A vehicle trajectory prediction method using Markov chain fusion LSTM and GPR

Abstract

The present invention relates to a vehicle trajectory prediction method using Markov chain to fuse LSTM and GPR, the method comprising the following steps: firstly, processing the original data set; constructing an LSTM network model, and performing trajectory prediction according to the LSTM network model; constructing a Gaussian process regression model, and performing trajectory prediction based on the Gaussian process regression model; finally, in each time step of the future predicted trajectory, the prediction results of the LSTM network model and the Gaussian process regression model are fused using the transition probability matrix trained by the Markov chain, so as to obtain the predicted trajectory of the target vehicle under the influence of multiple factors. The present invention gives full play to the advantages of the LSTM neural network in short-term prediction, and at the same time considers the influence of vehicle condition information and road condition information on the prediction of the future trajectory of the vehicle in combination with GPR. The prediction results of the two are effectively fused through the Markov chain, which further improves the accuracy and flexibility of trajectory prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent vehicle trajectory prediction, and in particular to a vehicle trajectory prediction method using a Markov chain fusing LSTM and GPR. Background Art

[0002] LSTM has been shown to have excellent ability in extracting time series data and is therefore widely used in trajectory prediction tasks for autonomous vehicles. Although LSTM has high accuracy in short-term predictions, it is not very effective in long-term predictions due to its limited memory capacity, gradient decay and gradient explosion.

[0003] Currently, most of the work that relies on LSTM for vehicle trajectory prediction only uses the historical position information of the vehicle (world coordinate system or polar coordinate system), speed information and acceleration information To predict the future short-term trajectory of the vehicle, although this information can characterize the vehicle's motion characteristics to a certain extent, in actual driving scenarios, non-time series data such as vehicle condition information and road condition information also have a very large impact on the future trajectory of the intelligent vehicle.

[0004] To this end, the present invention provides a vehicle trajectory prediction method that integrates LSTM and GPR to solve the above problems. Summary of the invention

[0005] The present invention provides a vehicle trajectory prediction method using Markov chain to fuse LSTM and GPR. Through GPR, not only the position, speed and acceleration information of the intelligent vehicle are considered, but also the relevant vehicle condition information and road condition information are introduced, and the uncertainty of trajectory prediction can be quantified. Finally, the probability of the prediction results of the two is allocated through the Markov chain, so that the final prediction can fully utilize the high-precision advantage of LSTM in short-term prediction, and fully consider the impact of vehicle condition and road condition information on future trajectory.

[0006] Terminology explanation:

[0007] Trajectory Prediction: In this article, trajectory prediction refers to predicting the movement trajectory of traffic vehicles in the future. This process usually involves predicting the position and state of the vehicle at different points in time. Trajectory prediction not only has important applications in the field of autonomous driving, but also plays a key role in intelligent transportation systems, robot navigation, and drone control.

[0008] LSTM: Long Short-Term Memory (LSTM) is a special type of recurrent neural network (RNN), which is often used to deal with problems related to time series data. The state at each moment is affected by the current input state and the previous input state, and has a memory effect.

[0009] GPR: Gaussian Process Regression (GPR) is a non-parametric regression method that uses Gaussian processes to model and predict data. Unlike traditional regression methods, GPR does not make any assumptions about the data, but instead defines a kernel function to measure the similarity between data points to infer the distribution of the target variable.

[0010] Markov Chain: Markov Chain is a random process with Markov properties. It is usually used to model a series of state sequences that develop according to a random process. The transition probability between these states depends only on the current state and has nothing to do with time.

[0011] A vehicle trajectory prediction method using Markov chain fusion LSTM and GPR, the method mainly consists of a data processing module, an LSTM neural network prediction module, a GPR prediction module and a Markov chain training module, and the method includes the following steps:

[0012] First, the original data set is processed;

[0013] Constructing an LSTM network model, and performing trajectory prediction based on the LSTM network model;

[0014] Constructing a Gaussian process regression model, and performing trajectory prediction based on the Gaussian process regression model;

[0015] Finally, in each time step of the future predicted trajectory, the prediction results of the LSTM network model and the Gaussian process regression model are fused using the transition probability matrix trained by the Markov chain to obtain the predicted trajectory of the target vehicle considering multiple factors.

[0016] As a further technical solution of the present invention, the original data set includes time series characteristic driving trajectory data, vehicle condition information and road condition information, wherein the time series characteristic driving trajectory data includes the vehicle's historical position, speed and acceleration, the vehicle condition information is the vehicle's own attribute information, including vehicle length, vehicle width and vehicle heading angle, and the road condition information includes traffic lights, weather conditions and road surface conditions.

[0017] As a further technical solution of the present invention, the step of processing the time series characteristic driving trajectory data includes:

[0018] 1.1 Data cleaning: remove missing values ​​and outliers in time series data;

[0019] 1.2 Data standardization: Normalize or standardize the data to make it suitable for wavelet analysis;

[0020] 1.3 Wavelet transform: Perform wavelet transform on the standardized data, select appropriate wavelet basis functions and decomposition layers to decompose the parts of different frequency domains in the data;

[0021]

[0022] It is a set of functions generated by the displacement and expansion of the mother wavelet, called wavelet basis function. a is the scale factor, which realizes the expansion and contraction transformation of the mother wavelet function, and b realizes the displacement transformation of the mother wavelet function.

[0023] 1.4 Select a suitable threshold (soft threshold or hard threshold is acceptable) and remove the noise components in the wavelet coefficients through threshold processing;

[0024] 1.5 Reconstruct the data through inverse wavelet transform so as to reconstruct the denoised wavelet coefficients into time series data;

[0025] 1.6 Post-process the denoised data to make it suitable for subsequent analysis and modeling;

[0026] As a further technical solution of the present invention, the step of data processing of the vehicle condition information and the road condition information includes:

[0027] 2.1 Observe and record the road condition information of the target driving area;

[0028] 2.2 Synchronize the time series, align the road condition information with the driving trajectory data recorded in the first phase of observation in the time domain to ensure that there is corresponding road condition information at each time point;

[0029] 2.3 Feature encoding, convert the road condition variables into numerical features, encode the weather conditions as sunny = 0, rainy = 1, snowy = 2; encode the road conditions as dry road surface = 3, water on the road surface = 4, snow on the road surface = 5; encode the traffic light information as red light = 6, green light = 7, yellow light = 8; in this way, a variety of environmental conditions that may affect the future trajectory of the vehicle can be considered.

[0030] As a further technical solution of the present invention, the step of constructing an LSTM network model and performing trajectory prediction according to the LSTM network model includes:

[0031] Build the LSTM model:

[0032]

[0033] in, is the output of the forget gate, is the output of the input gate, is the output of the output gate, It is a sigmoid activation function with an output range between 0 and 1, which determines the degree to which past information needs to be forgotten. It is also an activation function, whose output range is between -1 and 1, and is used to capture small changes in data. is the weight parameter of the forget gate, is the input gate weight parameter, is the memory state weight parameter, is the output gate weight parameter, is the hidden state at the previous time step, is the hidden state at the current time step, is the input feature information, is the bias parameter of the forget gate, is the bias parameter of the input gate, is the memory state bias parameter, is the output gate bias parameter, is a candidate memory state, and They are the memory state of the current time step and the memory state of the previous time step respectively;

[0034] Input the time series data of historical time step vehicles into the LSTM network:

[0035]

[0036] in, Vehicle time series data representing historical time steps; Represents the state information of target vehicle i at time t, including the position, speed and acceleration information at this time; i is the number of the target vehicle; is the time step; is the x-coordinate of the target vehicle's position at time t, is the y coordinate of the target vehicle position at time t; The target vehicle is Speed ​​in direction, The target vehicle is Speed ​​in direction; The target vehicle is The acceleration in the direction, The target vehicle is acceleration in direction;

[0037] ;

[0038] ;

[0039] in, Represents the target vehicle trajectory and speed information for the predicted future time step; is the prediction time step; Represents the predicted trajectory and speed information of the target vehicle at each time step; m represents a time step within the future prediction time step; is the x-coordinate of the target vehicle’s position at time m, is the y coordinate of the target vehicle position at time m; The target vehicle is Speed ​​in direction, The target vehicle is Speed ​​in direction.

[0040] As a further technical solution of the present invention, the step of constructing a Gaussian process regression model and performing trajectory prediction based on the Gaussian process regression model includes:

[0041] Define the input and output of Gaussian process regression (GPR). The input is defined as the time series features and related environmental feature data in the complete observation data after encoding and processing in step 1, and the output is the future trajectory and speed of the vehicle. ;

[0042] Formula 13: ;

[0043] ;

[0044] ;

[0045] ;

[0046] Formula 17: ;

[0047] ;

[0048] in, Represents all data related to the target vehicle in the historical time step of the complete observation data; Indicates that at time T (Note: The same meaning, only distinguishing between time series feature data and complete observation data) (Note: The meaning is the same as that of the state information of the time series feature data and the complete observation data); and They are heading angle, vehicle width, vehicle length, encoded weather, road conditions and traffic light information; Represents the target vehicle trajectory and velocity information for predicted future time steps; is the prediction time step; Represents the predicted trajectory and speed information of the target vehicle at each time step; is the x-coordinate of the target vehicle position at the future time M predicted by the GPR method, is the y coordinate of the target vehicle position at time M predicted by the GPR method; is the target vehicle at time M predicted by the GPR method. Speed ​​in direction, is the target vehicle at time M predicted by the GPR method. Speed ​​in direction;

[0049] Choose an appropriate Gaussian kernel function , construct the covariance matrix ,Commonly used kernel functions include radial basis function (RBF) kernel, linear kernel and Matern kernel.

[0050] Train the GPR model and use maximum likelihood estimation (MLE) to optimize the hyperparameters of the kernel function:

[0051] ;

[0052] in, is the input feature data matrix, is the output value vector, is the covariance matrix, is a constant used to calculate the normalization factor of the likelihood function;

[0053] For the input data, make a prediction and compute its predicted distribution.

[0054] As a further technical solution of the present invention, the step of finally fusing the prediction results of the LSTM network model and the Gaussian process regression model using the transition probability matrix trained by the Markov chain in each time step of the future predicted trajectory to obtain the predicted trajectory of the target vehicle under the influence of multiple factors includes:

[0055] Set the LSTM network model to predict the future time step of the target vehicle The prediction information is , and the Gaussian process regression model is used to predict the future time step of the target vehicle The result is , then the state transition probability matrix between LSTM and GPR prediction can be defined as:

[0056] ;

[0057] Among them, taking the first state transfer as an example, Indicates the probability that at a certain prediction time step in the LSTM prediction result, the prediction value of the next time step still adopts the value in the LSTM prediction result; is the probability that the predicted value of the next time step will adopt the value in the GPR prediction result at a certain prediction time step in the LSTM prediction result; is the probability that the predicted value of the next time step will adopt the value in the LSTM prediction result at a certain prediction time step in the GPR prediction result; The probability that the predicted value of the next time step still uses the value in the GPR prediction result at a certain prediction time step in the GPR prediction result;

[0058] In the process of predicting the future trajectory of the target vehicle, as the state transition probability matrix is ​​continuously trained and updated, the data of each prediction time step will be flexibly selected between the LSTM and GPR methods based on the time series data as well as the vehicle condition information and road condition information.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention gives full play to the advantages of LSTM neural network in short-term prediction, and at the same time considers the influence of vehicle condition information and road condition information on the prediction of future vehicle trajectory in combination with GPR. The prediction results of the two are effectively integrated through the Markov chain, further improving the accuracy and flexibility of trajectory prediction. The present invention not only demonstrates the excellent performance of LSTM neural network in short-term prediction, but also makes up for the shortcomings of LSTM in considering complex external factors through the processing of non-time series data by GPR. The overall method shows higher robustness and adaptability in complex and changeable traffic scenarios, thereby achieving more accurate vehicle trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 The overall framework diagram of the LSTM and GPR trajectory prediction method based on Markov chain fusion provided in an embodiment of the present invention.

[0061] Figure 2 The LSTM network model provided by the embodiment of the present invention.

[0062] Figure 3 A diagram of the Markov chain fusion state transition process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The technical solution of the present invention is further described in detail below in conjunction with specific implementation methods.

[0064] In conjunction with the accompanying drawings Figures 1 to 3 , a specific implementation case is introduced through detailed description, and the process of the present invention is further described. It is particularly noted that unless otherwise specifically and detailed, the operation flow, presentation form and numerical calculation of the nodes and implementation steps described in the embodiments of the present invention do not limit the scope of the present disclosure. The specific embodiments described in the present invention are only used to explain the present invention and are not used to limit the present invention.

[0065] A vehicle trajectory prediction method using Markov chain fusion LSTM and GPR, such as Figure 1 As shown, the method is mainly composed of a data processing module, an LSTM neural network prediction module, a GPR prediction module and a Markov chain training module, and the method includes the following steps:

[0066] 1. First, process the original data set;

[0067] The original data set includes time series characteristic driving trajectory data, vehicle condition information and road condition information. The time series characteristic driving trajectory data includes the vehicle's historical position, speed and acceleration. The vehicle condition information is the vehicle's own attribute information, including vehicle length, vehicle width and vehicle heading angle. The road condition information includes traffic lights, weather conditions and road surface conditions.

[0068] The step of processing the time series characteristic driving trajectory data comprises:

[0069] 1.1 Data cleaning: remove missing values ​​and outliers in time series data;

[0070] 1.2 Data standardization: Normalize or standardize the data to make it suitable for wavelet analysis;

[0071] 1.3 Wavelet transform: Perform wavelet transform on the standardized data, select appropriate wavelet basis functions and decomposition layers to decompose the parts of different frequency domains in the data. Common wavelet basis functions are shown in Table 1;

[0072]

[0073] It is a set of functions generated by the displacement and expansion of the mother wavelet, called wavelet basis function. a is the scale factor, which realizes the expansion and contraction transformation of the mother wavelet function, and b realizes the displacement transformation of the mother wavelet function.

[0074] 1.4 Select a suitable threshold (soft threshold or hard threshold is acceptable) and remove the noise components in the wavelet coefficients through threshold processing;

[0075] 1.5 Reconstruct the data through inverse wavelet transform so as to reconstruct the denoised wavelet coefficients into time series data;

[0076] 1.6 Post-process the denoised data to make it suitable for subsequent analysis and modeling;

[0077] The step of data processing of the vehicle condition information and the road condition information comprises:

[0078] 2.1 Observe and record the road condition information of the target driving area;

[0079] 2.2 Synchronize the time series, align the road condition information with the driving trajectory data recorded in the first phase of observation in the time domain to ensure that there is corresponding road condition information at each time point;

[0080] 2.3 Feature encoding, convert the road condition variables into numerical features, encode the weather conditions as sunny = 0, rainy = 1, snowy = 2; encode the road conditions as dry road surface = 3, water on the road surface = 4, snow on the road surface = 5; encode the traffic light information as red light = 6, green light = 7, yellow light = 8; in this way, a variety of environmental conditions that may affect the future trajectory of the vehicle can be considered.

[0081] 2. Build an LSTM network model and perform trajectory prediction based on the LSTM network model;

[0082] Build the LSTM model:

[0083]

[0084]

[0085] in, is the output of the forget gate, is the output of the input gate, is the output of the output gate, It is a sigmoid activation function with an output range between 0 and 1, which determines the degree to which past information needs to be forgotten. It is also an activation function, whose output range is between -1 and 1, and is used to capture small changes in data. is the weight parameter of the forget gate, is the input gate weight parameter, is the memory state weight parameter, is the output gate weight parameter, is the hidden state at the previous time step, is the hidden state at the current time step, is the input feature information, is the bias parameter of the forget gate, is the bias parameter of the input gate, is the memory state bias parameter, is the output gate bias parameter, is a candidate memory state, and are the memory state of the current time step and the memory state of the previous time step respectively; the LSTM network model is as follows Figure 2 As shown;

[0086] Input the time series data of historical time step vehicles into the LSTM network:

[0087]

[0088] in, Vehicle time series data representing historical time steps; Represents the state information of target vehicle i at time t, including the position, speed and acceleration information at this time; i is the number of the target vehicle; is the time step; is the x-coordinate of the target vehicle's position at time t, is the y coordinate of the target vehicle position at time t; The target vehicle is Speed ​​in direction, The target vehicle is Speed ​​in direction; The target vehicle is The acceleration in the direction, The target vehicle is acceleration in direction;

[0089] ;

[0090] ;

[0091] in, Represents the target vehicle trajectory and speed information for the predicted future time step; is the prediction time step; Represents the predicted trajectory and speed information of the target vehicle at each time step. m represents a time step within the future prediction time step; is the x-coordinate of the target vehicle’s position at time m, is the y coordinate of the target vehicle position at time m; The target vehicle is Speed ​​in direction, The target vehicle is Speed ​​in direction.

[0092] 3. Constructing a Gaussian process regression model, and performing trajectory prediction based on the Gaussian process regression model;

[0093] Define the input and output of Gaussian process regression (GPR). The input is defined as the time series features and related environmental feature data in the complete observation data after encoding and processing in step 1, and the output is the future trajectory and speed of the vehicle. ;

[0094] Formula 13: ;

[0095] ;

[0096] ;

[0097] ;

[0098] Formula 17: ;

[0099] ;

[0100] in, Represents all data related to the target vehicle in the historical time step of the complete observation data; Indicates that at time T (Note: The same meaning, only distinguishing between time series feature data and complete observation data) (Note: The meaning is the same as that of the state information of the time series feature data and the complete observation data); and They are (vehicle condition information) heading angle, vehicle width, vehicle length, encoded weather, road conditions and traffic light information; Represents the target vehicle trajectory and velocity information for predicted future time steps; is the prediction time step; Represents the predicted trajectory and speed information of the target vehicle at each time step; is the x-coordinate of the target vehicle position at the future time M predicted by the GPR method, is the y coordinate of the target vehicle position at time M predicted by the GPR method; is the target vehicle at time M predicted by the GPR method. Speed ​​in direction, is the target vehicle at time M predicted by the GPR method. Speed ​​in direction.

[0101] Select an appropriate Gaussian kernel function and construct the covariance matrix ,Commonly used kernel functions include radial basis function (RBF) kernel, linear kernel and Matern kernel.

[0102] Train the GPR model and use maximum likelihood estimation (MLE) to optimize the hyperparameters of the kernel function:

[0103] ;

[0104] in, is the input feature data matrix, is the output value vector, is the covariance matrix, is a constant used to calculate the normalization factor of the likelihood function;

[0105] For the input data, make a prediction and compute its predicted distribution.

[0106] 4. Finally, in each time step of the future predicted trajectory, the prediction results of the LSTM network model and the Gaussian process regression model are fused using the transition probability matrix trained by the Markov chain to obtain the predicted trajectory of the target vehicle considering multiple factors.

[0107] Markov Chain is a random process used to describe the transition of a system between different states. Its characteristic is that the future state of the system depends only on the current state and is not affected by the past state. This property is called "Markov property" or "no aftereffect".

[0108] The present invention proposes a prediction architecture that uses LSTM and GPR in parallel for time series data sets and comprehensive data sets. After obtaining the prediction results of LSTM and GPR, the state transition probability matrix is ​​trained through Markov chain, thereby achieving an effective fusion of the two prediction methods.

[0109] According to the above introduction, set the LSTM network model to predict the future time step of the target vehicle The prediction information is , and the Gaussian process regression model is used to predict the future time step of the target vehicle The result is , define the state transition probability matrix between LSTM and GPR prediction as:

[0110] ;

[0111] State transitions such as Figure 3 As shown: Taking the first state transfer as an example, Indicates the probability that at a certain prediction time step in the LSTM prediction result, the prediction value of the next time step still adopts the value in the LSTM prediction result; is the probability that the predicted value of the next time step will adopt the value in the GPR prediction result at a certain prediction time step in the LSTM prediction result; is the probability that the predicted value of the next time step will adopt the value in the LSTM prediction result at a certain prediction time step in the GPR prediction result; The probability that the predicted value of the next time step still uses the value in the GPR prediction result at a certain prediction time step in the GPR prediction result;

[0112] In the process of predicting the future trajectory of the target vehicle, as the state transition probability matrix is ​​continuously trained and updated, the data of each prediction time step will be flexibly selected between the LSTM and GPR methods based on the time series data as well as the vehicle condition information and road condition information.

[0113] In summary, the highlights of the trajectory prediction method architecture design proposed in the present invention are:

[0114] 1. When predicting the trajectory of the target vehicle, not only the influence of position, speed and acceleration information on the future trajectory is considered, but also the influence of vehicle attribute information (such as width, length and heading angle) and environmental information (such as weather conditions, road conditions and traffic light status) on the prediction accuracy of the vehicle's future trajectory, which makes the prediction more accurate.

[0115] 2. The present invention adopts the state transition probability matrix method of Markov chain to fuse the prediction results of LSTM and GPR based on parallel architecture. The information of each time step in the future trajectory of the target vehicle predicted by LSTM and GPR is regarded as a state, and the prediction results of LSTM and GPR are cleverly integrated through the state transition probability matrix. This fusion strategy not only makes full use of the advantages of LSTM in short-term trajectory prediction, but also considers multiple factors through GPR, significantly improving the accuracy of trajectory prediction. Especially in the long-term trajectory prediction, although the accuracy of LSTM is slightly inferior to that of GPR, the organic combination of the two effectively makes up for the disadvantages of LSTM in long-term prediction, and comprehensively improves the accuracy of trajectory prediction.

[0116] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0117] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A vehicle trajectory prediction method using Markov chain fusion LSTM and GPR, characterized in that: The method comprises the following steps: First, the original data set is processed; Constructing an LSTM network model, and performing trajectory prediction based on the LSTM network model; Constructing a Gaussian process regression model, and performing trajectory prediction based on the Gaussian process regression model; Finally, in each time step of the future predicted trajectory, the prediction results of the LSTM network model and the Gaussian process regression model are fused using the transition probability matrix trained by the Markov chain to obtain the predicted trajectory of the target vehicle considering multiple factors.

2. A vehicle trajectory prediction method using Markov chain fusion LSTM and GPR according to claim 1, characterized in that: The original data set includes time series characteristic driving trajectory data, vehicle condition information and road condition information, wherein the time series characteristic driving trajectory data includes the vehicle's historical position, speed and acceleration, the vehicle condition information is the vehicle's own attribute information, including vehicle length, vehicle width and vehicle heading angle, and the road condition information includes traffic lights, weather conditions and road surface conditions.

3. A vehicle trajectory prediction method using Markov chain fusion LSTM and GPR according to claim 2, characterized in that: The step of processing the time series characteristic driving trajectory data comprises: Data cleaning: remove missing values ​​and outliers from time series data; Data standardization: normalize or standardize the data to make it suitable for wavelet analysis; Wavelet transform: Perform wavelet transform on the standardized data, select appropriate wavelet basis functions and decomposition layers to decompose the parts of different frequency domains in the data; Formula 1: ; It is a set of functions generated by displacement and expansion of the mother wavelet, called wavelet basis function, a is the scale factor, and b realizes the displacement transformation of the mother wavelet function; Select a suitable threshold and remove the noise components in the wavelet coefficients through threshold processing; Reconstruct the data through inverse wavelet transform so as to reconstruct the denoised wavelet coefficients into time series data; The denoised data is post-processed to make it suitable for subsequent analysis and modeling.

4. A vehicle trajectory prediction method using Markov chain fusion LSTM and GPR according to claim 2, characterized in that: The step of data processing of the vehicle condition information and the road condition information comprises: Observe and record the road condition information of the target driving area; Synchronize the time series and align the road condition information with the driving trajectory data recorded in the first phase of observation in the time domain to ensure that there is corresponding road condition information at each time point; Feature encoding, converting road condition variables into numerical features, encoding weather conditions as sunny = 0, rainy = 1, snowy = 2; encoding road conditions as dry road surface = 3, water on the road surface = 4, snow on the road surface = 5; encoding traffic light information as red light = 6, green light = 7, yellow light = 8.

5. The vehicle trajectory prediction method using Markov chain fusion LSTM and GPR according to claim 1 is characterized in that: The step of constructing an LSTM network model and performing trajectory prediction according to the LSTM network model comprises: Build the LSTM model: Formula 2: ; ； ； ； ； ； in, is the output of the forget gate, is the output of the input gate, is the output of the output gate, It is a sigmoid activation function with an output range between 0 and 1, which determines the degree to which past information needs to be forgotten. It is also an activation function, whose output range is between -1 and 1, and is used to capture small changes in data. is the weight parameter of the forget gate, is the input gate weight parameter, is the memory state weight parameter, is the output gate weight parameter, is the hidden state at the previous time step, is the hidden state at the current time step, is the input feature information, is the bias parameter of the forget gate, is the bias parameter of the input gate, is the memory state bias parameter, is the output gate bias parameter, is a candidate memory state, and They are the memory state of the current time step and the memory state of the previous time step respectively; Input the time series data of historical time step vehicles into the LSTM network: ； ； in, Vehicle time series data representing historical time steps; Represents the state information of target vehicle i at time t, including the position, speed and acceleration information at this time; i is the number of the target vehicle; is the time step; is the x-coordinate of the target vehicle position at time t, is the y coordinate of the target vehicle position at time t; The target vehicle is Speed ​​in direction, The target vehicle is Speed ​​in direction; The target vehicle is The acceleration in the direction, The target vehicle is acceleration in direction; ； ； in, Represents the target vehicle trajectory and speed information for the predicted future time step; is the prediction time step; Represents the predicted trajectory and speed information of the target vehicle at each time step; m represents a time step within the future prediction time step; is the x-coordinate of the target vehicle’s position at time m, is the y coordinate of the target vehicle position at time m; The target vehicle is Speed ​​in direction, The target vehicle is Speed ​​in direction.

6. The vehicle trajectory prediction method using Markov chain fusion LSTM and GPR according to claim 1 is characterized in that: The step of constructing a Gaussian process regression model and performing trajectory prediction based on the Gaussian process regression model comprises: Define the inputs and outputs of Gaussian process regression; ； Formula 13: ; ； ； ； Formula 17: ; ； in, Represents all data related to the target vehicle in the historical time step of the complete observation data; Indicates the target vehicle at time T Status information; and They are heading angle, vehicle width, vehicle length, encoded weather, road conditions and traffic light information; Represents the target vehicle trajectory and velocity information for the predicted future time step; is the prediction time step; Represents the predicted trajectory and speed information of the target vehicle at each time step; is the x-coordinate of the target vehicle position at the future time M predicted by the GPR method, is the y coordinate of the target vehicle position at time M predicted by the GPR method; is the target vehicle at time M predicted by the GPR method. Speed ​​in direction, is the target vehicle at time M predicted by the GPR method. Speed ​​in direction; Select Gaussian kernel function and construct covariance matrix ,Commonly used kernel functions include radial basis function kernel, linear kernel and Matern kernel; Train the GPR model and use maximum likelihood estimation to optimize the kernel function hyperparameters: ； in, is the input feature data matrix, is the output value vector, is the covariance matrix, is a constant used to calculate the normalization factor of the likelihood function; For the input data, make a prediction and compute its predicted distribution.

7. The vehicle trajectory prediction method using Markov chain fusion LSTM and GPR according to claim 1 is characterized in that: The step of finally fusing the prediction results of the LSTM network model and the Gaussian process regression model using the transition probability matrix trained by the Markov chain in each time step of the future predicted trajectory to obtain the predicted trajectory of the target vehicle under the influence of multiple factors includes: Set the LSTM network model to predict the future time step of the target vehicle The prediction information is , and the Gaussian process regression model is used to predict the future time step of the target vehicle The result is , define the state transition probability matrix between LSTM and GPR prediction as: ； in, Indicates the probability that at a certain prediction time step in the LSTM prediction result, the prediction value of the next time step still adopts the value in the LSTM prediction result; is the probability that the predicted value of the next time step will adopt the value in the GPR prediction result at a certain prediction time step in the LSTM prediction result; is the probability that the predicted value of the next time step will adopt the value in the LSTM prediction result at a certain prediction time step in the GPR prediction result; The probability that the predicted value of the next time step still uses the value in the GPR prediction result at a certain prediction time step in the GPR prediction result; In the process of predicting the future trajectory of the target vehicle, as the state transition probability matrix is ​​continuously trained and updated, the data of each prediction time step will be flexibly selected between the LSTM and GPR methods based on the time series data as well as the vehicle condition information and road condition information.

Images