A trajectory prediction model training sample correction method and device and a storage medium

Abstract

The application discloses a trajectory prediction model training sample correction method and device and a storage medium, and comprises the following steps: obtaining a training sample, obtaining tracking trajectory data, analyzing all trajectory points, calculating the curvature of the trajectory points and the curvature change amount of adjacent trajectory points, and screening out the training sample with abnormal trajectory points for interpolation repair. The trajectory prediction model training sample correction method can obtain smoother tracking trajectory, improve the quality of the training sample, prevent the model from fitting incorrect true values during training, reduce model prediction problems caused by the quality problems of the training sample itself, and thus improve the accuracy and stability of trajectory prediction.

Description

Technical Field

[0001] This invention belongs to the field of autonomous vehicle trajectory prediction, specifically involving a method for repairing training samples for trajectory prediction models. Background Technology

[0002] Trajectory prediction technology based on deep learning models is currently the mainstream trajectory prediction technique. There are two main types of training sample representation methods for trajectory prediction models: one uses images or point clouds as training samples, and the other vectorizes the trajectories of road traffic participants and map elements, such as Waymo's behavior prediction algorithm VectorNet, which stores training samples using a graph data structure. Compared to the former, vectorized data is more interpretable, facilitates feature construction, and has lower computational cost for the model. Therefore, this paper mainly studies deep learning models using vectorized training samples.

[0003] For trajectory prediction models, the training data mainly consists of three parts: the vehicle to be predicted (Agent), obstacles around the vehicle (Actor), and map elements (lane graph). For the prediction task, the historical trajectory of the vehicle and its surrounding obstacles over a period of x seconds, along with the map elements around the vehicle, are typically selected as input. The trajectory of the vehicle over a period of y seconds is selected as the ground truth (GT) to train the neural network. During trajectory prediction model training, it was found that the predicted trajectory ground truth can exhibit anomalies such as excessive curvature or abrupt changes. Such trajectories can lead to misjudgments by downstream modules, causing intermittent braking or takeover, thus affecting the performance of the autonomous vehicle. Summary of the Invention

[0004] To address the problem that excessive curvature or abrupt changes in the ground truth of training samples cause trajectory prediction models to fit incorrect ground truth values ​​during training, this invention proposes a method to improve the stability of trajectory prediction models by correcting the training samples.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for correcting training samples of a trajectory prediction model, characterized by comprising: Obtain training samples from the trajectory prediction training sample set; Obtain tracking trajectory data from training samples; Parse the tracking trajectory data to obtain all trajectory points corresponding to the true value of the trajectory; Traverse all trajectory points and calculate the curvature of each trajectory point and the change in curvature between adjacent trajectory points; Training samples whose absolute curvature of trajectory points is greater than the curvature threshold or whose curvature of trajectory points exhibits abrupt changes are selected as training samples to be repaired. Delete abnormal trajectory points in the tracking trajectory corresponding to the training sample to be repaired; Interpolation is used to repair the deleted abnormal trajectory points.

[0006] To achieve the above objectives, embodiments of the present invention also provide a device for correcting training samples of a trajectory prediction model, characterized by comprising: The training sample extraction module is used to obtain training samples from the trajectory prediction training sample set; The tracking trajectory data extraction module is used to obtain tracking trajectory data from training samples; The trajectory parsing module is used to parse all trajectory points corresponding to the true value of the trajectory from the tracking trajectory data; The first calculation module is used to traverse all trajectory points and calculate the curvature of the trajectory points; The second calculation module is used to traverse all trajectory points and calculate the curvature change of adjacent trajectory points. The filtering module is used to filter out training samples whose absolute value of the curvature of the trajectory points is greater than the curvature threshold or whose curvature of the trajectory points exhibits abrupt changes. The deletion module is used to delete abnormal trajectory points in the tracking trajectory corresponding to the training sample to be repaired; The repair module performs interpolation repair on deleted abnormal trajectory points.

[0007] To achieve the above objectives, embodiments of the present invention also provide an apparatus, including a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that when the program is executed by the processor, it implements the steps of the method for correcting training samples of the trajectory prediction model.

[0008] To achieve the above objectives, embodiments of the present invention also provide a storage medium, characterized in that the storage medium stores at least one program, which can be executed by at least one processor to implement the steps of the trajectory prediction model training sample correction method.

[0009] Beneficial effects of this invention: The trajectory prediction model training sample correction method of the present invention uses interpolation to correct the position, velocity, shape, acceleration and other features of abnormal trajectory points for training samples with abnormal ground truth curvature. This corrects the erroneous ground truth, improves the online performance of the model, reduces model prediction problems caused by the quality of the training samples themselves, and improves the accuracy and stability of trajectory prediction. Attached Figure Description

[0010] Figure 1 The flowchart shows the correction method for training samples of the trajectory prediction model in Example 1. Figure 2 This is a flowchart of the correction device for the training samples of the trajectory prediction model in Example 2. Detailed Implementation

[0011] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0012] Example 1 This embodiment provides a method for correcting training samples of a trajectory prediction model, such as... Figure 1 As shown, it includes: Step 101: Obtain training samples from the training sample set of the deep learning-based trajectory prediction model. The training samples mainly consist of the following data: the historical x-second trajectory of the vehicle to be predicted and the obstacles around the vehicle, the map elements around the vehicle, and the future y-second trajectory of the vehicle. The historical x-second trajectory of the vehicle to be predicted and the obstacles around the vehicle, as well as the map elements around the vehicle, serve as the input to the trajectory prediction model, and the future y-second trajectory of the vehicle serves as the ground truth output of the trajectory prediction model.

[0013] The `get_agent_ids()` interface can be used to obtain the tracking trajectory IDs of all training samples in the database. Each tracking trajectory ID corresponds one-to-one with a training sample, thus allowing you to retrieve the corresponding training sample.

[0014] Step 102: Obtain tracking trajectory data from the training samples, i.e., the trajectory data of the vehicle to be predicted in the next y seconds.

[0015] By inputting the tracking track ID into the get_track_by_id() interface, the tracking track data corresponding to the training sample can be returned from the database.

[0016] Step 103: From the tracking trajectory data of the training samples, use existing analytical methods to obtain all trajectory points corresponding to the true trajectory values.

[0017] Step 104: Traverse all trajectory points, calculate the curvature of each trajectory point, and calculate the change in curvature between every two adjacent trajectory points. Specifically, the curvature calculation for each trajectory point involves grouping three adjacent trajectory points into a circle for calculation. The absolute value of the difference between the curvatures of two adjacent trajectory points is used to obtain the change in curvature between those two adjacent trajectory points.

[0018] Step 105: Compare the curvature of each calculated trajectory point with a preset curvature threshold, and compare the calculated curvature change between every two adjacent trajectory points with a preset curvature change threshold. Mark abnormal trajectory points, i.e., trajectory points with an absolute curvature value greater than the curvature threshold or trajectory points with abrupt curvature changes (i.e., the latter of two adjacent trajectory points with a curvature change greater than the curvature change threshold). Select the training samples corresponding to the abnormal trajectory points as training samples to be repaired.

[0019] Step 106: Delete the abnormal trajectory points in the tracking trajectory corresponding to the selected training samples to be repaired.

[0020] Step 107: Perform interpolation repair on the deleted abnormal trajectory points. Specific interpolation repair methods include: When the deleted abnormal trajectory point is a single trajectory point, and both its preceding and following trajectory points are correct trajectory points, the true values ​​of the preceding and following trajectory points of the abnormal trajectory point are averaged and interpolated to repair it, thereby correcting the position, velocity, shape, acceleration and other characteristics of the abnormal trajectory point.

[0021] When the deleted abnormal trajectory points are multiple consecutive trajectory points, and the trajectory points before and after them are all correct trajectory points, the number of trajectory points to be repaired is calculated, the step size is determined, the true values ​​of the trajectory points before and after the multiple consecutive abnormal trajectory points are averaged, and each trajectory point is interpolated and repaired according to the step size, thereby correcting the position, velocity, shape, acceleration and other characteristics of the abnormal trajectory points.

[0022] When the deleted abnormal trajectory point is a single trajectory point and the true values ​​of its preceding and following trajectory points are missing, interpolation is performed using extrapolation to repair the abnormal trajectory point, thereby correcting its position, velocity, shape, acceleration, and other characteristics.

[0023] When the deleted abnormal trajectory points are multiple consecutive trajectory points, and the correct ground truth values ​​of the preceding and following trajectory points are missing, the corresponding training samples are deleted.

[0024] Step 108: Obtain the next training sample. Repeat steps 102-107 above until all training samples have been traversed.

[0025] Curvature threshold and curvature change threshold can be reasonably set according to the desired trajectory tracking objective. By removing abnormal trajectory points with curvature greater than the curvature threshold or with abrupt curvature changes from the training samples, and performing interpolation repair based on the classification of the removed abnormal trajectory points, a smoother tracking trajectory is obtained. This improves the quality of the training samples, preventing the model from fitting incorrect ground truth values ​​during training. It also reduces model prediction problems caused by the quality of the training samples themselves, thereby improving the accuracy and stability of trajectory prediction.

[0026] Example 2 This embodiment provides a device for correcting training samples of a trajectory prediction model, such as... Figure 2 As shown, the system includes a training sample extraction module 1, a tracking trajectory data extraction module 2, a trajectory parsing module 3, a first calculation module 4, a second calculation module 5, a filtering module 6, a deletion module 7, and a repair module 8. The training sample extraction module 1 obtains training samples from the trajectory prediction training sample set and inputs them into the tracking trajectory data extraction module 2. The tracking trajectory data extraction module 2 then obtains tracking trajectory data from the training samples and inputs it into the trajectory parsing module 3. The trajectory parsing module 3 uses existing technology to parse all trajectory points corresponding to the true trajectory values ​​from the tracking trajectory data. The first calculation module 4 iterates through all trajectory points and calculates the curvature of each trajectory point. The second calculation module 5 iterates through all trajectory points and calculates the change in curvature between every two adjacent trajectory points. The filtering module 6 filters out training samples whose absolute curvature value is greater than a curvature threshold or whose curvature exhibits abrupt changes as training samples to be repaired. Specifically, the filtering module 6 compares the curvature of each trajectory point with a preset curvature threshold, and compares the curvature change between any two adjacent trajectory points with a preset curvature change threshold. Abnormal trajectory points are marked as those with an absolute curvature value greater than the curvature threshold or those with abrupt curvature changes (i.e., the latter of two adjacent trajectory points whose curvature change exceeds the curvature change threshold). The training samples corresponding to these abnormal trajectory points are then selected as training samples to be repaired. The deletion module 7 deletes the abnormal trajectory points from the tracking trajectory corresponding to the training samples to be repaired, so that the repair module 8 can perform interpolation repair on the deleted abnormal trajectory points.

[0027] Specifically, when the deleted abnormal trajectory point is a single trajectory point, and both its preceding and following trajectory points are correct trajectory points, the repair module averages the ground truth values ​​of the preceding and following trajectory points for interpolation repair. When the deleted abnormal trajectory point is a series of consecutive trajectory points, and both its preceding and following trajectory points are correct trajectory points, the repair module calculates the number of trajectory points to be repaired, determines the step size, averages the ground truth values ​​of the preceding and following trajectory points for the series of consecutive abnormal trajectory points, and performs interpolation repair on each trajectory point according to the step size. When the deleted abnormal trajectory point is a single trajectory point, and the correct ground truth values ​​of its preceding and following trajectory points are missing, the repair module performs interpolation repair using extrapolation. When the deleted abnormal trajectory point is a series of consecutive trajectory points, and the correct ground truth values ​​of its preceding and following trajectory points are missing, the repair module deletes the corresponding training samples.

[0028] Example 3 This embodiment provides an apparatus including a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the steps of the trajectory prediction model training sample correction method in the above embodiment.

[0029] Example 4 This embodiment provides a storage medium storing at least one program, which can be executed by at least one processor to implement the steps of the trajectory prediction model training sample correction method in the above embodiment.

[0030] It will be understood by those skilled in the art that the above descriptions are merely preferred embodiments of this application and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for correcting training samples of a trajectory prediction model, characterized in that... include: Training samples are obtained from the trajectory prediction training sample set. The training samples include: the historical x-second trajectory of the vehicle to be predicted and the obstacles around the vehicle, the map elements around the vehicle to be predicted, and the future y-second trajectory of the vehicle to be predicted. The historical x-second trajectory of the vehicle to be predicted and the obstacles around the vehicle, and the map elements around the vehicle to be predicted are used as inputs to the trajectory prediction model, and the future y-second trajectory of the vehicle to be predicted is used as the true value output of the trajectory prediction model. Obtain tracking trajectory data from the training samples; All trajectory points corresponding to the true value of the trajectory are obtained by parsing the tracking trajectory data; Traverse all the trajectory points and calculate the curvature of the trajectory points and the change in curvature between adjacent trajectory points; Training samples with absolute curvature values ​​greater than curvature thresholds or abrupt changes in curvature are selected as training samples to be repaired. Among them, the trajectory point with abrupt curvature changes is the second trajectory point among two adjacent trajectory points whose curvature change amount is greater than the curvature change amount threshold. Delete the abnormal trajectory points in the tracking trajectory corresponding to the training sample to be repaired; Interpolation repair is performed on the deleted abnormal trajectory points, including: When both the trajectory points before and after the abnormal trajectory point are correct trajectory points, the abnormal trajectory point is repaired by interpolation. When the correct ground truth values ​​of the trajectory points before and after the abnormal trajectory point are missing, and the abnormal trajectory point is a series of consecutive trajectory points, the corresponding training samples to be repaired are deleted. When the deleted abnormal trajectory points are multiple consecutive trajectory points, and the trajectory points before and after them are all correct trajectory points, calculate the number of trajectory points to be repaired, determine the step size, take the average of the true values ​​of the before and after trajectory points, and perform interpolation repair on each trajectory point according to the step size. When the deleted abnormal trajectory point is a single trajectory point and the correct ground truth values ​​of its preceding and following trajectory points are missing, interpolation is performed to repair it.

2. The method for correcting training samples of the trajectory prediction model as described in claim 1, characterized in that: When the deleted abnormal trajectory point is a single trajectory point, and both its preceding and following trajectory points are correct trajectory points, the true values ​​of the preceding and following trajectory points are averaged for interpolation repair.

3. A device for correcting training samples of a trajectory prediction model, characterized in that... include: The training sample extraction module is used to obtain training samples from the trajectory prediction training sample set. The training samples include: the historical x-second trajectory of the vehicle to be predicted and the obstacles around the vehicle, the map elements around the vehicle, and the future y-second trajectory of the vehicle. The historical x-second trajectory of the vehicle to be predicted and the obstacles around the vehicle, and the map elements around the vehicle are used as inputs to the trajectory prediction model, and the future y-second trajectory of the vehicle is used as the ground truth output of the trajectory prediction model. The tracking trajectory data extraction module is used to obtain tracking trajectory data from the training samples; The trajectory parsing module is used to parse all trajectory points corresponding to the true value of the trajectory from the tracking trajectory data; The first calculation module is used to traverse all the trajectory points and calculate the curvature of the trajectory points; The second calculation module is used to traverse all the trajectory points and calculate the curvature change of adjacent trajectory points; The filtering module is used to filter out training samples whose absolute curvature value of trajectory points is greater than the curvature threshold or whose curvature of trajectory points has abrupt changes as training samples to be repaired. Among them, the trajectory point with abrupt curvature changes is the second trajectory point among two adjacent trajectory points whose curvature change is greater than the curvature change threshold. The deletion module is used to delete abnormal trajectory points in the tracking trajectory corresponding to the training sample to be repaired. The repair module performs interpolation repair on the deleted abnormal trajectory points, including: When both the trajectory points before and after the abnormal trajectory point are correct trajectory points, the abnormal trajectory point is repaired by interpolation. When the correct ground truth values ​​of the trajectory points before and after the abnormal trajectory point are missing, and the abnormal trajectory point is a series of consecutive trajectory points, the corresponding training samples to be repaired are deleted. When the deleted abnormal trajectory points are multiple consecutive trajectory points, and the trajectory points before and after them are all correct trajectory points, calculate the number of trajectory points to be repaired, determine the step size, take the average of the true values ​​of the before and after trajectory points, and perform interpolation repair on each trajectory point according to the step size. When the deleted abnormal trajectory point is a single trajectory point and the correct ground truth values ​​of its preceding and following trajectory points are missing, interpolation is performed to repair it.

4. The correction device for training samples of the trajectory prediction model as described in claim 3, characterized in that... When the deleted abnormal trajectory point is a single trajectory point, and both its preceding and following trajectory points are correct trajectory points, the repair module performs interpolation repair by averaging the true values ​​of the preceding and following trajectory points.

5. A device for correcting training samples of a trajectory prediction model, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that... When the program is executed by the processor, it implements the steps of the method for correcting training samples of the trajectory prediction model as described in claim 1 or 2.

6. A storage medium, characterized in that... The storage medium stores at least one program, which can be executed by at least one processor to implement the steps of the method for correcting training samples of the trajectory prediction model as described in claim 1 or 2.

Images