time t, where S t represents the state of the decoder hidden layer at time t: P(y t |y1,y2,...,y t-1 ,c) = g(y t-1 ,s t ,c) (6) In formula (5) and formula (6), P(y) represents the probability distribution of the future trajectory y, P(y t |y1,y2,...,y t-1 , c) represents the prediction y at time t t The probability of y t-1 Represents the prediction result at time t-1.
5. The on-vehicle trajectory prediction and correction method based on an encoder-decoder neural network according to claim 4, characterized in that, In the said Step 2, the introduction of the attention mechanism for semantic recognition specifically includes: Add attention to the i-th layer. The conditional probability of the future trajectory y at time t is: P(y t |y1, y2..., y t-1 , x) = g(y t-1 , s t , c t ) (7) s t = f(s t-1 y t-1 C t ) (8) e tj = score(s t-1 , h j ) = Vσ(W s-1 + Uh j ) (11) In formulas (7) to (11), x represents the input data, and c t [[ID=ST2]]represents the attention vector, a tj represents the weight at time t, h j represents the state of the j-th hidden layer, e tj represents the relevance score of the j-th part of the input data x at time t, score represents the scoring function, s t-1 represents the decoder hidden state at time t - 1, V, W, and U respectively represent the weight matrices to be trained, and σ represents the activation function.
6. The on-vehicle trajectory prediction and correction method based on an encoder-decoder neural network according to claim 5, wherein In the said Step 2, the determination of the optimal network structure specifically includes: Increase the number of label data coordinates in different training steps; before training, normalize the data through the MinMaxScaler function to accelerate the model convergence speed and eliminate the differences between data: In Equation (12), xscaled represents the normalized value, and axis = 0 represents row normalization; select the mean square error MSE as the loss function. The mean square error MSE is: In formula (13), y i real represents the true result, and y i pre represents the predicted result, where m represents the number of samples in the dataset.
7. The vehicle trajectory prediction and correction method based on an encoder-decoder neural network according to claim 1, wherein The said Step 3 specifically includes: Step 3-1: Form a vehicle state matrix according to the vehicle motion state in different sections as the input of the deep neural network. Step 3-2: Calculate the trajectory prediction error between the actual trajectory and the predicted trajectory of the vehicle to form a trajectory error matrix as the label of the deep neural network. Step 3-3: After the deep neural network training is completed, calculate the correction matrix using the new vehicle state matrix, and use the correction matrix to correct the predicted trajectory.
8. The on-vehicle trajectory prediction and correction method based on an encoder-decoder neural network according to claim 7, wherein, In the said Step 3-3, a structure layer of a 3-6 layer deep neural network is used for training.
9. The vehicle trajectory prediction and correction method based on an encoder-decoder neural network according to claim 1, characterized in that The said Step 4 specifically includes: Step 4-1: Train and test the deep encoder-decoder network and the deep neural network respectively. Step 4-2: Use the output of the deep encoder-decoder network as the input of the deep neural network, and use the output of the deep neural network as the final prediction result. Step 4-3: Use the test data set to verify the overall performance of the model and compare it with other prediction models.