An online trajectory prediction method for uncertain and incomplete data

Abstract

The application discloses an online trajectory prediction method for uncertain and incomplete data. The method comprises the following steps: obtaining historical geographic position information of a target to generate historical trajectory data; performing data preprocessing to obtain preprocessed data; constructing a DoubleMLP model, inputting the preprocessed data into training, processing based on a CVAE method, and obtaining a trained model; inputting the historical trajectory data of a target to be predicted into the model after data preprocessing, processing based on the CVAE method, and outputting a future trajectory; performing coordinate conversion to obtain predicted geographic position information, and completing online trajectory prediction. The method can preprocess noisy and missing data under uncertain and incomplete sensing conditions, so that the trajectory prediction has high reliability, provides a reliable solution for multi-target trajectory real-time prediction, has high prediction real-time performance, can meet real-time prediction tasks in multiple scenes, and is suitable for fields such as traffic management and automatic driving.

Description

Technical Field

[0001] This invention relates to an online trajectory prediction method, specifically to an online trajectory prediction method for uncertain and incomplete data. Background Technology

[0002] In the field of autonomous driving technology, a deep understanding of various traffic participants (such as pedestrians and vehicles) and their behavior around autonomous driving systems has become indispensable. This is because autonomous driving systems not only need to perceive these targets but also predict their future movements in order to make safe, efficient, and traffic-compliant decisions. It plays a pivotal role in connecting environmental perception with path planning and is one of the key issues in autonomous driving technology and driverless vehicles. However, in the process of trajectory prediction, under uncertain and incomplete perception conditions, the data transmitted by sensors may contain noise or be missing. Noise refers to unwanted additional signals or interference, while missing data points may be lost or unavailable during the acquisition process. This greatly increases the challenge of trajectory prediction in autonomous driving or driverless vehicle systems.

[0003] Currently, there are various trajectory prediction methods. For example, the method for predicting target trajectory (publication number CN 115797892 A) extracts feature information from multiple dimensions, including time, space, and motion characteristics, based on the target's historical trajectory to achieve high-precision prediction of the target's future trajectory. Additionally, the method and apparatus for predicting trajectory (publication number CN 116740664A) predict the target's future trajectory using the path's feature vector, the path's interaction feature vector, and the feature vector of the historical trajectory. Furthermore, the method and apparatus for predicting motion trajectory (publication number CN 115690845A) utilizes the pedestrian features and corresponding motion features of the image sequence to be processed, along with the trajectory pedestrian bounding boxes, to predict the future trajectory of the target pedestrian, thus improving the accuracy of pedestrian trajectory prediction.

[0004] However, existing trajectory prediction methods have at least the following drawbacks: 1. In some cases, sensor perception conditions are uncertain or incomplete, and existing methods rarely take into account the presence of noise and missing data in the received data; 2. In some cases, the current methods are not real-time, making it difficult to achieve online prediction and unable to meet the real-time prediction tasks in multiple scenarios. Summary of the Invention

[0005] To address the problems existing in the background art, the present invention provides an online trajectory prediction method for uncertain and incomplete data.

[0006] The technical solution adopted in this invention is:

[0007] The online trajectory prediction method for uncertain and incomplete data of the present invention includes:

[0008] Step 1: Obtain historical geographic location information of several targets, thereby generating historical trajectory data for each target and sending it to the workbench via the User Datagram Protocol (UDP).

[0009] Step 2: After receiving the historical trajectory data of each target, the workbench performs data preprocessing to obtain the historical preprocessed trajectory data of each target.

[0010] Step 3: Construct the target trajectory prediction model DoubleMLP (Double Multilayer Perceptron). Input the historical preprocessed trajectory data of each target into the target trajectory prediction model DoubleMLP for training. At the same time, process the data based on the Conditional Variational Autoencoder (CVAE) method to learn the conditional distribution of the historical data and obtain the trained target trajectory prediction model DoubleMLP.

[0011] Step 4: Collect historical geographical location information of the target to be predicted, thereby generating historical trajectory data of the target to be predicted. After performing the same data preprocessing as in Step 2, the historical trajectory data of the target to be predicted is input into the trained target trajectory prediction model DoubleMLP. At the same time, it is processed based on the Conditional Variational Autoencoder (CVAE) method to generate trajectory data that meets the conditions. The target trajectory prediction model DoubleMLP outputs several future trajectories of the target to be predicted.

[0012] Step 5: After performing coordinate transformation on each future trajectory of the target to be predicted output in Step 4, obtain the predicted geographical location information of the target to be predicted, send the data to the workbench and receive it, and complete the online trajectory prediction.

[0013] In step one, the geographic location information includes the timestamps, latitude and longitude coordinates, and orientation of each geographic location of the target. The trajectory data is specifically generated based on each geographic location within a certain time period.

[0014] In step two, data preprocessing includes performing coordinate transformation, reducing or filtering data noise, filling in missing data points, format transformation, and Z-score standardization.

[0015] The coordinate transformation specifically involves converting the latitude and longitude coordinates in the historical geographical location information of the target's historical trajectory data into coordinates in a two-dimensional rectangular coordinate system, which facilitates subsequent data analysis and processing.

[0016] The reduction or filtering of data noise specifically employs the Kalman filtering method.

[0017] Reducing or filtering data noise refers to using Kalman filtering to reduce or remove noise in the presence of noise. By taking into account the relationship between observed data and system state, Kalman filtering can effectively filter out noise from sensors and improve data quality.

[0018] Kalman filtering is employed to remove noise and uncertainty from trajectory data. First, state variables are defined based on historical trajectory data, along with a state transition matrix F, an observation matrix H, an observation noise covariance matrix R, and a process noise covariance matrix Q. Then, the state estimates and covariance matrices of the Kalman filter are initialized. During prediction, at each time step, the state transition matrix is ​​used to predict the estimated value and covariance matrix of the next state. After receiving new observation data, the uncertainty of the observation data is considered to calculate updated state estimates and covariance matrices. These prediction and update steps are repeated to gradually improve the accuracy of trajectory estimation. Finally, the noise-filtered trajectory data is obtained from the state estimates.

[0019] Specifically, the missing data points are filled by using linear interpolation to fill in the missing data for targets moving in a straight line, and by using polynomial fitting to fill in the missing data for targets moving in a non-linear line. By fitting a polynomial model, the values ​​of the missing data points are estimated.

[0020] For targets with nonlinear motion, the coordinates of missing trajectory points are estimated by fitting the trajectory data into a quadratic curve.

[0021] The formal transformation specifically involves converting the target's historical trajectory data, after filling in the missing data points, into a relative position vector, as detailed below:

[0022]

[0023]

[0024]

[0025] Where ΔX is the relative position vector of the i-th target; ΔP i 1 Let be the position difference of the i-th target at time 2 and time 1. Let be the position difference of the i-th target at time t+1 and time t. Let $i$ be the position difference of the i-th target at time $obs+1$ and time $obs$. and These are the historical trajectory data of the i-th target at time t+1 and time t, respectively. and Let x and y be the x-axis coordinates and y-axis coordinates of the i-th target at time t in a two-dimensional rectangular coordinate system, respectively; R is a real number.

[0026] The historical time range of the observations is expressed as t1~t obs The predicted future time range is represented as t. obs+1 ~t obs+k Where, obs is the length of the historical trajectory, t obs Here, k is the current time, and k is the trajectory length to be predicted. This invention uses a relative vector to represent the input data and calculates the position difference between time t+1 and time t for the i-th target. The input for trajectory prediction is represented as Indicates the time interval from t1 to t2. obs The relative position vector of the i-th target within the observation period.

[0027] The input relative vector data is subjected to Z-score standardization, which involves vectorizing and standardizing the input historical trajectory data. Based on the target type, the mean μ and standard deviation σ of the relative position vectors are calculated, and then standardization is performed to normalize the input to a normal distribution.

[0028] In step three, the target trajectory prediction model DoubleMLP, based on the ResMLP fully connected neural network with residuals, includes a historical trajectory feature extraction module, an inter-frame deviation feature extraction module, a mean feature extraction module, and a future trajectory feature decoding module. The historical trajectory feature extraction module comprises four sequentially connected fully connected layers, the inter-frame deviation feature extraction module comprises five sequentially connected fully connected layers, the mean feature extraction module comprises three sequentially connected fully connected layers, and the future trajectory feature decoding module comprises five sequentially connected fully connected layers. The input to the DoubleMLP target trajectory prediction model is first processed in the historical trajectory feature extraction module. The outputs of the first and third fully connected layers in the historical trajectory feature extraction module are added to form a residual connection, which is then processed by the fourth fully connected layer and its output becomes the output of the historical trajectory feature extraction module. The output of the historical trajectory feature extraction module is concatenated with the inter-frame deviation noise Z and then input into the inter-frame deviation feature extraction module. In the inter-frame deviation feature extraction module, the outputs of the first and fourth fully connected layers are added to form a residual connection, which is then input into the fifth fully connected layer. The output of this residual connection serves as the output fully connected layer of the inter-frame deviation feature extraction module, i.e., the inter-frame average deviation prediction value fully connected layer. The output of the inter-frame deviation feature extraction module is then input into the mean feature extraction module. The outputs of the first and second fully connected layers in the mean feature extraction module are added to form a residual connection, which is then input into the third fully connected layer. The output of this residual connection serves as the output of the mean feature extraction module. The outputs of the historical trajectory feature extraction module and the mean feature extraction module are concatenated and then input into the future trajectory feature decoding module. The outputs of the first and fourth fully connected layers in the future trajectory feature decoding module are added to form a residual connection, which is then input into the fifth fully connected layer. The output of this residual connection serves as the output of the target trajectory prediction model, DoubleMLP.

[0029] The DoubleMLP target trajectory prediction model treats historical trajectory information as a complete block, extracts features from historical trajectories, and then reconstructs future trajectory features to achieve trajectory prediction. The predicted future trajectory features are obtained by combining historical trajectory features with the mean inter-frame deviation features. On one hand, the standardized historical trajectory relative vector sequence is input and processed by the ResMLP network to extract features, obtaining the historical trajectory features. On the other hand, the historical trajectory features are concatenated with the deviation noise Z, and after passing through the ResMLP network, the predicted future inter-frame deviation mean features are output. The combination of historical trajectory features and future inter-frame deviation mean features constitutes the feature representation of the future trajectory. Finally, the model decodes these features into the coordinates of the prediction sequence through the decoding module, thereby generating the predicted trajectory. Using a fully connected MLP network and leveraging historical trajectory information to predict future trajectory information improves processing speed and ensures that trajectory features at all times are represented. At the same time, the use of a residual structure makes the network more robust, improving prediction accuracy.

[0030] The inter-frame deviation noise Z is specifically the noise extracted from historical preprocessed trajectory data using the Conditional Variational Autoencoder (CVAE) method.

[0031] Multi-trajectory prediction based on the CVAE method refers to generating different predicted trajectories by sampling the feature bias noise Z. Z represents the bias noise that predicts the mean of future inter-frame bias from historical trajectory data. This feature is random and follows a normal distribution. The feature of Z can be trained by the network using historical trajectory features and the true mean of inter-frame bias. During the learning process, constraints are formed by the mean vector μ and variance vector σ of Z, and this constraint condition is constructed using KLDloss relative entropy loss.

[0032] The beneficial effects of this invention are:

[0033] 1. The method of the present invention can preprocess noisy and missing data under uncertain and incomplete perception conditions, so that trajectory prediction has high reliability.

[0034] 2. The prediction model used in this invention does not treat the problem as a time-by-time generation task of sequential relationships. Instead, it extracts features representing future trajectories from historical trajectories and introduces feature bias noise Z to achieve multi-trajectory prediction.

[0035] 3. The method of the present invention can realize online trajectory prediction with high real-time performance and can meet the real-time prediction tasks in multiple scenarios.

[0036] In summary, the method of this invention can effectively handle uncertain and incomplete data, providing a reliable solution for real-time prediction of multi-target trajectories, and is applicable to fields such as traffic management and autonomous driving. Attached Figure Description

[0037] Figure 1 This is a flowchart of the workflow of an online trajectory prediction system;

[0038] Figure 2 This is a diagram of the DoubleMLP model architecture;

[0039] Figure 3 This is a graph showing the results of processing uncertain and incomplete data over eight time periods to predict the target. Figure 3 (a) is the result of processing the historical trajectory data of the first segment of the predicted target. Figure 3 (b) is the result of processing the historical trajectory data of the second segment of the predicted target. Figure 3 (c) is the result of processing the historical trajectory data of the predicted target's third segment. Figure 3 (d) is the result of processing the historical trajectory data of the fourth segment of the predicted target. Figure 3 (e) is the result of processing the historical trajectory data of the fifth segment of the predicted target. Figure 3 (f) is the result of processing the historical trajectory data of the sixth segment of the predicted target. Figure 3 (g) is the result of processing the historical trajectory data of the seventh segment of the predicted target. Figure 3 (h) is the effect diagram of the historical trajectory data processing of the predicted target's eighth segment. Detailed Implementation

[0040] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] like Figure 1 As shown, the online trajectory prediction method for uncertain and incomplete data of the present invention includes:

[0042] Step 1: Obtain historical geographic location information of several targets, thereby generating historical trajectory data for each target and sending it to the workbench via User Datagram Protocol (UDP) communication protocol.

[0043] In step one, the geographic location information includes the timestamps, latitude and longitude coordinates, and orientation of each geographic location of the target. The trajectory data is specifically generated based on each geographic location within a certain time period.

[0044] Step 2: After receiving the historical trajectory data of each target, the workbench performs data preprocessing to obtain the historical preprocessed trajectory data of each target.

[0045] In step two, data preprocessing includes performing coordinate transformation, reducing or filtering data noise, filling in missing data points, format transformation, and Z-score standardization.

[0046] Coordinate transformation specifically involves converting the latitude and longitude coordinates in the historical geographical location information of the target into coordinates in a two-dimensional rectangular coordinate system, which facilitates subsequent data analysis and processing.

[0047] Kalman filtering is used to reduce or filter out data noise.

[0048] Reducing or filtering data noise refers to using Kalman filtering to reduce or remove noise in the presence of noise. By taking into account the relationship between observed data and system state, Kalman filtering can effectively filter out noise from sensors and improve data quality.

[0049] Kalman filtering is employed to remove noise and uncertainty from trajectory data. First, state variables are defined based on historical trajectory data, along with a state transition matrix F, an observation matrix H, an observation noise covariance matrix R, and a process noise covariance matrix Q. Then, the state estimates and covariance matrices of the Kalman filter are initialized. During prediction, at each time step, the state transition matrix is ​​used to predict the estimated value and covariance matrix of the next state. After receiving new observation data, the uncertainty of the observation data is considered to calculate updated state estimates and covariance matrices. These prediction and update steps are repeated to gradually improve the accuracy of trajectory estimation. Finally, the noise-filtered trajectory data is obtained from the state estimates.

[0050] Specifically, to fill in missing data points, linear interpolation is used to fill in missing data for targets moving in a straight line; for targets moving in a non-linear manner, polynomial fitting is used to fill in missing data, and the values ​​of missing data points are estimated by fitting a polynomial model.

[0051] For targets with nonlinear motion, the coordinates of missing trajectory points are estimated by fitting the trajectory data into a quadratic curve.

[0052] The formal transformation specifically involves converting the target's historical trajectory data, after filling in the missing data points, into a relative position vector, as follows:

[0053]

[0054]

[0055]

[0056] Where ΔX is the relative position vector of the i-th target; Let be the position difference of the i-th target at time 2 and time 1. Let be the position difference of the i-th target at time t+1 and time t. Let $i$ be the position difference of the i-th target at time $obs+1$ and time $obs$. and These are the historical trajectory data of the i-th target at time t+1 and time t, respectively. and Let x and y be the x-axis coordinates and y-axis coordinates of the i-th target at time t in a two-dimensional rectangular coordinate system, respectively; R is a real number.

[0057] The historical time range of the observations is expressed as t1~t obs The predicted future time range is represented as t. obs+1 ~t obs+k Where, obs is the length of the historical trajectory, t obs Here, k is the current time, and k is the trajectory length to be predicted. This invention uses a relative vector to represent the input data and calculates the position difference between time t+1 and time t for the i-th target. The input for trajectory prediction is represented as Indicates the time interval from t1 to t2. obs The relative position vector of the i-th target within the observation period.

[0058] The input relative vector data is subjected to Z-score standardization, which involves vectorizing and standardizing the input historical trajectory data. Based on the target type, the mean μ and standard deviation σ of the relative position vectors are calculated, and then standardization is performed to normalize the input to a normal distribution.

[0059] Step 3: Construct the target trajectory prediction model DoubleMLP. Input the historical preprocessed trajectory data of each target into the target trajectory prediction model DoubleMLP for training. At the same time, process the data based on the conditional variational autoencoder (CVAE) method to learn the conditional distribution of the historical data and obtain the trained target trajectory prediction model DoubleMLP.

[0060] like Figure 2As shown, in step three, the target trajectory prediction model DoubleMLP, based on the ResMLP fully connected neural network with residuals, includes a historical trajectory feature extraction module, an inter-frame deviation feature extraction module, a mean feature extraction module, and a future trajectory feature decoding module. The historical trajectory feature extraction module includes four fully connected layers connected in sequence, the inter-frame deviation feature extraction module includes five fully connected layers connected in sequence, the mean feature extraction module includes three fully connected layers connected in sequence, and the future trajectory feature decoding module includes five fully connected layers connected in sequence. The input of the target trajectory prediction model DoubleMLP is first fed into the historical trajectory feature extraction module for processing. The outputs of the first and third fully connected layers in the historical trajectory feature extraction module are added to form a residual connection, which is then fed into the fourth fully connected layer for processing, and the output becomes the output of the historical trajectory feature extraction module. The output of the historical trajectory feature extraction module is concatenated with the inter-frame deviation noise Z and then fed into the inter-frame deviation feature extraction module. In the inter-frame deviation feature extraction module, the outputs of the first and fourth fully connected layers are added to form a residual connection, which is then input into the fifth fully connected layer. The output of this residual connection serves as the output fully connected layer of the inter-frame deviation feature extraction module, i.e., the inter-frame average deviation prediction value fully connected layer. The output of the inter-frame deviation feature extraction module is then input into the mean feature extraction module. The outputs of the first and second fully connected layers in the mean feature extraction module are added to form a residual connection, which is then input into the third fully connected layer. The output of this residual connection serves as the output of the mean feature extraction module. The outputs of the historical trajectory feature extraction module and the mean feature extraction module are concatenated and then input into the future trajectory feature decoding module. The outputs of the first and fourth fully connected layers in the future trajectory feature decoding module are added to form a residual connection, which is then input into the fifth fully connected layer. The output of this residual connection serves as the output of the target trajectory prediction model, DoubleMLP.

[0061] The DoubleMLP target trajectory prediction model treats historical trajectory information as a complete block, extracts features from historical trajectories, and then reconstructs future trajectory features to achieve trajectory prediction. The predicted future trajectory features are obtained by combining historical trajectory features with the mean inter-frame deviation features. On one hand, the standardized historical trajectory relative vector sequence is input and processed by the ResMLP network to extract features, obtaining the historical trajectory features. On the other hand, the historical trajectory features are concatenated with the deviation noise Z, and after passing through the ResMLP network, the predicted future inter-frame deviation mean features are output. The combination of historical trajectory features and future inter-frame deviation mean features constitutes the feature representation of the future trajectory. Finally, the model decodes these features into the coordinates of the prediction sequence through the decoding module, thereby generating the predicted trajectory. Using a fully connected MLP network and leveraging historical trajectory information to predict future trajectory information improves processing speed and ensures that trajectory features at all times are represented. At the same time, the use of a residual structure makes the network more robust, improving prediction accuracy.

[0062] The inter-frame deviation noise Z is specifically the noise extracted from historical preprocessed trajectory data using the conditional variational autoencoder (CVAE) method.

[0063] Multi-trajectory prediction based on the CVAE method refers to generating different predicted trajectories by sampling the feature bias noise Z. Z represents the bias noise that predicts the mean of future inter-frame bias from historical trajectory data. This feature is random and follows a normal distribution. The feature of Z can be trained by the network using historical trajectory features and the true mean of inter-frame bias. During the learning process, constraints are formed by the mean vector μ and variance vector σ of Z, and this constraint condition is constructed using KLDloss relative entropy loss.

[0064] Step 4: Collect historical geographical location information of the target to be predicted, thereby generating historical trajectory data of the target to be predicted. After performing the same data preprocessing as in Step 2, the historical trajectory data of the target to be predicted is input into the trained target trajectory prediction model DoubleMLP. At the same time, it is processed based on the Conditional Variational Autoencoder (CVAE) method to generate trajectory data that meets the conditions. The target trajectory prediction model DoubleMLP outputs several future trajectories of the target to be predicted.

[0065] Step 5: After performing coordinate transformation on each future trajectory of the target to be predicted output in Step 4, obtain the predicted geographical location information of the target to be predicted, send the data to the workbench and receive it, and complete the online trajectory prediction.

[0066] like Figure 2The diagram shows the architecture of the DoubleMLP model. The features of the future trajectory are obtained by combining historical trajectory features and the mean inter-frame deviation features. The extraction of historical trajectory features first involves unfolding the 2D coordinates along historical time into a 2×obs vector, which is then projected onto a 16-dimensional feature space through a fully connected layer. Then, a feature extraction network (16D-512D-256D-16D ResMLP) extracts the 16-dimensional historical trajectory features. Obtaining the mean inter-frame deviation features is more complex. First, the 16-dimensional historical trajectory features are concatenated with the 8-dimensional deviation noise Z, then processed through a 24D-1024D-512D-1024D-24D ResMLP network to extract features. Finally, a fully connected layer outputs the predicted 2D mean inter-frame deviation value. This predicted mean inter-frame deviation is then processed by the feature extraction network (16D-8D-16D ResMLP network) to obtain the 16-dimensional predicted mean inter-frame deviation features. The 16-dimensional historical trajectory features are concatenated with the 16-dimensional mean inter-frame deviation features to form a 32-dimensional future trajectory feature. This future feature is then decoded by a ResMLP with a structure of 32-1024-512-256-32 dimensions, outputting a k×2-dimensional vector as the coordinates of the corresponding predicted sequence.

[0067] like Figure 3 of (a), Figure 3 (b) Figure 3 (c) Figure 3 (d) Figure 3 of (e) Figure 3 of (f), Figure 3 of (g), Figure 3 As shown in (h), this diagram illustrates the effect of preprocessing uncertain and incomplete data. This invention uses 8 frames of historical trajectory data to predict the trajectory data of the target for the next 3 frames. Figure 3 The document includes eight smaller graphs, each displaying historical trajectory data of the same target at different time points, preprocessed historical trajectory data, and 20 possible future trajectory data. As can be seen from the graphs, regardless of whether the movement is linear or curvilinear, the method employed in this invention can reduce noise interference and fill in missing points, thus achieving the goal of processing uncertain and incomplete data.

Claims

1. An online trajectory prediction method for uncertain and incomplete data, characterized in that, include: Step 1: Obtain historical geographic location information for several targets, thereby generating historical trajectory data for each target and sending it to the workbench; Step 2: After receiving the historical trajectory data of each target, the workbench performs data preprocessing to obtain the historical preprocessed trajectory data of each target. Step 3: Construct the target trajectory prediction model DoubleMLP. Input the historical preprocessed trajectory data of each target into the target trajectory prediction model DoubleMLP for training. At the same time, process the data based on the conditional variational autoencoder (CVAE) method to obtain the trained target trajectory prediction model DoubleMLP. Step 4: Collect historical geographic location information of the target to be predicted, thereby generating historical trajectory data of the target to be predicted. After performing the same data preprocessing as in Step 2, input the historical trajectory data of the target to be predicted into the trained target trajectory prediction model DoubleMLP. At the same time, it is processed based on the Conditional Variational Autoencoder (CVAE) method. The target trajectory prediction model DoubleMLP outputs several future trajectories of the target to be predicted. Step 5: After performing coordinate transformation on each future trajectory of the target to be predicted output in Step 4, obtain the predicted geographical location information of the target to be predicted, and complete the online trajectory prediction. In step three, the target trajectory prediction model DoubleMLP includes a historical trajectory feature extraction module, an inter-frame deviation feature extraction module, a mean feature extraction module, and a future trajectory feature decoding module. The historical trajectory feature extraction module includes four fully connected layers connected in sequence, the inter-frame deviation feature extraction module includes five fully connected layers connected in sequence, the mean feature extraction module includes three fully connected layers connected in sequence, and the future trajectory feature decoding module includes five fully connected layers connected in sequence. The input of the target trajectory prediction model DoubleMLP is first processed in the historical trajectory feature extraction module. The outputs of the first and third fully connected layers in the historical trajectory feature extraction module are added to form a residual connection, which is then processed by the fourth fully connected layer and its output is used as the output of the historical trajectory feature extraction module. The output of the historical trajectory feature extraction module is concatenated with the inter-frame deviation noise Z and then input into the inter-frame deviation feature extraction module. The inter-frame deviation feature extraction module's outputs are processed as follows: the outputs from the first and fourth fully connected layers are summed to form a residual connection, which is then input into the fifth fully connected layer for further processing, and the output becomes the output of the inter-frame deviation feature extraction module. The output of the inter-frame deviation feature extraction module is then input into the mean feature extraction module for further processing. The outputs from the first and second fully connected layers in the mean feature extraction module are summed to form a residual connection, which is then input into the third fully connected layer for further processing, and the output becomes the output of the mean feature extraction module. The outputs from the historical trajectory feature extraction module and the mean feature extraction module are concatenated and then input into the future trajectory feature decoding module for further processing. The outputs from the first and fourth fully connected layers in the future trajectory feature decoding module are summed to form a residual connection, which is then input into the fifth fully connected layer for further processing, and the output becomes the output of the target trajectory prediction model, DoubleMLP.

2. The online trajectory prediction method for uncertain and incomplete data according to claim 1, characterized in that: In step one, the geographic location information includes the timestamps, latitude and longitude coordinates, and orientation of each geographic location of the target. The trajectory data is specifically generated based on each geographic location within a certain time period.

3. The online trajectory prediction method for uncertain and incomplete data according to claim 1, characterized in that: In step two, data preprocessing includes performing coordinate transformation, reducing or filtering data noise, filling in missing data points, format transformation, and Z-score standardization in sequence.

4. The online trajectory prediction method for uncertain and incomplete data according to claim 3, characterized in that: The coordinate transformation specifically involves converting the latitude and longitude coordinates in the historical geographical location information of the target's historical trajectory data into coordinates in a two-dimensional rectangular coordinate system.

5. The online trajectory prediction method for uncertain and incomplete data according to claim 3, characterized in that: The reduction or filtering of data noise specifically employs the Kalman filtering method.

6. The online trajectory prediction method for uncertain and incomplete data according to claim 3, characterized in that: Specifically, the missing data points are filled by using linear interpolation to fill in the missing data for targets moving in a straight line, and by using polynomial fitting to fill in the missing data for targets moving in a non-linear line.

7. The online trajectory prediction method for uncertain and incomplete data according to claim 3, characterized in that: The formal transformation specifically involves converting the target's historical trajectory data, after filling in the missing data points, into a relative position vector, as detailed below: in, For the first The relative position vectors of the targets; For the first The position difference of the target at time 2 and time 1 For the first The goal is at a certain moment and time Position difference, For the first The goal is at any time and time Positional difference; and They are time points and time The next Historical trajectory data of each target and They are time points The next The target in a two-dimensional rectangular coordinate system Axis coordinates and Axis coordinates; It is a real number.

8. The online trajectory prediction method for uncertain and incomplete data according to claim 1, characterized in that: The inter-frame deviation noise Z is specifically the noise extracted from historical preprocessed trajectory data using the Conditional Variational Autoencoder (CVAE) method.

Images