WO2023010827A1 - 基于迁移场景的轨迹预测模型的训练方法及装置 - Google Patents

基于迁移场景的轨迹预测模型的训练方法及装置 Download PDF

Info

Publication number
WO2023010827A1
WO2023010827A1 PCT/CN2022/076624 CN2022076624W WO2023010827A1 WO 2023010827 A1 WO2023010827 A1 WO 2023010827A1 CN 2022076624 W CN2022076624 W CN 2022076624W WO 2023010827 A1 WO2023010827 A1 WO 2023010827A1
Authority
WO
WIPO (PCT)
Prior art keywords
trajectory
training samples
prediction model
candidate
trajectory prediction
Prior art date
Application number
PCT/CN2022/076624
Other languages
English (en)
French (fr)
Inventor
樊明宇
黄佳雯
任冬淳
夏华夏
徐一
Original Assignee
北京三快在线科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京三快在线科技有限公司 filed Critical 北京三快在线科技有限公司
Priority to EP22851554.0A priority Critical patent/EP4383037A1/en
Publication of WO2023010827A1 publication Critical patent/WO2023010827A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05D—SYSTEMS FOR CONTROLLING OR REGULATING NON-ELECTRIC VARIABLES
    • G05D1/00—Control of position, course, altitude or attitude of land, water, air or space vehicles, e.g. using automatic pilots
    • G05D1/02—Control of position or course in two dimensions
    • G05D1/021—Control of position or course in two dimensions specially adapted to land vehicles
    • G05D1/0212—Control of position or course in two dimensions specially adapted to land vehicles with means for defining a desired trajectory
    • G05D1/0221—Control of position or course in two dimensions specially adapted to land vehicles with means for defining a desired trajectory involving a learning process
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01C—MEASURING DISTANCES, LEVELS OR BEARINGS; SURVEYING; NAVIGATION; GYROSCOPIC INSTRUMENTS; PHOTOGRAMMETRY OR VIDEOGRAMMETRY
    • G01C21/00—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00
    • G01C21/26—Navigation; Navigational instruments not provided for in groups G01C1/00 - G01C19/00 specially adapted for navigation in a road network
    • G01C21/34—Route searching; Route guidance
    • G01C21/3407—Route searching; Route guidance specially adapted for specific applications
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements

Definitions

  • the present application relates to the field of unmanned driving, and in particular to a training method and device for a trajectory prediction model based on migration scenarios.
  • unmanned driving has initially entered people's lives, and in practical applications, ensuring the safe driving of unmanned equipment is a prerequisite for performing various tasks through unmanned equipment.
  • a large amount of trajectory data of obstacles can be obtained in history, and a model that can predict the trajectory of obstacles can be pre-trained through these trajectory data.
  • This application provides a training method and device for a trajectory prediction model based on migration scenarios, and the technical solution is as follows:
  • a training method for a trajectory prediction model based on migration scenarios including:
  • the first trajectory prediction model is a trajectory prediction model obtained through training of training samples in a preset geographic scene
  • the plurality of candidate training samples are training samples in a migration scenario A sample, the geographical area where the migration scene is located is different from the preset geographical scene, and the candidate training samples include historical driving trajectories of obstacles around the target device;
  • the reference value corresponding to the candidate training samples is used to characterize the degree of difference between the driving characteristics of the historical driving trajectory included in the candidate training samples and the driving characteristics corresponding to the preset geographical scene;
  • the first trajectory prediction model is trained according to the target training samples to obtain a second trajectory prediction model, and the second trajectory prediction model is used to predict the driving trajectory of each obstacle in the migration scene.
  • a training device for a trajectory prediction model based on a migration scenario including:
  • the obtaining module is used to obtain a first trajectory prediction model and a plurality of candidate training samples, the first trajectory prediction model is a trajectory prediction model obtained through training samples in a preset geographic scene, and the plurality of candidate training samples are Migrating the training samples under the scenario, where the geographic area where the migration scenario is located is different from the preset geographic scenario, and the candidate training samples include historical driving trajectories of obstacles around the target device;
  • a determining module configured to, for any of the candidate training samples, determine the candidate training samples according to the trajectory features corresponding to the candidate training samples, and/or the prediction result of the first trajectory prediction model on the candidate training samples.
  • a reference value corresponding to the sample, the reference value corresponding to the candidate training sample is used to characterize the degree of difference between the driving characteristics of the historical driving trajectory included in the candidate training samples and the driving characteristics corresponding to the preset geographical scene;
  • a selection module configured to select a target training sample from the plurality of candidate training samples according to the reference values corresponding to the plurality of candidate training samples;
  • a training module configured to train the first trajectory prediction model according to the target training samples to obtain a second trajectory prediction model, and the second trajectory prediction model is used to predict the movement of each obstacle in the migration scene track.
  • a method for controlling an unmanned device including:
  • the driving trajectory is input into the pre-trained second trajectory prediction model, and the obstacle trajectory is output through the second trajectory model, and the second trajectory prediction model is obtained by training through the above-mentioned training method of the trajectory prediction model based on the migration scene ;
  • the unmanned driving device is controlled.
  • a control device for unmanned equipment including:
  • An acquisition module configured to acquire the driving track of the obstacles around the unmanned equipment, the unmanned equipment driving in the migration scene
  • the input module is used to input the driving trajectory into the pre-trained second trajectory prediction model, and output the obstacle trajectory through the second trajectory model, and the second trajectory prediction model passes the above-mentioned trajectory prediction model based on the migration scene.
  • the training method is trained to obtain;
  • a control module configured to control the unmanned device according to the predicted trajectory of the obstacle.
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:
  • the first trajectory prediction model is a trajectory prediction model obtained through training of training samples in a preset geographic scene
  • the plurality of candidate training samples are training samples in a migration scenario A sample, the geographical area where the migration scene is located is different from the preset geographical scene, and the candidate training samples include historical driving trajectories of obstacles around the target device;
  • the reference value corresponding to the candidate training samples is used to characterize the degree of difference between the driving characteristics of the historical driving trajectory included in the candidate training samples and the driving characteristics corresponding to the preset geographical scene;
  • the first trajectory prediction model is trained according to the target training samples to obtain a second trajectory prediction model, and the second trajectory prediction model is used to predict the driving trajectory of each obstacle in the migration scene.
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:
  • the driving trajectory is input into the pre-trained second trajectory prediction model, and the obstacle trajectory is output through the second trajectory model, and the second trajectory prediction model is obtained by training through the above-mentioned training method of the trajectory prediction model based on the migration scene ;
  • the unmanned driving device is controlled.
  • An unmanned driving device comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, and the processor implements the following steps when executing the program:
  • the first trajectory prediction model is a trajectory prediction model obtained through training of training samples in a preset geographic scene
  • the plurality of candidate training samples are training samples in a migration scenario A sample, the geographical area where the migration scene is located is different from the preset geographical scene, and the candidate training samples include historical driving trajectories of obstacles around the target device;
  • the reference value corresponding to the candidate training samples is used to characterize the degree of difference between the driving characteristics of the historical driving trajectory included in the candidate training samples and the driving characteristics corresponding to the preset geographical scene;
  • the first trajectory prediction model is trained according to the target training samples to obtain a second trajectory prediction model, and the second trajectory prediction model is used to predict the driving trajectory of each obstacle in the migration scene.
  • An unmanned driving device comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, and the processor implements the following steps when executing the program:
  • the driving trajectory is input into the pre-trained second trajectory prediction model, and the obstacle trajectory is output through the second trajectory model, and the second trajectory prediction model is obtained by training through the above-mentioned training method of the trajectory prediction model based on the migration scene ;
  • the unmanned driving device is controlled.
  • FIG. 1 is a schematic flowchart of a training method for a trajectory prediction model based on a migration scenario provided in an embodiment of the present application;
  • FIG. 2 is a schematic diagram of iterative training of a trajectory prediction model provided by an embodiment of the present application
  • FIG. 3 is a schematic flowchart of a control method for an unmanned device provided in an embodiment of the present application
  • FIG. 4 is a schematic diagram of a training device for a trajectory prediction model based on a migration scene provided by an embodiment of the present application
  • FIG. 5 is a schematic diagram of a control device for an unmanned driving device provided in an embodiment of the present application.
  • FIG. 6 is a schematic diagram of an unmanned driving device corresponding to FIG. 1 or FIG. 3 provided by an embodiment of the present application.
  • Fig. 1 is a schematic flow chart of a training method for a trajectory prediction model based on a migration scenario in the present application, including the following steps:
  • the business platform acquires a first trajectory prediction model and multiple candidate training samples.
  • the first trajectory prediction model is a trajectory prediction model obtained by training training samples in a preset geographical scene.
  • the multiple candidate training samples are The geographical area where the migration scene is located is different from the preset geographical scene, and the candidate training samples include historical driving trajectories of obstacles around the target device.
  • the business platform needs to operate unmanned driving equipment in a geographical scene, it often needs to train the trajectory prediction model through the driving trajectories of various targets (such as vehicles, pedestrians, etc.) in this geographical scene,
  • the trajectory prediction model can predict the obstacle trajectory of each target object based on the driving trajectory of each target object in this geographical scene, and the unmanned driving device can avoid each target object in the geographic scene based on the obstacle trajectory .
  • the previously trained trajectory prediction model usually cannot be directly adapted to the trajectory prediction in another geographic scenario. In this way, it is necessary to Train a trajectory prediction model for trajectory prediction in another geographic scene.
  • the target is also the obstacle.
  • the service platform can obtain the first trajectory prediction model trained in advance by training samples in a preset geographical scene.
  • the first trajectory prediction model is a trajectory prediction model trained by the business platform based on training samples in a preset geographic scenario, or a trajectory obtained by other business platforms trained in training samples in a preset geographic scenario
  • the prediction model is not limited in this embodiment of the present application.
  • the business platform obtains multiple training samples in the migration scenario, and uses the multiple training samples as multiple candidate training samples, wherein the geographic area of the migration scenario is different from the preset geographic scenario, and each candidate training sample includes obstacles around the target device historical trajectories.
  • the migration scene is different from the geographical area of the preset geographical scene.
  • the migration scene and the preset geographical scene are different cities.
  • the migration scene refers to city A
  • the preset geographical scene is Refers to city B.
  • the preset geographical scene refers to the city
  • the migration scene refers to the countryside.
  • the driving habits of obstacles (such as vehicles, pedestrians, etc.) in the migration scene may be different from the preset geographical scene.
  • a human-driven device refers to, for example, an unmanned vehicle, or an ordinary vehicle, etc.
  • the target device is also referred to as a specified device.
  • the multiple training samples in the migration scenario acquired by the service platform are prepared for further training of the above-mentioned first trajectory prediction model. Since the first trajectory prediction model is obtained by training the training samples in the preset geographic scene, the trajectory prediction effect of the first trajectory prediction model in the migration scenario may be poor, so it is necessary to use the migration scenario The training samples further train the trajectory prediction model.
  • the execution subject of the training trajectory prediction model may refer to the server in the business platform, the server contained in the unmanned driving device itself, etc.
  • the business platform is used as the execution subject below, and the model training Each step is explained.
  • the service platform determines the reference value corresponding to the candidate training sample according to the trajectory feature corresponding to the candidate training sample and/or the prediction result of the first trajectory prediction model for the candidate training sample, the The reference value corresponding to the candidate training sample is used to characterize the degree of difference between the driving characteristics of the historical driving trajectory included in the candidate training sample and the driving characteristics corresponding to the preset geographic scene.
  • the service platform after the service platform acquires multiple candidate training samples, it can perform training for each candidate training sample, for example, the service platform can perform training according to the trajectory features and/or the trajectory corresponding to the historical driving trajectory included in the candidate training samples.
  • the prediction model determines the reference value corresponding to the candidate training sample based on the prediction result of the candidate training sample, and the reference value corresponding to the candidate training sample is used to represent the driving characteristics of the historical driving trajectory included in the candidate training sample and the preset geographical scene The degree of differentiation between the corresponding driving characteristics.
  • the business platform can determine the reference value only through the trajectory features corresponding to multiple candidate training samples, or only through the prediction results of the trajectory prediction model for each candidate training sample, or a combination of the two, and then can be selected based on the reference value Get the target training samples.
  • the service platform determines the reference value.
  • the service platform determines that the candidate training sample corresponds to The first reference value is used to characterize the degree of uncertainty of the trajectory prediction model based on the candidate training sample for trajectory prediction.
  • the first reference value is also the reference value corresponding to the candidate training sample.
  • the service platform determines the degree of difference between the prediction result of the first trajectory prediction model for the candidate training sample and the labeling result corresponding to the candidate training sample.
  • the service platform determines the first reference value corresponding to the candidate training sample according to the degree of difference.
  • the service platform may also determine the confidence level of the prediction result of the candidate training sample by the first trajectory prediction model, and use the confidence level as the first reference value.
  • the service platform determines the second reference value corresponding to the candidate training sample according to the trajectory characteristics corresponding to the candidate training sample, and the second reference value indicates that the candidate training sample is used for training The diversity of the training samples of the trajectory prediction model.
  • the purpose of determining the second reference value is to hope that the training samples selected for training the trajectory prediction model can have diversity, so that the selected training samples
  • the included driving trajectories can be of various kinds, so that it can also ensure that training samples different from the driving characteristics of the preset geographic scene are selected from the candidate training samples.
  • the second reference value is also the reference value corresponding to the candidate training sample.
  • the service platform may determine, for each candidate training sample, that the trajectory feature corresponding to the historical driving trajectory included in the candidate training sample corresponds to the training sample under the preset geographic scene.
  • the degree of difference between trajectory features is used as the second reference value corresponding to the candidate training sample.
  • the trajectory features mentioned here may refer to feature vectors corresponding to historical driving trajectories included in the training samples.
  • the service platform After the service platform determines the above-mentioned first reference value and/or second reference value in the above-mentioned way, it can determine the candidate training data through the first reference value of each candidate training sample and/or the second reference value of each candidate training sample.
  • the reference value corresponding to the sample that is, the reference value can be obtained by integrating the above-mentioned first reference value and the second reference value, or the reference value can be determined only by the first reference value or the second reference value, and if the first reference value is integrated and the second reference value to calculate the above reference value, different weights may be set for the first reference value and the second reference value, so as to determine the reference value.
  • the service platform selects a target training sample from the multiple candidate training samples according to the reference values corresponding to the multiple candidate training samples.
  • the service platform can select training samples that are valuable to the first trajectory prediction model as target training samples, and use the target training samples to train the trajectory prediction model, thereby improving The generalization ability of the first trajectory prediction model to expand the scope of application of the first trajectory prediction model. Therefore, candidate training samples whose driving characteristics are different from those corresponding to the preset geographical scene can be selected as target training samples.
  • These target training samples can represent the unique driving characteristics of the migration scene, so that the first trajectory prediction model
  • the target training samples in the migration scenario can be learned to improve the accuracy of the trajectory prediction model for the trajectory prediction of the migration scenario. In other words, the target training samples can reflect the difference between the migration scene and the preset geographic scene.
  • the target training sample is determined based on the first reference value, that is, The first trajectory prediction model has candidate training samples with certain uncertainty when performing trajectory prediction. Using these target training samples to train the trajectory prediction model can improve the accuracy of the trajectory prediction model, and these target training samples are likely to be different from the driving characteristics corresponding to the preset geographical scenes, so the training of the trajectory prediction model is valuable.
  • the target training samples are selected only by the first reference value, there may be abnormal training samples in the selected target training samples, for example, there may be certain noise in the historical driving trajectory included in the selected target training samples, That is, the wrong trajectory may be collected due to the instability of the signal when collecting the driving trajectory.
  • the driving trajectory in the selected target training samples may be special, such as emergency stop, speeding, etc.
  • the service platform can also combine other types of reference values to select target training samples. For example, the service platform can determine the target training sample through the second reference value of the candidate training sample, or determine the target training sample through the reference value obtained from the first reference value and the second reference value. This embodiment of the present application does not do this limited.
  • the service platform trains the first trajectory prediction model according to the target training sample to obtain a second trajectory prediction model, and the second trajectory prediction model is used to predict the driving trajectory of each obstacle in the migration scene.
  • the service platform may select a target training sample from each candidate training sample, and train the first trajectory prediction model according to the target training sample. That is to say, the service platform can select candidate training samples with higher reference values as the selected target training samples, and perform model training based on the target training samples.
  • the service platform may preset the ratio of selected target training samples to candidate training samples, and select according to this ratio. Alternatively, the service platform may sort the candidate training samples according to the descending order of the reference values, and select the candidate training samples that are ranked before the set ranking as the target training samples.
  • the business platform selects target training samples, it can select all the required target training samples at one time according to the reference value, or it can select some target training samples each time during iterative training, and pass these target training samples. After a round of training is performed on the first trajectory prediction model, the target training samples are selected, and the next round of training is performed through the selected target training samples until the first trajectory prediction model is trained until convergence.
  • the service platform determines a plurality of remaining training samples among the plurality of candidate training samples,
  • the multiple remaining training samples are candidate training samples other than those selected as target training samples in the previous N-1 rounds among the multiple candidate training samples, and N is a positive integer;
  • the service platform re-determines that the multiple remaining training samples correspond to The reference value;
  • the business platform selects the target training samples required for the Nth round of training from the multiple remaining training samples according to the reference values corresponding to the multiple remaining training samples re-determined;
  • the required target training samples are used to perform N rounds of training on the first trajectory prediction model until the first trajectory prediction model meets the preset training target to obtain the second trajectory prediction model.
  • FIG. 2 is a schematic diagram of iterative training for a trajectory prediction model provided by the present application.
  • the business platform can select a certain number of target training samples from the candidate training samples each time, and use these target training samples to carry out a stage of training on the first trajectory prediction model, and then, in the next When selecting the target training samples, re-determine the reference values of the remaining training samples (that is, the candidate training samples that have not been selected as the target training samples among the candidate training samples), and then select the remaining training samples by using the reference values of the remaining training samples.
  • the service platform inputs the target training sample into the first trajectory prediction model, and the first trajectory prediction model performs prediction based on the target training sample, and outputs the predicted obstacle trajectory.
  • the service platform trains the first trajectory prediction model based on the difference virtuality between the marked obstacle trajectory corresponding to the target training sample and the predicted obstacle trajectory.
  • the process of training the first trajectory prediction model is also the process of updating the model parameters of the first trajectory prediction model.
  • both the first reference value and the second reference value are used when determining the initial reference value of each candidate training sample, each time the reference value corresponding to the remaining training samples is re-determined, Both the first reference value and the second reference value can be re-determined. Of course, if only one reference value is used, only this reference value can be re-determined.
  • the first trajectory prediction model re-predicts the trajectory of the remaining training samples to obtain a prediction result.
  • the service platform re-determines the first reference value according to the prediction result.
  • the method of determining the second reference value can be adjusted as follows, that is, for each remaining training sample, the business platform determines the trajectory characteristics corresponding to the remaining training sample and the selected target The degree of difference between the trajectory features corresponding to the training samples. The service platform determines the degree of difference as the re-determined second reference value corresponding to the remaining training samples, and obtains the reference value corresponding to the remaining training samples according to the re-determined second reference value.
  • the business platform can use the degree of difference between the trajectory features corresponding to the historical driving trajectories included in the candidate training samples and the trajectory features corresponding to the training samples in the preset geographical scene as the initialization value of the second reference value .
  • the second reference value is set to a fixed value. The reason why the second reference value is determined in this way is to make the selected target training samples themselves have diversity. .
  • the trajectory prediction model needs to be applied to the trajectory of the unmanned driving device driving in the migration scenario.
  • the trained second trajectory prediction model can be configured on the unmanned driving device for unmanned driving in the migration scenario. Trajectory prediction for driving devices. Therefore, the unmanned driving device can obtain the driving trajectory of the surrounding obstacles, and input the driving trajectory of the surrounding obstacles into the second trajectory prediction model after the training is completed, and obtain the predicted obstacle trajectory, and according to the predicted The obstacle track is used to control the unmanned equipment.
  • the unmanned device drives at least in the migration scene, that is, the unmanned device can not only drive in the migration scene, but also in other geographic scenes, and the unmanned device When driving in the migration scene, the trajectory prediction can be performed through the above-mentioned trained second trajectory prediction model.
  • FIG. 3 is a schematic flowchart of a control method for an unmanned driving device provided in the present application.
  • the service platform acquires the driving trajectory of obstacles around the unmanned driving device, and the unmanned driving device is driving in a migration scene.
  • the business platform inputs the driving trajectory into the pre-trained second trajectory prediction model, outputs the obstacle trajectory through the second trajectory prediction model, and the second trajectory prediction model is trained by the training method of the trajectory prediction model based on the migration scene get.
  • the second trajectory model is obtained through training in the above steps 101-104.
  • the service platform controls the unmanned device according to the trajectory of the obstacle.
  • the unmanned equipment mentioned above can refer to unmanned vehicles, unmanned aerial vehicles, automatic distribution equipment and other equipment that can realize automatic driving. Based on this, the training method of the trajectory prediction model based on the migration scene and the control method of the unmanned device provided by this application can be used for the trajectory prediction of the unmanned device in the migration scenario.
  • the unmanned device can be specifically applied to The field of distribution through unmanned equipment, such as the business scenario of using unmanned equipment for express delivery, logistics, and takeaway.
  • the service platform can further train the first trajectory prediction model that has been trained with the training samples in the preset geographical scene by using the training samples in the migration scenario. .
  • suitable training samples can be selected from the training samples in the migration scenario as target training samples, and the first trajectory prediction model is trained using the target training samples.
  • the strategy selected by the service platform may be based on the first reference value, that is, the inaccuracy of the trajectory prediction model for the candidate training samples, so that the trajectory prediction model can accurately predict the migration scenario, and through the second
  • the second reference value can select various target training samples with diversity, thereby reducing the proportion of abnormal samples included in the training samples used to train the second trajectory prediction model, and further improving the trajectory prediction of the second trajectory prediction model. accuracy.
  • the embodiment of the present application also provides a corresponding model training device. And the control device of unmanned equipment, as shown in Figure 4 and 5.
  • FIG. 4 is a schematic diagram of a training device for a trajectory prediction model based on a migration scenario provided in an embodiment of the present application, including:
  • the obtaining module 401 is used to obtain a first trajectory prediction model and a plurality of candidate training samples, the first trajectory prediction model is a trajectory prediction model obtained through training samples in a preset geographical scene, and the plurality of candidate training samples are migration A training sample in a scene where the geographical area where the migration scene is located is different from the preset geographical scene, and the candidate training sample includes historical driving trajectories of obstacles around the target device.
  • the determining module 402 is configured to, for any of the candidate training samples, determine the reference corresponding to the candidate training sample according to the trajectory characteristics corresponding to the candidate training sample, and/or the prediction result of the first trajectory prediction model for the candidate training sample.
  • the reference value corresponding to the candidate training sample is used to characterize the degree of difference between the driving characteristics of the historical driving trajectory included in the candidate training sample and the driving characteristics corresponding to the preset geographic scene.
  • the selection module 403 is configured to select a target training sample from the multiple candidate training samples according to the reference values corresponding to the multiple candidate training samples.
  • the training module 404 is configured to train the first trajectory prediction model according to the target training sample to obtain a second trajectory prediction model, and the second trajectory prediction model is used to predict the driving trajectory of each obstacle in the migration scene.
  • the determining module 402 is configured to determine a first reference value corresponding to the candidate training sample according to the prediction result of the first trajectory prediction model for the candidate training sample, and the first reference value is used to characterize the The first trajectory prediction model performs trajectory prediction uncertainty based on the candidate training samples. According to the trajectory characteristics corresponding to the candidate training samples, determine the second reference value corresponding to the candidate training samples, the second reference value is used to characterize the diversity of the candidate training samples in the training samples used to train the first trajectory prediction model sex. The reference value corresponding to the candidate training sample is determined according to the first reference value corresponding to the candidate training sample and/or the second reference value corresponding to the candidate training sample.
  • the determining module 402 is configured to determine the degree of difference between the prediction result of the first trajectory prediction model on the candidate training sample and the labeling result corresponding to the candidate training sample. According to the degree of difference, the first reference value corresponding to the candidate training sample is determined.
  • the determining module 402 is configured to determine the degree of difference between the trajectory feature corresponding to the candidate training sample and the trajectory feature corresponding to the training sample under the preset geographic scene as the candidate training sample corresponding Second reference value.
  • the training module 404 is configured to, after training the first trajectory prediction model with the target training samples selected in the N-1th round of training, determine how many of the multiple candidate training samples remaining training samples, where the multiple remaining training samples are candidate training samples other than those selected as target training samples in the previous N-1 rounds of the multiple candidate training samples, and N is a positive integer. Re-determine the reference values corresponding to the plurality of remaining training samples. According to the re-determined reference values corresponding to the plurality of remaining training samples, the target training samples required for the Nth round of training are selected from the plurality of remaining training samples. According to the target training samples required by the N-th round of training, N-th rounds of training are performed on the first trajectory prediction model until the first trajectory prediction model meets a preset training target to obtain the second trajectory prediction model.
  • the determining module 402 is configured to, for any of the remaining training samples, determine a second reference value corresponding to the remaining training samples, where the second reference value is the trajectory feature corresponding to the remaining training samples and the selected The degree of difference between trajectory features corresponding to the obtained target training samples, and the second reference value is used to characterize the diversity of the remaining training samples in the training samples used to train the first trajectory prediction model.
  • the reference value corresponding to the remaining training sample is determined according to the second reference value corresponding to the remaining training sample.
  • FIG. 5 is a schematic diagram of a control device for an unmanned device provided in an embodiment of the present application, including:
  • the obtaining module 501 is used to obtain the driving track of the obstacles around the unmanned equipment, and the unmanned equipment is driving in the migration scene;
  • the input module 502 is used to input the driving trajectory into the pre-trained second trajectory prediction model, and output the obstacle trajectory through the second trajectory model, and the second trajectory prediction model passes the above-mentioned trajectory prediction model based on the migration scene
  • the training method is trained to obtain;
  • the control module 503 is configured to control the unmanned device according to the obstacle track.
  • the embodiment of the present application also provides a computer-readable storage medium, the storage medium stores a computer program, and the computer program can be used to execute the training method of a trajectory prediction model based on a migration scene provided in FIG. 1 or FIG. A control method for a human-driven device.
  • the embodiment of the present application also provides a schematic structural diagram of an unmanned driving device 600 shown in FIG. 6 corresponding to FIG. 1 or FIG. 3 .
  • the driverless device includes a processor 601 , an internal bus 602 , a network interface 603 , a memory 604 and a non-volatile memory 605 , and of course it may also include hardware required by other services.
  • the processor 601 reads the corresponding computer program from the non-volatile memory 605 into the memory and then runs it to realize the training method of the trajectory prediction model based on the migration scene described in FIG. 1 or FIG. 3 and the implementation of the unmanned device. Control Method.
  • a typical implementing device is a computer.
  • the computer may be, for example, a vehicle terminal, a personal computer, a laptop computer, a cell phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet A computer, a wearable device, or a combination of any of these devices.
  • embodiments of the present application may be provided as methods, devices or computer program products. Therefore, the embodiment of the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, embodiments of the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.
  • computer-usable storage media including but not limited to disk storage, CD-ROM, optical storage, etc.
  • a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned training method of the trajectory prediction model based on the migration scene and the control method of the unmanned equipment are realized.

Landscapes

  • Engineering & Computer Science (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Traffic Control Systems (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种基于迁移场景的轨迹预测模型的训练方法，包括：(101)获取第一轨迹预测模型以及多个候选训练样本，第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型；(102)针对任一候选训练样本，根据候选训练样本对应的轨迹特征，和/或第一轨迹预测模型对候选训练样本的预测结果，确定候选训练样本对应的参考值；(103)根据多个候选训练样本对应的参考值，从多个候选训练样本中选取目标训练样本；(104)根据目标训练样本，对第一轨迹预测模型进行训练，得到第二轨迹预测模型，第二轨迹预测模型用于预测迁移场景中各障碍物的行驶轨迹。

Description

基于迁移场景的轨迹预测模型的训练方法及装置
本申请要求于2021年8月2日提交的申请号为202110878026.6、发明名称为“基于迁移场景用于预测障碍物轨迹的模型训练方法”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及无人驾驶领域，尤其涉及一种基于迁移场景的轨迹预测模型的训练方法及装置。
背景技术
随着信息技术的不断发展，无人驾驶已初步进入人们的生活，而在实际应用中，保证无人驾驶设备的安全行驶，则是通过无人驾驶设备执行各项任务的前提条件。为了无人驾驶设备的安全行驶，需要使无人驾驶设备对周围的障碍物未来的轨迹进行预测，从而无人驾驶设备可以提前制定行驶策略，达到安全行驶的目的。
在相关技术中，可以通过历史上获取到大量障碍物的轨迹数据，并通过这些轨迹数据预先训练出可以预测障碍物轨迹的模型。
发明内容
本申请提供一种基于迁移场景的轨迹预测模型的训练方法及装置，技术方案如下：
提供了一种基于迁移场景的轨迹预测模型的训练方法，包括：
获取第一轨迹预测模型以及多个候选训练样本，所述第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，所述多个候选训练样本为迁移场景下的训练样本，所述迁移场景所在的地理区域与所述预设地理场景不同，所述候选训练样本包括目标设备周围障碍物的历史行驶轨迹；
针对任一所述候选训练样本，根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值，所述候选训练样本对应的参考值用于表征所述候选训练样本包括的历史行驶轨迹的行驶特点与所述预设地理场景对应的行驶特点之间的区别程度；
根据所述多个候选训练样本对应的参考值，从所述多个候选训练样本中选取目标训练样本；
根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型，所述第二轨迹预测模型用于预测所述迁移场景中各障碍物的行驶轨迹。
提供了一种基于迁移场景的轨迹预测模型的训练装置，包括：
获取模块，用于获取第一轨迹预测模型以及多个候选训练样本，所述第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，所述多个候选训练样本为迁移场景下的训练样本，所述迁移场景所在的地理区域与所述预设地理场景不同，所述候 选训练样本包括目标设备周围障碍物的历史行驶轨迹；
确定模块，用于针对任一所述候选训练样本，根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值，所述候选训练样本对应的参考值用于表征所述候选训练样本包括的历史行驶轨迹的行驶特点与所述预设地理场景对应的行驶特点之间的区别程度；
选取模块，用于根据所述多个候选训练样本对应的参考值，从所述多个候选训练样本中选取目标训练样本；
训练模块，用于根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型，所述第二轨迹预测模型用于预测所述迁移场景中各障碍物的行驶轨迹。
提供了一种无人驾驶设备的控制方法，包括：
获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过上述基于迁移场景的轨迹预测模型的训练方法进行训练得到；
根据所述障碍物轨迹，对所述无人驾驶设备进行控制。
提供了一种无人驾驶设备的控制装置，包括：
获取模块，用于获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
输入模块，用于将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过上述基于迁移场景的轨迹预测模型的训练方法进行训练得到；
控制模块，用于根据所述预测出的障碍物轨迹，对所述无人驾驶设备进行控制。
提供了一种计算机可读存储介质，所述存储介质存储有计算机程序，所述计算机程序被处理器执行时实现下述步骤：
获取第一轨迹预测模型以及多个候选训练样本，所述第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，所述多个候选训练样本为迁移场景下的训练样本，所述迁移场景所在的地理区域与所述预设地理场景不同，所述候选训练样本包括目标设备周围障碍物的历史行驶轨迹；
针对任一所述候选训练样本，根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值，所述候选训练样本对应的参考值用于表征所述候选训练样本包括的历史行驶轨迹的行驶特点与所述预设地理场景对应的行驶特点之间的区别程度；
根据所述多个候选训练样本对应的参考值，从所述多个候选训练样本中选取目标训练样本；
根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型，所述第二轨迹预测模型用于预测所述迁移场景中各障碍物的行驶轨迹。
提供了一种计算机可读存储介质，所述存储介质存储有计算机程序，所述计算机程序被处理器执行时实现下述步骤：
获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过上述基于迁移场景的轨迹预测模型的训练方法进行训练得到；
根据所述障碍物轨迹，对所述无人驾驶设备进行控制。
提供了一种无人驾驶设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述程序时实现下述步骤：
获取第一轨迹预测模型以及多个候选训练样本，所述第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，所述多个候选训练样本为迁移场景下的训练样本，所述迁移场景所在的地理区域与所述预设地理场景不同，所述候选训练样本包括目标设备周围障碍物的历史行驶轨迹；
针对任一所述候选训练样本，根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值，所述候选训练样本对应的参考值用于表征所述候选训练样本包括的历史行驶轨迹的行驶特点与所述预设地理场景对应的行驶特点之间的区别程度；
根据所述多个候选训练样本对应的参考值，从所述多个候选训练样本中选取目标训练样本；
根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型，所述第二轨迹预测模型用于预测所述迁移场景中各障碍物的行驶轨迹。
提供了一种无人驾驶设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述程序时实现下述步骤：
获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过上述基于迁移场景的轨迹预测模型的训练方法进行训练得到；
根据所述障碍物轨迹，对所述无人驾驶设备进行控制。
附图说明
此处所说明的附图用来提供对本申请的进一步理解，构成本申请的一部分。
图1为本申请实施例提供的一种基于迁移场景的轨迹预测模型的训练方法的流程示意图；
图2为本申请实施例提供的一种对轨迹预测模型进行迭代训练的示意图；
图3为本申请实施例提供的一种无人驾驶设备的控制方法的流程示意图；
图4为本申请实施例提供的一种基于迁移场景的轨迹预测模型的训练装置的示意图；
图5为本申请实施例提供的无人驾驶设备的控制装置的示意图；
图6为本申请实施例提供的一种对应于图1或图3的无人驾驶设备的示意图。
具体实施方式
为使本申请的目的、技术方案和优点更加清楚，下面将结合本申请的具体实施例及相应的附图对本申请提供的技术方案进行清楚、完整地描述。显然，所描述的实施例仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
相关技术中的方式，往往对无人驾驶设备在不同地理场景行驶时存在不同的效果，例如，可以通过大量的A地的轨迹数据训练出模型，而该模型若应用到无人驾驶设备在B地行驶时的轨迹预测，预测出的轨迹可能会不准确，因此，现有技术这种方式存在一定的不准确性，而若直接通过B地的全部轨迹数据重新训练轨迹预测模型，则所需的数据量较大效率较低。
所以，如何提高在不同地理场景下轨迹预测的准确性以及提高模型训练的效率，则是一个亟待解决的问题。
以下结合附图，详细说明本申请各实施例提供的技术方案。
图1为本申请中一种基于迁移场景的轨迹预测模型的训练方法的流程示意图，包括以下步骤：
101、业务平台获取第一轨迹预测模型以及多个候选训练样本，该第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，该多个候选训练样本为迁移场景下的训练样本，该迁移场景所在的地理区域与该预设地理场景不同，该候选训练样本包括目标设备周围障碍物的历史行驶轨迹。
在一些实施例中，业务平台若需要在一种地理场景内运营无人驾驶设备，往往需要通过这种地理场景内的各目标物(如车辆、行人等)的行驶轨迹来训练轨迹预测模型，通过该轨迹预测模型能够基于这种地理场景内的各目标物的行驶轨迹来预测各目标物的障碍物轨迹，该无人驾驶设备能够基于该障碍物轨迹来规避该地理场景内的各目标物。在业务平台之后想要在另一种地理场景内运营无人驾驶设备的情况下，之前训练的轨迹预测模型通常无法直接适应于在另一种地理场景内的轨迹预测，这样一来，就需要训练出在另一种地理场景内进行轨迹预测的轨迹预测模型。在上述说明中，目标物也即是障碍物。
基于此，业务平台能够获取预先通过预设地理场景下的训练样本进行训练的第一轨迹预测模型。在一些实施例中，该第一轨迹预测模型是该业务平台基于预设地理场景下的训练样本训练得到的轨迹预测模型，或者是其他业务平台于预设地理场景下的训练样本训练得到的轨迹预测模型，本申请实施例对此不做限定。业务平台获取迁移场景下的多个训练样本，将该多个训练样本作为多个候选训练样本，其中，迁移场景的地理区域与预设地理场景不同，每个候选训练样本包括目标设备周围障碍物的历史行驶轨迹。
在一些实施例中，迁移场景与预设地理场景的地理区域不同的情况可以有多种，例如，迁移场景和预设地理场景为不同城市，比如迁移场景是指A市，预设地理场景是指B市。再例如，预设地理场景是指城市，迁移场景是指乡村，迁移场景中障碍物(如车辆、行人等)的行驶习惯可以与预设地理场景不同，上述提到的目标设备可以是指无人驾驶设备， 比如是指无人驾驶车辆，也可以是指普通车辆等，在一些实施例中，该目标设备也被称为指定设备。
在一些实施例中，业务平台获取到的迁移场景下的多个训练样本，是为上述第一轨迹预测模型进行进一步地训练准备的。由于该第一轨迹预测模型是通过预设地理场景下的训练样本进行训练得到，因此，该第一轨迹预测模型在迁移场景中进行轨迹预测的效果可能较差，所以，需要通过迁移场景下的训练样本对该轨迹预测模型进行进一步地训练。
在本申请实施例中，训练轨迹预测模型的执行主体可以是指业务平台中的服务器、无人驾驶设备自身包含的服务器等，为了便于描述，下面均以业务平台为执行主体，对模型训练中的每个步骤进行说明。
102、针对任一候选训练样本，业务平台根据该候选训练样本对应的轨迹特征，和/或该第一轨迹预测模型对该候选训练样本的预测结果，确定该候选训练样本对应的参考值，该候选训练样本对应的参考值用于表征该候选训练样本包括的历史行驶轨迹的行驶特点与该预设地理场景对应的行驶特点之间的区别程度。
在一些实施例中，业务平台获取到多个候选训练样本后，可以针对每个候选训练样本进行训练，比如，业务平台根据该候选训练样本包括的历史行驶轨迹对应的轨迹特征和/或该轨迹预测模型对该候选训练样本的预测结果，确定出该候选训练样本对应的参考值，该候选训练样本对应的参考值用于表征该候选训练样本包括的历史行驶轨迹的行驶特点与预设地理场景对应的行驶特点之间的区别程度。
其中，业务平台可以仅通过多个候选训练样本对应的轨迹特征，也可以仅通过该轨迹预测模型对各候选训练样本的预测结果，或结合两者来确定参考值，后续能够基于参考值来挑选出目标训练样本。业务平台确定参考值的方式可以有多种，在一些实施例中，针对任一候选训练样本，业务平台根据该第一轨迹预测模型对该候选训练样本的预测结果，确定出该候选训练样本对应的第一参考值，该第一参考值用于表征该轨迹预测模型基于该候选训练样本进行轨迹预测的不确定程度。在仅通过该轨迹预测模型对各候选训练样本的预测结果确定该候选训练样本对应的参考值的情况下，该第一参考值也即是该候选训练样本对应的参考值。
在一些实施例中，对于一个候选训练样本，业务平台确定该第一轨迹预测模型对该候选训练样本的预测结果与该候选训练样本对应的标注结果之间的差异程度。业务平台根据该差异程度，确定出该候选训练样本对应的第一参考值。或者，业务平台也可以确定出该第一轨迹预测模型对该候选训练样本的预测结果的置信度，并将该置信度作为该第一参考值。
在一些实施例中，针对任一候选训练样本，业务平台根据该候选训练样本对应的轨迹特征，确定该候选训练样本对应的第二参考值，第二参考值表征该候选训练样本在用于训练轨迹预测模型的训练样本中的多样性，在一些实施例中，确定第二参考值的目的是希望选取出的用于训练该轨迹预测模型的训练样本能够具有多样性，使得选取出的训练样本包括的行驶轨迹可以是多种的，这样一来，也可以保证从候选训练样本中选取出与预设地理场景的行驶特点不同的训练样本。在仅通过候选训练样本对应的轨迹特征确定该候选训练样本对应的参考值的情况下，该第二参考值也即是该候选训练样本对应的参考值。
在一些实施例中，在确定第二参考值时，业务平台可以针对每个候选训练样本，确定 出该候选训练样本包括的历史行驶轨迹对应的轨迹特征与预设地理场景下的训练样本对应的轨迹特征之间的差异程度，并将该差异程度作为该候选训练样本对应的第二参考值。其中，这里提到的轨迹特征可以是指训练样本包括的历史行驶轨迹对应的特征向量。
业务平台通过上述方式确定出上述第一参考值和/或第二参考值后，可以通过各候选训练样本的第一参考值和/或各候选训练样本的第二参考值，确定出各候选训练样本对应的参考值，即，可以综合上述第一参考值以及第二参考值得到参考值，也可以仅通过第一参考值或第二参考值来确定参考值，并且，若综合第一参考值以及第二参考值来计算上述参考值，可以针对第一参考值以及第二参考值设置不同的权重，从而确定出参考值。
103、业务平台根据多个候选训练样本对应的参考值，从所述多个候选训练样本中选取目标训练样本。
通过步骤103，在迁移场景下的全部训练样本中，业务平台可以挑选出对第一轨迹预测模型有价值的训练样本作为目标训练样本，并通过目标训练样本对该轨迹预测模型进行训练，从而提高第一轨迹预测模型的泛化能力，以扩大第一轨迹预测模型的适用范围。因此，可以选取出行驶特点与预设地理场景对应的行驶特点具有一定区别的候选训练样本作为目标训练样本，这些目标训练样本能够表示出迁移场景独特的行驶特点，从而使得该第一轨迹预测模型能够针对迁移场景所具备的目标训练样本进行学习，以提升该轨迹预测模型对迁移场景轨迹预测的准确性。或者说，目标训练样本能够反映迁移场景和预设地理场景之间的差异。
在一些实施例中，在业务平台基于该第一轨迹预测模型对该候选训练样本的预测结果的置信度，确定第一参考值的情况下，基于该第一参考值确定目标训练样本也即是该第一轨迹预测模型在进行轨迹预测时存在一定不确定性的候选训练样本。通过这些目标训练样本来训练轨迹预测模型则可以提升该轨迹预测模型的准确性，并且这些目标训练样本很可能与预设地理场景对应的行驶特点存在区别，对该轨迹预测模型的训练具有价值。
但是，若仅通过第一参考值来选取目标训练样本，则可能选出的目标训练样本中存在异常的训练样本，例如，可能选取出的目标训练样本包括的历史行驶轨迹中存在一定的噪声，即，可能由于采集行驶轨迹时的信号不稳定等问题导致采集到了错误的轨迹，再例如，可能选取出的目标训练样本中的行驶轨迹较为特殊，如，急停、超速等，所以，除了上述第一参考值的方式，业务平台还可以结合其他类型的参考值来选取目标训练样本。比如，业务平台能够通过该候选训练样本的第二参考值来确定目标训练样本，或者通过第一参考值和第二参考值得到的参考值来确定目标训练样本，本申请实施例对此不做限定。
104、业务平台根据该目标训练样本，对该第一轨迹预测模型进行训练，得到第二轨迹预测模型，该第二轨迹预测模型用于预测迁移场景中各障碍物的行驶轨迹。
在一些实施例中，业务平台确定出各候选训练样本对应的参考值后，可以从各候选训练样本中选取目标训练样本，并根据目标训练样本对该第一轨迹预测模型进行训练。也就是说，业务平台可以选取参考值较高的候选训练样本作为选取出的目标训练样本，并基于目标训练样本进行模型训练。在一些实施例中，业务平台可以预先设置选取出的目标训练样本占候选训练样本的比例，并按照该比例进行选取。或者，业务平台可以根据参考值从大到小的顺序对各候选训练样本进行排序，并将排在设定排位前的候选训练样本选取为目标训练样本。
需要说明的是，业务平台在选取目标训练样本时，可以按照参考值一次性选取出所需的全部目标训练样本，也可以在迭代训练中，每次选取出一些目标训练样本，通过这些目标训练样本对第一轨迹预测模型进行一轮训练后，再进行目标训练样本的选取，并通过选取出的目标训练样本进行下一轮训练，直到将第一轨迹预测模型训练到收敛为止。
在一些实施例中，在通过第N-1轮训练中选取出的目标训练样本，对该第一轨迹预测模型进行训练后，业务平台确定该多个候选训练样本中的多个剩余训练样本，该多个剩余训练样本为该多个候选训练样本中除在前N-1轮被选为目标训练样本之外的候选训练样本，N为正整数；业务平台重新确定该多个剩余训练样本对应的参考值；业务平台根据重新确定出的该多个剩余训练样本对应的参考值，从该多个剩余训练样本中选取第N轮训练所需的目标训练样本；业务平台根据该第N轮训练所需的目标训练样本，对该第一轨迹预测模型进行第N轮训练，直到该第一轨迹预测模型符合预设的训练目标，得到该第二轨迹预测模型。
上述训练过程参见图2，图2为本申请提供的一种对轨迹预测模型进行迭代训练的示意图。
从图2中可以看出，业务平台每次可以从候选训练样本中选取出一定数量的目标训练样本，并通过这些目标训练样本对该第一轨迹预测模型进行一个阶段的训练，而后，在下次选取目标训练样本时，重新确定剩余训练样本(即，候选训练样本中还未被选为目标训练样本的候选训练样本)的参考值，再通过剩余训练样本的参考值，来选取出这一次剩余训练样本中的目标训练样本，以此类推，直到将该第一轨迹预测模型训练到能够在迁移场景下进行准确的轨迹预测。
在一些实施例中，在每次迭代训练过程中，业务平台将该目标训练样本输入该第一轨迹预测模型，通过该第一轨迹预测模型基于该目标训练样本进行预测，输出预测障碍物轨迹。业务平台基于该目标训练样本对应的标注障碍物轨迹与该预测障碍物轨迹之间的差异虚拟性，对该第一轨迹预测模型进行训练。其中，对该第一轨迹预测模型进行训练的过程，也即是对该第一轨迹预测模型的模型参数进行更新的过程。
在一些实施例中，若在确定各候选训练样本初始的参考值时，既采用了第一参考值，也采用了第二参考值，则在每次重新确定剩余训练样本对应的参考值时，可以将第一参考值以及第二参考值均重新确定出来，当然，若只采用了一种参考值，则可以只重新确定这种参考值。
其中，在每次重新确定第一参考值时，由于该第一轨迹预测模型在之前进行了一轮的迭代训练，那么该第一轨迹预测模型的准确性有了一定的提升，因此，可以通过此时的第一轨迹预测模型对该剩余训练样本重新进行轨迹预测，得到预测结果。业务平台根据该预测结果，重新确定第一参考值。
而在重新确定第二参考值时，则可以将确定第二参考值的方式做以下调整，即，业务平台针对每个剩余训练样本，确定该剩余训练样本对应的轨迹特征与已选取出的目标训练样本对应的轨迹特征之间的差异程度。业务平台将该差异程度确定为该剩余训练样本对应的重新确定的第二参考值，并根据该重新确定的第二参考值，得到该剩余训练样本对应的参考值。
也就是说，业务平台可以将上述候选训练样本包括的历史行驶轨迹对应的轨迹特征， 与预设地理场景下的训练样本对应的轨迹特征之间的差异程度，作为第二参考值的初始化的值。从第二轮选取目标训练样本开始，使用剩余训练样本对应的轨迹特征与已选取出的目标训练样本对应的轨迹特征之间的差异程度，来确定出目标训练样本的第二参考值，当然，初始化该第二参考值的方式还可以有多种，如，将该第二参考值设为固定值，之所以这样来确定第二参考值，是希望使得选取出的目标训练样本本身具有多样性。
上述均是以对该轨迹预测模型进行训练的角度，对本申请提供的基于迁移场景的轨迹预测模型的训练方法进行说明，而该轨迹预测模型需要应用于无人驾驶设备行驶在迁移场景中的轨迹预测中，也就是说，通过上述方式将对第一轨迹预测模型进行了训练后，则可以将训练完毕的第二轨迹预测模型配置在无人驾驶设备上，用于在迁移场景下的无人驾驶设备的轨迹预测。因此，无人驾驶设备可以获取周围障碍物的行驶轨迹，并将周围障碍物的行驶轨迹输入到训练完成后的第二轨迹预测模型中，得到的预测出的障碍物轨迹，并根据预测出的障碍物轨迹，对该无人驾驶设备进行控制。在一些实施例中，该无人驾驶设备至少在迁移场景中行驶，也就是说，该无人驾驶设备不仅可以在迁移场景中行驶，也可以在其他地理场景中行驶，而该无人驾驶设备在迁移场景中行驶时，可以通过上述训练后的第二轨迹预测模型来进行轨迹预测。
因此，下面以无人驾驶设备的角度进行说明，如图3所示。
图3为本申请提供的一种无人驾驶设备的控制方法的流程示意图。
301、业务平台获取无人驾驶设备周围障碍物的行驶轨迹，该无人驾驶设备在迁移场景中行驶。
302、业务平台将该行驶轨迹输入到预先训练的第二轨迹预测模型，通过该第二轨迹预测模型输出障碍物轨迹，该第二轨迹预测模型通过基于迁移场景的轨迹预测模型的训练方法进行训练得到。
也即是，该第二轨迹模型是通过上述步骤101-104训练得到的。
303、业务平台根据该障碍物轨迹，对该无人驾驶设备进行控制。
上述提到的无人驾驶设备可以是指无人车、无人机、自动配送设备等能够实现自动驾驶的设备。基于此，通过本申请提供的基于迁移场景的轨迹预测模型的训练方法以及无人驾驶设备的控制方法可以用于无人驾驶设备在迁移场景中进行轨迹预测，该无人驾驶设备具体可应用于通过无人驾驶设备进行配送的领域，如，使用无人驾驶设备进行快递、物流、外卖等配送的业务场景。
从上述方法中可以看出，业务平台可以对预先通过预设地理场景下的训练样本进行过训练的第一轨迹预测模型，采用迁移场景下的训练样本对该第一轨迹预测模型进行进一步地训练。在训练过程中，可以在迁移场景下的训练样本选取出适合的训练样本，作为目标训练样本，采用目标训练样本对该第一轨迹预测模型进行训练。在一些实施例中，业务平台选取的策略可以依据第一参考值，即，轨迹预测模型对候选训练样本进行轨迹预测的不准确性，使得轨迹预测模型能够针对迁移场景进行准确预测，以及通过第二参考值可以选取出具备多样性的各目标训练样本，从而降低了用于训练该第二轨迹预测模型的训练样本包括的异常样本的比例，进而进一步地提高了第二轨迹预测模型进行轨迹预测的准确性。
以上为本申请的一个或多个实施例提供的基于迁移场景的轨迹预测模型的训练方法以及无人驾驶设备的控制方法，基于同样的思路，本申请实施例还提供了相应的模型训练的 装置以及无人驾驶设备的控制装置，如图4、5所示。
图4为本申请实施例提供的一种基于迁移场景的轨迹预测模型的训练装置的示意图，包括：
获取模块401，用于获取第一轨迹预测模型以及多个候选训练样本，该第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，该多个候选训练样本为迁移场景下的训练样本，该迁移场景所在的地理区域与该预设地理场景不同，该候选训练样本包括目标设备周围障碍物的历史行驶轨迹。
确定模块402，用于针对任一该候选训练样本，根据该候选训练样本对应的轨迹特征，和/或该第一轨迹预测模型对该候选训练样本的预测结果，确定该候选训练样本对应的参考值，该候选训练样本对应的参考值用于表征该候选训练样本包括的历史行驶轨迹的行驶特点与该预设地理场景对应的行驶特点之间的区别程度。
选取模块403，用于根据该多个候选训练样本对应的参考值，从该多个候选训练样本中选取目标训练样本。
训练模块404，用于根据该目标训练样本，对该第一轨迹预测模型进行训练，得到第二轨迹预测模型，该第二轨迹预测模型用于预测该迁移场景中各障碍物的行驶轨迹。
在一些实施例中，该确定模块402用于，根据该第一轨迹预测模型对该候选训练样本的预测结果，确定该候选训练样本对应的第一参考值，该第一参考值用于表征该第一轨迹预测模型基于该候选训练样本进行轨迹预测的不确定程度。根据该候选训练样本对应的轨迹特征，确定该候选训练样本对应的第二参考值，该第二参考值用于表征该候选训练样本在用于训练该第一轨迹预测模型的训练样本中的多样性。根据该候选训练样本对应的第一参考值和/或该候选训练样本对应的第二参考值，确定该候选训练样本对应的参考值。
在一些实施例中，该确定模块402用于，确定该第一轨迹预测模型对该候选训练样本的预测结果与该候选训练样本对应的标注结果之间的差异程度。根据该差异程度，确定该候选训练样本对应的第一参考值。
在一些实施例中，该确定模块402用于，将该候选训练样本对应的轨迹特征与该预设地理场景下的训练样本对应的轨迹特征之间的差异程度，确定为该候选训练样本对应的第二参考值。
在一些实施例中，该训练模块404用于，在通过第N-1轮训练中选取出的目标训练样本，对该第一轨迹预测模型进行训练后，确定该多个候选训练样本中的多个剩余训练样本，该多个剩余训练样本为该多个候选训练样本中除在前N-1轮被选为目标训练样本之外的候选训练样本，N为正整数。重新确定该多个剩余训练样本对应的参考值。根据重新确定出的该多个剩余训练样本对应的参考值，从该多个剩余训练样本中选取第N轮训练所需的目标训练样本。根据该第N轮训练所需的目标训练样本，对该第一轨迹预测模型进行第N轮训练，直到该第一轨迹预测模型符合预设的训练目标，得到该第二轨迹预测模型。
在一些实施例中，该确定模块402用于，针对任一该剩余训练样本，确定该剩余训练样本对应的第二参考值，该第二参考值为该剩余训练样本对应的轨迹特征与已选取出的目标训练样本对应的轨迹特征之间的差异程度，该第二参考值用于表征该剩余训练样本在用于训练该第一轨迹预测模型的训练样本中的多样性。根据该剩余训练样本对应的第二参考值，确定该剩余训练样本对应的参考值。
图5为本申请实施例提供的一种无人驾驶设备的控制装置的示意图，包括：
获取模块501，用于获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
输入模块502，用于将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过上述基于迁移场景的轨迹预测模型的训练方法进行训练得到；
控制模块503，用于根据该障碍物轨迹，对该无人驾驶设备进行控制。
本申请实施例还提供了一种计算机可读存储介质，该存储介质存储有计算机程序，计算机程序可用于执行上述图1或图3提供的一种基于迁移场景的轨迹预测模型的训练方法和无人驾驶设备的控制方法。
本申请实施例还提供了图6所示的一种对应于图1或图3的无人驾驶设备600的示意结构图。如图6所述，在硬件层面，该无人驾驶设备包括处理器601、内部总线602、网络接口603、内存604以及非易失性存储器605，当然还可能包括其他业务所需要的硬件。处理器601从非易失性存储器605中读取对应的计算机程序到内存中然后运行，以实现上述图1或图3所述的基于迁移场景的轨迹预测模型的训练方法和无人驾驶设备的控制方法。当然，除了软件实现方式之外，本申请并不排除其他实现方式，比如逻辑器件抑或软硬件结合的方式等等，也就是说以下处理流程的执行主体并不限定于各个逻辑单元，也可以是硬件或逻辑器件。
上述实施例阐明的系统、装置、模块或单元，可以由计算机芯片或实体实现，或者由具有某种功能的产品来实现。一种典型的实现设备为计算机。在一些实施例中，计算机例如可以为车载终端、个人计算机、膝上型计算机、蜂窝电话、相机电话、智能电话、个人数字助理、媒体播放器、导航设备、电子邮件设备、游戏控制台、平板计算机、可穿戴设备或者这些设备中的任何设备的组合。
为了描述的方便，描述以上装置时以功能分为各种单元分别描述。当然，在实施本申请时可以把各单元的功能在同一个或多个软件和/或硬件中实现。
本领域内的技术人员应明白，本申请实施例可提供为方法、装置或计算机程序产品。因此，本申请实施例可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请实施例可采用在一个或多个其中有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
提供了一种计算机可读存储介质，所述存储介质存储有计算机程序，所述计算机程序被处理器执行时实现上述基于迁移场景的轨迹预测模型的训练方法和无人驾驶设备的控制方法。
以上所述仅为本申请的实施例而已，并不用于限制本申请。对于本领域技术人员来说，本申请可以有各种更改和变化。凡在本申请的精神和原理之内所作的任何修改、等同替换、改进等，均应包含在本申请的权利要求范围之内。

Claims (13)

  1. 一种基于迁移场景的轨迹预测模型的训练方法，包括：
    获取第一轨迹预测模型以及多个候选训练样本，所述第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，所述多个候选训练样本为迁移场景下的训练样本，所述迁移场景所在的地理区域与所述预设地理场景不同，所述候选训练样本包括目标设备周围障碍物的历史行驶轨迹；
    针对任一所述候选训练样本，根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值，所述候选训练样本对应的参考值用于表征所述候选训练样本包括的历史行驶轨迹的行驶特点与所述预设地理场景对应的行驶特点之间的区别程度；
    根据所述多个候选训练样本对应的参考值，从所述多个候选训练样本中选取目标训练样本；
    根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型，所述第二轨迹预测模型用于预测所述迁移场景中各障碍物的行驶轨迹。
  2. 如权利要求1所述的方法，其中，所述根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值包括：
    根据所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的第一参考值，所述第一参考值用于表征所述第一轨迹预测模型基于所述候选训练样本进行轨迹预测的不确定程度；
    根据所述候选训练样本对应的轨迹特征，确定所述候选训练样本对应的第二参考值，所述第二参考值用于表征所述候选训练样本在用于训练所述第一轨迹预测模型的训练样本中的多样性；
    根据所述候选训练样本对应的第一参考值和/或所述候选训练样本对应的第二参考值，确定所述候选训练样本对应的参考值。
  3. 如权利要求2所述的方法，其中，所述根据所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的第一参考值包括：
    确定所述第一轨迹预测模型对所述候选训练样本的预测结果与所述候选训练样本对应的标注结果之间的差异程度；
    根据所述差异程度，确定所述候选训练样本对应的第一参考值。
  4. 如权利要求2所述的方法，其中，所述根据所述候选训练样本对应的轨迹特征，确定所述候选训练样本对应的第二参考值包括：
    将所述候选训练样本对应的轨迹特征与所述预设地理场景下的训练样本对应的轨迹特征之间的差异程度，确定为所述候选训练样本对应的第二参考值。
  5. 如权利要求1～4任一项所述的方法，其中，所述根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型包括：
    在通过第N-1轮训练中选取出的目标训练样本，对所述第一轨迹预测模型进行训练后，确定所述多个候选训练样本中的多个剩余训练样本，所述多个剩余训练样本为所述多个候选训练样本中除在前N-1轮被选为目标训练样本之外的候选训练样本，N为正整数；
    重新确定所述多个剩余训练样本对应的参考值；
    根据重新确定出的所述多个剩余训练样本对应的参考值，从所述多个剩余训练样本中选取第N轮训练所需的目标训练样本；
    根据所述第N轮训练所需的目标训练样本，对所述第一轨迹预测模型进行第N轮训练，直到所述第一轨迹预测模型符合预设的训练目标，得到所述第二轨迹预测模型。
  6. 如权利要求5所述的方法，其中，重新确定所述多个剩余训练样本对应的参考值包括：
    针对任一所述剩余训练样本，确定所述剩余训练样本对应的第二参考值，所述第二参考值为所述剩余训练样本对应的轨迹特征与已选取出的目标训练样本对应的轨迹特征之间的差异程度，所述第二参考值用于表征所述剩余训练样本在用于训练所述第一轨迹预测模型的训练样本中的多样性；
    根据所述剩余训练样本对应的第二参考值，确定所述剩余训练样本对应的参考值。
  7. 一种无人驾驶设备的控制方法，包括：
    获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
    将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过权利要求1～6任一项所述的方法进行训练得到；
    根据所述障碍物轨迹，对所述无人驾驶设备进行控制。
  8. 一种基于迁移场景的轨迹预测模型的训练装置，包括：
    获取模块，用于获取第一轨迹预测模型以及多个候选训练样本，所述第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，所述多个候选训练样本为迁移场景下的训练样本，所述迁移场景所在的地理区域与所述预设地理场景不同，所述候选训练样本包括目标设备周围障碍物的历史行驶轨迹；
    确定模块，用于针对任一所述候选训练样本，根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值，所述候选训练样本对应的参考值用于表征所述候选训练样本包括的历史行驶轨迹的行驶特点与所述预设地理场景对应的行驶特点之间的区别程度；
    选取模块，用于根据所述多个候选训练样本对应的参考值，从所述多个候选训练样本中选取目标训练样本；
    训练模块，用于根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型，所述第二轨迹预测模型用于预测所述迁移场景中各障碍物的行驶轨迹。
  9. 一种无人驾驶设备的控制装置，包括：
    获取模块，用于获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
    输入模块，用于将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过权利要求1～6任一项所述的方法进行训练得到；
    控制模块，用于根据所述预测出的障碍物轨迹，对所述无人驾驶设备进行控制。
  10. 一种计算机可读存储介质，所述存储介质存储有计算机程序，所述计算机程序被处理器执行时实现下述步骤：
    获取第一轨迹预测模型以及多个候选训练样本，所述第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，所述多个候选训练样本为迁移场景下的训练样本，所述迁移场景所在的地理区域与所述预设地理场景不同，所述候选训练样本包括目标设备周围障碍物的历史行驶轨迹；
    针对任一所述候选训练样本，根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值，所述候选训练样本对应的参考值用于表征所述候选训练样本包括的历史行驶轨迹的行驶特点与所述预设地理场景对应的行驶特点之间的区别程度；
    根据所述多个候选训练样本对应的参考值，从所述多个候选训练样本中，选取目标训练样本；
    根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型，所述第二轨迹预测模型用于预测所述迁移场景中各障碍物的行驶轨迹。
  11. 一种计算机可读存储介质，所述存储介质存储有计算机程序，所述计算机程序被处理器执行时实现下述步骤：
    获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
    将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过权利要求1～6任一项所述的方法进行训练得到；
    根据所述障碍物轨迹，对所述无人驾驶设备进行控制。
  12. 一种无人驾驶设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述程序时实现下述步骤：
    获取第一轨迹预测模型以及多个候选训练样本，所述第一轨迹预测模型是通过预设地理场景下的训练样本训练得到的轨迹预测模型，所述多个候选训练样本为迁移场景下的训练样本，所述迁移场景所在的地理区域与所述预设地理场景不同，所述候选训练样本包括目标设备周围障碍物的历史行驶轨迹；
    针对任一所述候选训练样本，根据所述候选训练样本对应的轨迹特征，和/或所述第一轨迹预测模型对所述候选训练样本的预测结果，确定所述候选训练样本对应的参考值，所述候选训练样本对应的参考值用于表征所述候选训练样本包括的历史行驶轨迹的行驶特点与所述预设地理场景对应的行驶特点之间的区别程度；
    根据所述多个候选训练样本对应的参考值，从所述多个候选训练样本中，选取目标训练样本；
    根据所述目标训练样本，对所述第一轨迹预测模型进行训练，得到第二轨迹预测模型，所述第二轨迹预测模型用于预测所述迁移场景中各障碍物的行驶轨迹。
  13. 一种无人驾驶设备，包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述程序时实现下述步骤：
    获取无人驾驶设备周围障碍物的行驶轨迹，所述无人驾驶设备在迁移场景中行驶；
    将所述行驶轨迹输入到预先训练的第二轨迹预测模型，通过所述第二轨迹模型输出障碍物轨迹，所述第二轨迹预测模型通过权利要求1～6任一项所述的方法进行训练得到；
    根据所述障碍物轨迹，对所述无人驾驶设备进行控制。
PCT/CN2022/076624 2021-08-02 2022-02-17 基于迁移场景的轨迹预测模型的训练方法及装置 WO2023010827A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
EP22851554.0A EP4383037A1 (en) 2021-08-02 2022-02-17 Migration scene-based trajectory prediction model training method and apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110878026.6 2021-08-02
CN202110878026.6A CN113325855B (zh) 2021-08-02 2021-08-02 基于迁移场景用于预测障碍物轨迹的模型训练方法

Publications (1)

Publication Number Publication Date
WO2023010827A1 true WO2023010827A1 (zh) 2023-02-09

Family

ID=77426743

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/076624 WO2023010827A1 (zh) 2021-08-02 2022-02-17 基于迁移场景的轨迹预测模型的训练方法及装置

Country Status (3)

Country Link
EP (1) EP4383037A1 (zh)
CN (1) CN113325855B (zh)
WO (1) WO2023010827A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117874529A (zh) * 2024-03-12 2024-04-12 浪潮电子信息产业股份有限公司 运动轨迹预测方法、模型训练方法、装置、设备及介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113325855B (zh) * 2021-08-02 2021-11-30 北京三快在线科技有限公司 基于迁移场景用于预测障碍物轨迹的模型训练方法
CN114675672A (zh) * 2022-04-02 2022-06-28 南京航空航天大学 一种基于迁移学习的无人机实时轨迹集成预测方法

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109359793A (zh) * 2018-08-03 2019-02-19 阿里巴巴集团控股有限公司 一种针对新场景的预测模型训练方法及装置
US20190129436A1 (en) * 2017-10-28 2019-05-02 TuSimple System and method for real world autonomous vehicle trajectory simulation
CN111626097A (zh) * 2020-04-09 2020-09-04 吉利汽车研究院（宁波）有限公司 一种障碍物未来轨迹的预测方法、装置、电子设备及存储介质
CN111881973A (zh) * 2020-07-24 2020-11-03 北京三快在线科技有限公司 一种样本选择方法、装置、存储介质及电子设备
WO2021062594A1 (en) * 2019-09-30 2021-04-08 Beijing Voyager Technology Co., Ltd. Systems and methods for jointly predicting trajectories of multiple moving objects
CN112990375A (zh) * 2021-04-29 2021-06-18 北京三快在线科技有限公司 一种模型训练方法、装置、存储介质及电子设备
CN113033364A (zh) * 2021-03-15 2021-06-25 商汤集团有限公司 轨迹预测、行驶控制方法、装置、电子设备及存储介质
CN113325855A (zh) * 2021-08-02 2021-08-31 北京三快在线科技有限公司 基于迁移场景用于预测障碍物轨迹的模型训练方法

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101210729B1 (ko) * 2012-03-06 2012-12-11 한국가스안전공사 지능형 위험전조 패턴 분석모델을 이용한 에너지 플랜트의 실시간 위험전조 예측방법 및 예측시스템
CN105573323A (zh) * 2016-01-12 2016-05-11 福州华鹰重工机械有限公司 自动驾驶轨迹生成方法及装置
CN110371112B (zh) * 2019-07-06 2021-10-01 深圳数翔科技有限公司 一种自动驾驶车辆的智能避障系统及方法
CN111332298B (zh) * 2020-02-19 2021-08-31 北京百度网讯科技有限公司 可行驶区域确定方法、装置、设备及存储介质
CN111208838B (zh) * 2020-04-20 2020-11-03 北京三快在线科技有限公司 一种无人驾驶设备的控制方法及装置
CN111639525A (zh) * 2020-04-22 2020-09-08 上海擎感智能科技有限公司 一种感知算法的训练方法、装置及计算机存储介质
CN111912423B (zh) * 2020-10-13 2021-02-02 北京三快在线科技有限公司 一种预测障碍物轨迹以及模型训练的方法及装置
CN112650220B (zh) * 2020-12-04 2022-03-25 东风汽车集团有限公司 一种车辆自动驾驶方法、车载控制器及系统
CN112364997B (zh) * 2020-12-08 2021-06-04 北京三快在线科技有限公司 一种障碍物的轨迹预测方法及装置
CN112712138B (zh) * 2021-01-19 2022-05-20 腾讯科技（深圳）有限公司 一种图像处理方法、装置、设备及存储介质
CN113110526B (zh) * 2021-06-15 2021-09-24 北京三快在线科技有限公司 一种模型训练的方法、无人驾驶设备的控制方法及装置

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190129436A1 (en) * 2017-10-28 2019-05-02 TuSimple System and method for real world autonomous vehicle trajectory simulation
CN109359793A (zh) * 2018-08-03 2019-02-19 阿里巴巴集团控股有限公司 一种针对新场景的预测模型训练方法及装置
WO2021062594A1 (en) * 2019-09-30 2021-04-08 Beijing Voyager Technology Co., Ltd. Systems and methods for jointly predicting trajectories of multiple moving objects
CN111626097A (zh) * 2020-04-09 2020-09-04 吉利汽车研究院（宁波）有限公司 一种障碍物未来轨迹的预测方法、装置、电子设备及存储介质
CN111881973A (zh) * 2020-07-24 2020-11-03 北京三快在线科技有限公司 一种样本选择方法、装置、存储介质及电子设备
CN113033364A (zh) * 2021-03-15 2021-06-25 商汤集团有限公司 轨迹预测、行驶控制方法、装置、电子设备及存储介质
CN112990375A (zh) * 2021-04-29 2021-06-18 北京三快在线科技有限公司 一种模型训练方法、装置、存储介质及电子设备
CN113325855A (zh) * 2021-08-02 2021-08-31 北京三快在线科技有限公司 基于迁移场景用于预测障碍物轨迹的模型训练方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117874529A (zh) * 2024-03-12 2024-04-12 浪潮电子信息产业股份有限公司 运动轨迹预测方法、模型训练方法、装置、设备及介质
CN117874529B (zh) * 2024-03-12 2024-05-28 浪潮电子信息产业股份有限公司 运动轨迹预测方法、模型训练方法、装置、设备及介质

Also Published As

Publication number Publication date
CN113325855A (zh) 2021-08-31
EP4383037A1 (en) 2024-06-12
CN113325855B (zh) 2021-11-30

Similar Documents

Publication Publication Date Title
WO2023010827A1 (zh) 基于迁移场景的轨迹预测模型的训练方法及装置
US11187548B2 (en) Planning vehicle computational unit migration based on mobility prediction
KR102373448B1 (ko) 사람의 주행 데이터를 트레이닝 데이터로서 사용하여 강화 학습을 지원함으로써 맞춤형 경로 플래닝을 수행하는 학습 방법 및 학습 장치
CN112766468B (zh) 一种轨迹预测方法、装置、存储介质及电子设备
KR20200047307A (ko) 유용한 학습 데이터를 취사 선별하기 위한 cnn 기반 학습 방법 및 학습 장치 그리고 이를 이용한 테스트 방법 및 테스트 장치
CN110225368B (zh) 一种视频定位方法、装置及电子设备
EP3710993B1 (en) Image segmentation using neural networks
CN111912423B (zh) 一种预测障碍物轨迹以及模型训练的方法及装置
US11462101B2 (en) Non-essential autonomous vehicle rerouting
CN111238523B (zh) 一种运动轨迹的预测方法及装置
CN112632380A (zh) 兴趣点推荐模型的训练方法和推荐兴趣点的方法
EP3588386A1 (fr) Plate-forme d'entraînement et/ou d'assistance au contrôle aérien via un système électronique de contrôle du trafic aérien, procédé associé
US11267128B2 (en) Online utility-driven spatially-referenced data collector for classification
CN111179610B (zh) 一种交通信号设备的控制方法及装置
US11093743B2 (en) Intelligent personalization of operations of an image capturing device
CN114372736A (zh) 一种路线推荐的方法、装置、存储介质及电子设备
CN114153207A (zh) 一种无人驾驶设备的控制方法及控制装置
CN114296456A (zh) 一种网络训练以及无人驾驶设备的控制方法及装置
CN114114954A (zh) 一种无人设备的控制方法及装置
CN113469204A (zh) 数据处理方法、装置、设备和计算机存储介质
CN113741459A (zh) 确定训练样本的方法和自动驾驶模型的训练方法、装置
CN113850407A (zh) 路径规划方法、装置、存储介质及电子设备
CN115577318B (zh) 基于半实物的数据融合评估方法、系统、设备及储存介质
US11609344B2 (en) Systems and methods for utilizing a machine learning model to determine a determined location of a vehicle based on a combination of a geographical location and a visual positioning system location
CN112822045B (zh) 基于多特征混合神经网络的内容传播热区预测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22851554

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2022851554

Country of ref document: EP

Effective date: 20240304