Trajectory prediction model training method and device, equipment and medium

Abstract

The embodiment of the invention provides a trajectory prediction model training method and device, equipment and a medium, and relates to the technical field of automatic driving. The method comprises the steps that multiple kinds of scene mixed training data are used for conducting multiple times of fine adjustment on a pre-training model, multiple intermediate models are obtained, and each kind of scene mixed training data at least comprises simulation track data in one kind of scene and real track data of a real vehicle in multiple scenes; selecting a target intermediate model from the plurality of intermediate models; performing fine adjustment on the target intermediate model by using the target scene mixed training data to obtain a trajectory prediction model; wherein the target scene mixed training data comprises simulation trajectory data in a newly-added scene, simulation trajectory data, used for training the target intermediate model, in multiple scenes and real trajectory data of a real vehicle in multiple scenes. According to the embodiment of the invention, the trajectory prediction precision of multiple scenes can be improved on the basis of not changing the existing algorithm.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a trajectory prediction model training method, device, equipment and medium. Background Art

[0002] Trajectory prediction plays a very important role in the process of autonomous driving algorithms. Accurately grasping the future driving trajectory of the target can effectively reduce problems such as false braking and missed braking of the vehicle and improve algorithm performance. Traditional trajectory prediction algorithms are mostly rule-based algorithms. With the development of autonomous driving technology, their limitations become more and more obvious as the complexity of the scene increases. Therefore, using deep learning algorithms to predict trajectories has become a better technical path at present.

[0003] Since some scenes often account for a small proportion of the full-scene data and are usually relatively simple, the full-scene model generated based on the deep learning algorithm has the problem of insufficient "learning" in a certain scene. If you want to make additional optimizations for certain scenes based on the full-scene trajectory prediction deep learning algorithm, modifying the existing algorithm can be one of the methods. However, modifying the existing algorithm is time-consuming and laborious. Therefore, how to improve the trajectory prediction accuracy of multiple scenes without changing the existing algorithm is a technical problem that needs to be solved urgently in the present invention. Summary of the invention

[0004] Embodiments of the present invention provide a trajectory prediction model training method, apparatus, device and medium to improve trajectory prediction accuracy in multiple scenarios.

[0005] A first aspect of an embodiment of the present invention provides a trajectory prediction model training method, the method comprising:

[0006] The pre-trained model is fine-tuned multiple times using mixed training data of multiple scenarios to obtain multiple intermediate models, wherein the mixed training data of each scenario includes at least: simulated trajectory data of one scenario and real trajectory data of the real vehicle in multiple scenarios;

[0007] selecting a target intermediate model from the plurality of intermediate models;

[0008] The target intermediate model is fine-tuned using target scene mixed training data to obtain a trajectory prediction model; the target scene mixed training data includes: simulated trajectory data in newly added scenarios, simulated trajectory data in multiple scenarios used to train the target intermediate model, and real trajectory data of the actual vehicle in multiple scenarios.

[0009] Optionally, the pre-trained model is fine-tuned multiple times using mixed training data of multiple scenarios to obtain multiple intermediate models, including:

[0010] Using first-scenario mixed training data, fine-tuning the pre-trained model to obtain a first-scenario model, wherein the first-scenario mixed training data includes: simulated trajectory data in the first scenario and real trajectory data of the real vehicle in multiple scenarios;

[0011] The first scene model is fine-tuned using a plurality of second scene mixed training data to obtain a plurality of second scene models; each second scene mixed training data includes at least: simulated trajectory data under the second scene and real trajectory data of the real vehicle under a plurality of scenes.

[0012] Optionally, a target intermediate model is selected from the multiple intermediate models, and the target intermediate model is fine-tuned using target scene mixed training data to obtain a trajectory prediction model, including:

[0013] The target intermediate model is determined from the multiple second scene models, and the target intermediate model is fine-tuned using the third scene mixed training data to obtain the trajectory prediction model; the third scene mixed training data includes: the first scene, the second scene, the simulated trajectory data under the third scene and the real trajectory data of the actual vehicle under multiple scenes.

[0014] Optionally, the plurality of second-scenario mixed training data correspond to at least a first training strategy and a second training strategy; determining the target intermediate model from the plurality of second-scenario models, and using the third-scenario mixed training data to fine-tune the target intermediate model to obtain the trajectory prediction model, including:

[0015] Determine a target intermediate model corresponding to the first training strategy from the multiple second scene models, and use the third scene mixed training data corresponding to the first training strategy to fine-tune the target intermediate model corresponding to the first training strategy to obtain the trajectory prediction model;

[0016] In the third-scenario mixed training data corresponding to the first training strategy, the simulated trajectory data under the third scenario corresponds to a first proportion, and the sum of the real trajectory data of the actual vehicle in multiple scenarios and the simulated trajectory data under the first scenario and the second scenario corresponds to a second proportion.

[0017] Optionally, the plurality of second-scenario mixed training data correspond to at least a first training strategy and a second training strategy; determining the target intermediate model from the plurality of second-scenario models, and using the third-scenario mixed training data to fine-tune the target intermediate model to obtain the trajectory prediction model, including:

[0018] Determine a target intermediate model corresponding to the second training strategy from a plurality of the second scene models, and use the third scene mixed training data corresponding to the second training strategy to fine-tune the target intermediate model corresponding to the second training strategy to obtain the trajectory prediction model;

[0019] In the third-scenario mixed training data corresponding to the second training strategy, the simulated trajectory data under the first scenario, the second scenario, and the third scenario each correspond to a third proportion, and the actual trajectory data of the actual vehicle under multiple scenarios corresponds to a fourth proportion.

[0020] Optionally, the plurality of second scene mixed training data correspond to the first training strategy, the second training strategy and the third training strategy; and the plurality of second scene mixed training data are used to fine-tune the first scene model to obtain the plurality of second scene models, including:

[0021] Using the second scene mixed training data corresponding to the first training strategy, fine-tuning the first scene model to obtain the second scene model corresponding to the first training strategy; in the second scene mixed training data corresponding to the first training strategy, the simulated trajectory data under the second scene corresponds to the fifth proportion, and the sum of the real trajectory data of the real vehicle under multiple scenes and the simulated trajectory data under the first scene corresponds to the sixth proportion;

[0022] Using the second scene mixed training data corresponding to the second training strategy, fine-tuning the first scene model to obtain the second scene model corresponding to the second training strategy; in the second scene mixed training data corresponding to the second training strategy, the simulation trajectory data in the first scene and the second scene respectively correspond to the seventh proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the eighth proportion;

[0023] Using the second scene mixed training data corresponding to the third training strategy, the first scene model is fine-tuned to obtain the second scene model corresponding to the third training strategy; in the second scene mixed training data corresponding to the third training strategy, the simulation trajectory data under the second scene corresponds to the ninth proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the tenth proportion.

[0024] A second aspect of an embodiment of the present invention provides a trajectory prediction model training device, the device comprising:

[0025] An intermediate model acquisition module is used to fine-tune the pre-trained model multiple times using mixed training data of multiple scenarios to obtain multiple intermediate models, where each mixed training data of each scenario includes at least: simulated trajectory data in one scenario and real trajectory data of the real vehicle in multiple scenarios;

[0026] An intermediate model selection module, used to select a target intermediate model from the multiple intermediate models;

[0027] The trajectory model determination module is used to fine-tune the target intermediate model using target scene mixed training data to obtain a trajectory prediction model; the target scene mixed training data includes: simulated trajectory data in a newly added scenario, simulated trajectory data in multiple scenarios used to train the target intermediate model, and real trajectory data of the actual vehicle in multiple scenarios.

[0028] Optionally, the intermediate model acquisition module includes:

[0029] A first scenario model training module is used to fine-tune the pre-trained model using first scenario mixed training data to obtain a first scenario model, wherein the first scenario mixed training data includes: simulated trajectory data in the first scenario and real trajectory data of the real vehicle in multiple scenarios;

[0030] The second scene model training module is used to use multiple second scene mixed training data to fine-tune the first scene model to obtain multiple second scene models; each second scene mixed training data at least includes: simulated trajectory data under the second scene and real trajectory data of the real vehicle under multiple scenes.

[0031] Optionally, the intermediate model selection module and the trajectory model determination module include:

[0032] The third scene model training module is used to determine the target intermediate model from the multiple second scene models, and use the third scene mixed training data to fine-tune the target intermediate model to obtain the trajectory prediction model; the third scene mixed training data includes: the first scene, the second scene, the third scene under the simulated trajectory data and the real trajectory data of the real vehicle in multiple scenes.

[0033] Optionally, the plurality of second scene mixed training data correspond to at least the first training strategy and the second training strategy; the third scene model training module includes:

[0034] a first strategy selection module, configured to determine a target intermediate model corresponding to the first training strategy from the plurality of second scene models, and fine-tune the target intermediate model corresponding to the first training strategy using the third scene mixed training data corresponding to the first training strategy to obtain the trajectory prediction model;

[0035] In the third-scenario mixed training data corresponding to the first training strategy, the simulated trajectory data under the third scenario corresponds to a first proportion, and the sum of the real trajectory data of the actual vehicle in multiple scenarios and the simulated trajectory data under the first scenario and the second scenario corresponds to a second proportion.

[0036] Optionally, the plurality of second scene mixed training data correspond to at least the first training strategy and the second training strategy; the third scene model training module includes:

[0037] a second strategy selection module, configured to determine a target intermediate model corresponding to the second training strategy from a plurality of the second scene models, and fine-tune the target intermediate model corresponding to the second training strategy using the third scene mixed training data corresponding to the second training strategy to obtain the trajectory prediction model;

[0038] In the third-scenario mixed training data corresponding to the second training strategy, the simulated trajectory data under the first scenario, the second scenario, and the third scenario each correspond to a third proportion, and the actual trajectory data of the actual vehicle under multiple scenarios corresponds to a fourth proportion.

[0039] Optionally, the plurality of second scene mixed training data correspond to the first training strategy, the second training strategy and the third training strategy; the second scene model training module includes:

[0040] A first strategy fine-tuning module is used to fine-tune the first scenario model using the second scenario mixed training data corresponding to the first training strategy to obtain the second scenario model corresponding to the first training strategy; in the second scenario mixed training data corresponding to the first training strategy, the simulated trajectory data under the second scenario corresponds to the fifth proportion, and the sum of the real trajectory data of the real vehicle under multiple scenarios and the simulated trajectory data under the first scenario corresponds to the sixth proportion;

[0041] A second strategy fine-tuning module is used to fine-tune the first scenario model using the second scenario mixed training data corresponding to the second training strategy to obtain the second scenario model corresponding to the second training strategy; in the second scenario mixed training data corresponding to the second training strategy, the simulated trajectory data under the first scenario and the second scenario respectively correspond to the seventh proportion, and the real trajectory data of the real vehicle under multiple scenarios corresponds to the eighth proportion;

[0042] The third strategy fine-tuning module is used to use the second scene mixed training data corresponding to the third training strategy to fine-tune the first scene model to obtain the second scene model corresponding to the third training strategy; in the second scene mixed training data corresponding to the third training strategy, the simulation trajectory data under the second scene corresponds to the ninth proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the tenth proportion.

[0043] A third aspect of an embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the computer program is executed by the processor, the trajectory prediction model training method according to the first aspect of the embodiment of the present invention is implemented.

[0044] A fourth aspect of an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the trajectory prediction model training method of the first aspect of the embodiment of the present invention is implemented.

[0045] In the trajectory prediction model training method provided by the embodiment of the present invention, first, the pre-trained model is fine-tuned multiple times using mixed training data of multiple scenes to obtain multiple intermediate models; then, a target intermediate model is selected from the multiple intermediate models; finally, the target intermediate model is fine-tuned using the mixed training data of the target scene to obtain the trajectory prediction model; wherein each mixed training data of the scene includes at least: simulation trajectory data in one scene and real trajectory data of the real vehicle in multiple scenes; the target scene mixed training data includes: simulation trajectory data in a newly added scene, simulation trajectory data in multiple scenes used for training the target intermediate model, and real trajectory data of the real vehicle in multiple scenes. The embodiment of the present invention combines the simulation trajectory data in multiple scenes with the real trajectory data of the real vehicle in multiple scenes to perform multi-scene fine-tuning on the basis of the pre-trained model, so that the existing full-scene deep learning trajectory prediction algorithm in the pre-trained model does not need to be changed, and the evaluation accuracy of the pre-trained model in the full scene will not be greatly reduced. The simulation trajectory data is used to perform multi-scene fine-tuning on the basis of any full-scene deep learning trajectory prediction algorithm to optimize and improve the trajectory prediction accuracy of the current full-scene trajectory prediction model in multiple scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0047] Figure 1 is a flow chart of a trajectory prediction model training method shown in one embodiment of the present invention;

[0048] Figure 2 It is a structural block diagram of a trajectory prediction model training device provided by an embodiment of the present invention;

[0049] Figure 3 It is a schematic diagram of an electronic device shown in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.

[0051] In one embodiment, reference Figure 1 , Figure 1 FIG. 1 is a flow chart of a trajectory prediction model training method according to an embodiment of the present invention. Figure 1 As shown, the trajectory prediction model training method of this embodiment may include the following steps:

[0052] Step S11: fine-tune the pre-trained model multiple times using mixed training data of multiple scenarios to obtain multiple intermediate models, wherein each mixed training data of each scenario includes at least: simulated trajectory data in one scenario and real trajectory data of the real vehicle in multiple scenarios.

[0053] In this embodiment, the pre-trained model is a pre-trained trajectory prediction model, which can be any pre-trained full-scenario deep learning trajectory prediction model, and can be trained using real trajectory data of a real vehicle (real vehicle) in multiple scenarios.

[0054] In this embodiment, the pre-trained model can be fine-tuned multiple times using mixed training data of multiple scenes to obtain multiple intermediate models. Among them, each scene mixed training data in the multiple scene mixed training data includes at least: simulation trajectory data in one scene and real trajectory data of the real vehicle in multiple scenes, and the real trajectory data of the real vehicle in multiple scenes is the data used to train the pre-trained model. In each fine-tuning, the model can be fine-tuned based on at least one simulation trajectory data in a newly added scene and the real trajectory data of the real vehicle in multiple scenes, so that the model can optimize the trajectory prediction for the newly added scene. In multiple fine-tuning, the pre-trained model or the intermediate model obtained after the previous fine-tuning can be fine-tuned for the newly added scene, so that the intermediate model can optimize the trajectory prediction in the multiple newly added scenes targeted by the fine-tuning. This embodiment does not limit the number of fine-tuning times for obtaining multiple intermediate models, and the number of fine-tuning times can be related to the number of newly added scenes.

[0055] Step S12: Selecting a target intermediate model from the multiple intermediate models.

[0056] In this embodiment, after obtaining multiple intermediate models, a target intermediate model can be selected from the multiple intermediate models to perform subsequent trajectory prediction fine-tuning in newly added scenarios.

[0057] Step S13: Use the target scene mixed training data to fine-tune the target intermediate model to obtain a trajectory prediction model; the target scene mixed training data includes: simulated trajectory data in a newly added scenario, simulated trajectory data in multiple scenarios used to train the target intermediate model, and real trajectory data of the actual vehicle in multiple scenarios.

[0058] In this embodiment, the target scene mixed training data can be used to fine-tune the selected target intermediate model to obtain a trajectory prediction model. The target scene mixed training data includes: simulated trajectory data in newly added scenarios, simulated trajectory data in multiple scenarios used to train the target intermediate model, and real trajectory data of the real vehicle in multiple scenarios. In this embodiment, the multiple scenarios in the real trajectory data of the real vehicle in multiple scenarios used to train the pre-trained model include: the newly added scenarios targeted for each fine-tuning.

[0059] The trajectory prediction model of this embodiment can be used for trajectory prediction in newly added scenarios. Since the trajectory prediction model of this embodiment is obtained by multiple fine-tuning based on the pre-trained model, the obtained trajectory prediction model can not only perform trajectory prediction in all scenarios, but also perform more accurate trajectory prediction for newly added scenarios that have been fine-tuned multiple times.

[0060] In this embodiment, multiple fine-tuning of multiple scenes is performed on the basis of the pre-trained model by combining the simulated trajectory data in multiple scenes with the real trajectory data of the actual vehicle in multiple scenes. Therefore, there is no need to change the existing full-scene deep learning trajectory prediction algorithm in the pre-trained model, and the evaluation accuracy of the pre-trained model in the full scene will not be significantly reduced. The simulated trajectory data is used to perform multi-scene fine-tuning on the basis of any full-scene deep learning trajectory prediction algorithm, thereby optimizing and improving the trajectory prediction accuracy of the current full-scene trajectory prediction model in multiple scenes.

[0061] In combination with the above embodiments, in one implementation, an embodiment of the present invention further provides a trajectory prediction model training method. In this method, the "using multiple scene mixed training data to fine-tune the pre-trained model multiple times to obtain multiple intermediate models" in the above step S11 can specifically include steps S21 and S22:

[0062] Step S21: fine-tune the pre-trained model using the first scene mixed training data to obtain a first scene model, wherein the first scene mixed training data includes: simulated trajectory data under the first scene and real trajectory data of the real vehicle under multiple scenes.

[0063] In this embodiment, the pre-trained model can be fine-tuned using the first scene mixed training data to obtain a first scene model, and the first scene model can be used to perform trajectory prediction in the first scene. Among them, the first scene mixed training data of this embodiment includes: simulated trajectory data in the first scene and real trajectory data of the real vehicle in multiple scenes. The first scene in this embodiment can be any one of the full scenes corresponding to the pre-trained model (i.e., multiple scenes corresponding to the real trajectory data of the real vehicle in multiple scenes), and this embodiment does not limit this. The simulation trajectory data of this embodiment can be obtained through the commercial software VTD.

[0064] Step S22: using a plurality of second scene mixed training data, fine-tuning the first scene model to obtain a plurality of second scene models; each second scene mixed training data at least includes: simulation trajectory data under the second scene and real trajectory data of the real vehicle under multiple scenes.

[0065] In this embodiment, after obtaining the first scene model, multiple second scene mixed training data can be used to fine-tune the first scene model to obtain multiple second scene models, and the second scene model can be used to perform trajectory prediction in the first scene and / or the second scene. Among them, each second scene mixed training data in the multiple second scene mixed training data includes at least: simulated trajectory data in the second scene and real trajectory data of the real vehicle in multiple scenes. The second scene in this embodiment can be any scene except the first scene in the full scene corresponding to the pre-trained model. The first scene is different from the second scene, and this embodiment does not limit this.

[0066] In combination with the above embodiments, in one implementation, the present invention also provides a trajectory prediction model training method. In this method, the above steps S12 and S13 may specifically include step S31:

[0067] Step S31: Determine the target intermediate model from the multiple second scene models, use the third scene mixed training data to fine-tune the target intermediate model, and obtain the trajectory prediction model; the third scene mixed training data includes: the first scene, the second scene, the simulated trajectory data under the third scene and the real trajectory data of the real vehicle in multiple scenes.

[0068] In this embodiment, after obtaining multiple second scene models, a target intermediate model can be determined from the multiple second scene models, and then the target intermediate model can be fine-tuned using the third scene mixed training data to obtain a trajectory prediction model. The trajectory prediction model of this embodiment can be used for trajectory prediction in the first scene and / or the second scene and / or the third scene. Among them, the third scene mixed training data includes: simulated trajectory data in the first scene, the second scene, and the third scene, and real trajectory data of the real vehicle in multiple scenes. The third scene in this embodiment can be any scene except the first scene and the second scene in the full scene corresponding to the pre-training model. The first scene, the second scene, and the third scene are different, and this embodiment does not limit this.

[0069] It should be noted that this embodiment takes the trajectory prediction optimization of three scenes as an example. In fact, the present invention can fine-tune the pre-trained model for two or more scenes to achieve trajectory prediction optimization for two or more scenes. The fine-tuning of more than three scenes is the same or similar to the fine-tuning of the three scenes provided in this embodiment. They are based on the same principle, and this embodiment will not go into details one by one.

[0070] In combination with the above embodiments, in one implementation, an embodiment of the present invention further provides a trajectory prediction model training method. In this method, the multiple second scene mixed training data correspond to at least the first training strategy and the second training strategy; and the above step S31 "determine the target intermediate model from the multiple second scene models, use the third scene mixed training data, fine-tune the target intermediate model, and obtain the trajectory prediction model" may specifically include step S41:

[0071] Step S41: Determine a target intermediate model corresponding to the first training strategy from the multiple second scene models, and use the third scene mixed training data corresponding to the first training strategy to fine-tune the target intermediate model corresponding to the first training strategy to obtain the trajectory prediction model.

[0072] In this embodiment, for the training of the trajectory prediction model, there are at least a first training strategy and a second training strategy, and the first training strategy and the second training strategy correspond to different second scene mixed training data and third scene mixed training data, respectively. A plurality of second scene mixed training data correspond to at least the first training strategy and the second training strategy, that is, the plurality of second scene mixed training data at least include: the second scene mixed training data corresponding to the first training strategy, and the first scene model can be fine-tuned based on the second scene mixed training data corresponding to the first training strategy to obtain the second scene model corresponding to the first training strategy.

[0073] This embodiment can determine the second scene model corresponding to the first training strategy from multiple second scene models, that is, determine the target intermediate model corresponding to the first training strategy, use the third scene mixed training data corresponding to the first training strategy, fine-tune the target intermediate model corresponding to the first training strategy, and obtain a trained trajectory prediction model.

[0074] In this embodiment, in the mixed training data of the third scene corresponding to the first training strategy, the simulated trajectory data in the third scene corresponds to the first proportion, and the sum of the real trajectory data of the real vehicle in multiple scenes and the simulated trajectory data in the first scene and the second scene corresponds to the second proportion. For example, the simulated trajectory data in the third scene corresponds to 30% (accounting for 30% of the mixed training data of the third scene), and the sum of the real trajectory data of the real vehicle in multiple scenes and the simulated trajectory data of the first scene and the simulated trajectory data in the second scene corresponds to 70% (accounting for 70% of the mixed training data of the third scene). The above is only an example, and this embodiment does not impose any restrictions on the specific values ​​of the first proportion and the second proportion.

[0075] In combination with the above embodiments, in one implementation, an embodiment of the present invention further provides a trajectory prediction model training method. In this method, the multiple second scene mixed training data correspond to at least the first training strategy and the second training strategy; and the above step S31 "determine the target intermediate model from the multiple second scene models, use the third scene mixed training data, fine-tune the target intermediate model, and obtain the trajectory prediction model" may specifically include step S51:

[0076] Step S51: Determine a target intermediate model corresponding to the second training strategy from multiple second scene models, use the third scene mixed training data corresponding to the second training strategy to fine-tune the target intermediate model corresponding to the second training strategy, and obtain the trajectory prediction model.

[0077] In this embodiment, for the training of the trajectory prediction model, there are at least a first training strategy and a second training strategy, and the first training strategy and the second training strategy correspond to different second scene mixed training data and third scene mixed training data, respectively. A plurality of second scene mixed training data correspond to at least the first training strategy and the second training strategy, that is, the plurality of second scene mixed training data at least include: second scene mixed training data corresponding to the second training strategy, and the first scene model can be fine-tuned based on the second scene mixed training data corresponding to the second training strategy to obtain a second scene model corresponding to the second training strategy.

[0078] This embodiment can determine the second scene model corresponding to the second training strategy from multiple second scene models, that is, determine the target intermediate model corresponding to the second training strategy, use the third scene mixed training data corresponding to the second training strategy, fine-tune the target intermediate model corresponding to the second training strategy, and obtain a trained trajectory prediction model.

[0079] In this embodiment, in the third scene mixed training data corresponding to the second training strategy, the simulated trajectory data in the first scene, the second scene, and the third scene each correspond to the third proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the fourth proportion. For example, the simulated trajectory data in the first scene corresponds to 10% (accounting for 10% of the mixed training data of the third scene), the simulated trajectory data in the second scene corresponds to 10% (accounting for 10% of the mixed training data of the third scene), and the simulated trajectory data in the third scene corresponds to 10% (accounting for 10% of the mixed training data of the third scene), and the real trajectory data of the real vehicle in multiple scenes corresponds to 70% (accounting for 70% of the mixed training data of the third scene). The above is only an example. This embodiment does not impose any restrictions on the specific values ​​of the third proportion and the fourth proportion. The first proportion, the second proportion, the third proportion, and the fourth proportion may be the same or different.

[0080] In combination with the above embodiments, in one implementation, an embodiment of the present invention further provides a trajectory prediction model training method. In this method, the plurality of second scene mixed training data corresponds to the first training strategy, the second training strategy and the third training strategy; and the above step S22 may specifically include steps S61 to S63:

[0081] Step S61: Use the second scene mixed training data corresponding to the first training strategy to fine-tune the first scene model to obtain the second scene model corresponding to the first training strategy; in the second scene mixed training data corresponding to the first training strategy, the simulation trajectory data under the second scene corresponds to the fifth proportion, and the sum of the real trajectory data of the actual vehicle in multiple scenes and the simulation trajectory data under the first scene corresponds to the sixth proportion.

[0082] In this embodiment, for the training of the trajectory prediction model, at least the first training strategy, the second training strategy and the third training strategy correspond to the same first scene mixed training data, in which the first scene mixed training data corresponds to the eleventh proportion of the simulated trajectory data in the first scene, and the real trajectory data of the real vehicle in multiple scenes corresponds to the twelfth proportion; for example, in the first scene mixed training data corresponding to the first training strategy, the second training strategy and the third training strategy, the simulated trajectory data in the first scene corresponds to 30% (accounting for 30% of the first scene mixed training data), and the real trajectory data of the real vehicle in multiple scenes corresponds to 70% (accounting for 70% of the first scene mixed training data). In addition, the first training strategy, the second training strategy and the third training strategy correspond to different second scene mixed training data and third scene mixed training data, respectively.

[0083] The first scene model can be fine-tuned using the second scene mixed training data corresponding to the first training strategy to obtain the second scene model corresponding to the first training strategy. Among them, in the second scene mixed training data corresponding to the first training strategy, the simulation trajectory data under the second scene corresponds to the fifth proportion, and the sum of the real trajectory data of the real vehicle under multiple scenes and the simulation trajectory data under the first scene corresponds to the sixth proportion. For example, the simulation trajectory data under the second scene corresponds to 30% (accounting for 30% of the second scene mixed training data), and the sum of the real trajectory data of the real vehicle under multiple scenes and the simulation trajectory data of the first scene corresponds to 70% (accounting for 70% of the second scene mixed training data). The above is only an example. This embodiment does not impose any restrictions on the specific values ​​of the fifth proportion and the sixth proportion, and the first proportion, the second proportion, the third proportion, the fourth proportion, the fifth proportion and the sixth proportion may be the same or different.

[0084] Step S62: Use the second scene mixed training data corresponding to the second training strategy to fine-tune the first scene model to obtain the second scene model corresponding to the second training strategy; in the second scene mixed training data corresponding to the second training strategy, the simulation trajectory data under the first scene and the second scene respectively correspond to the seventh proportion, and the real trajectory data of the actual vehicle in multiple scenes corresponds to the eighth proportion.

[0085] In this embodiment, the first scene model can be fine-tuned using the second scene mixed training data corresponding to the second training strategy to obtain the second scene model corresponding to the second training strategy. Among them, in the second scene mixed training data corresponding to the second training strategy, the simulation trajectory data under the first scene and the simulation trajectory data under the second scene each correspond to the seventh proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the eighth proportion. For example, the simulation trajectory data under the first scene corresponds to 10% (accounting for 10% of the second scene mixed training data), the simulation trajectory data under the second scene corresponds to 20% (accounting for 20% of the second scene mixed training data), and the real trajectory data of the real vehicle in multiple scenes corresponds to 70% (accounting for 70% of the second scene mixed training data). The above is only an example. This embodiment does not impose any restrictions on the specific values ​​of the seventh proportion and the eighth proportion. The first proportion, the second proportion, the third proportion, the fourth proportion, the fifth proportion, the sixth proportion, the seventh proportion and the eighth proportion can be the same or different.

[0086] Step S63: Use the second scene mixed training data corresponding to the third training strategy to fine-tune the first scene model to obtain the second scene model corresponding to the third training strategy; in the second scene mixed training data corresponding to the third training strategy, the simulation trajectory data under the second scene corresponds to the ninth proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the tenth proportion.

[0087] In this embodiment, the first scene model can be fine-tuned using the second scene mixed training data corresponding to the third training strategy to obtain the second scene model corresponding to the third training strategy. Among them, in the second scene mixed training data corresponding to the third training strategy, the simulation trajectory data under the second scene corresponds to the ninth proportion, and the real trajectory data of the real vehicle under multiple scenes corresponds to the tenth proportion. For example, the simulation trajectory data under the second scene corresponds to 30% (accounting for 20% of the second scene mixed training data), and the real trajectory data of the real vehicle under multiple scenes corresponds to 70% (accounting for 70% of the second scene mixed training data). The above is only an example. This embodiment does not impose any restrictions on the specific values ​​of the ninth proportion and the tenth proportion. The first proportion, the second proportion, the third proportion, the fourth proportion, the fifth proportion, the sixth proportion, the seventh proportion, the eighth proportion, the ninth proportion and the tenth proportion may be the same or different, respectively.

[0088] In an optional example, taking the first scene as the entry scene, the second scene as the import scene, and the third scene as the export scene as an example, the following are specific examples corresponding to the first training strategy, the second training strategy, and the third training strategy:

[0089] First training strategy: In the first scenario mixed training data of the first fine-tuning, the original real vehicle full-scene data (i.e., the real trajectory data of the real vehicle in multiple scenarios) accounts for 70% of the total sample volume, and the simulation cut-in scenario data (i.e., the simulated trajectory data in the first scenario) accounts for 30% of the total sample volume. In the second scenario mixed training data of the second fine-tuning, the sum of the original real vehicle full-scene data plus the simulation cut-in scenario data accounts for 70% of the total sample volume, and the simulation imported scenario data accounts for 30% of the total sample volume (i.e., the simulated trajectory data in the second scenario). In the third scenario mixed training data of the third fine-tuning, the original real vehicle full-scene data, the simulation cut-in scenario data, and the simulation imported scenario data account for 70% of the total sample volume, and the simulation exported scenario data (i.e., the simulated trajectory data in the third scenario) accounts for 30% of the total sample volume.

[0090] Second training strategy: In the first scenario mixed training data of the first fine-tuning, the original real vehicle full-scene data (i.e., the real trajectory data of the real vehicle in multiple scenarios) accounts for 70% of the total sample volume, and the simulation cut-in scenario data (i.e., the simulated trajectory data in the first scenario) accounts for 30% of the total sample volume. In the second scenario mixed training data of the second fine-tuning, the original real vehicle full-scene data accounts for 70% of the total sample volume, the simulation cut-in scenario data accounts for 10% of the total sample volume, and the simulation imported scenario data (i.e., the simulated trajectory data in the second scenario) accounts for 20% of the total sample volume; in the third scenario mixed training data of the third fine-tuning, the original real vehicle full-scene data accounts for 70% of the total sample volume, the simulation cut-in scenario data accounts for 10% of the total sample volume, the simulation imported scenario data accounts for 10% of the total sample volume, and the simulation exported scenario data (i.e., the simulated trajectory data in the third scenario) accounts for 10% of the total sample volume.

[0091] The third training strategy: In the first-scenario mixed training data of the first fine-tuning, the original real vehicle full-scenario data (i.e., the real trajectory data of the real vehicle in multiple scenarios) accounts for 70% of the total sample volume, and the simulation cut-in scene data (i.e., the simulated trajectory data in the first scenario) accounts for 30% of the total sample volume. In the second-scenario mixed training data of the second fine-tuning, the original real vehicle full-scenario data accounts for 70% of the total sample volume, and the simulation import scene data (i.e., the simulated trajectory data in the second scenario) accounts for 30% of the total sample volume. In the third-scenario mixed training data of the third fine-tuning, the original real vehicle full-scenario data accounts for 70% of the total sample volume, and the simulation export scene data (i.e., the simulated trajectory data in the third scenario) accounts for 30% of the total sample volume.

[0092] Among them, the cut-in scene data, the imported scene data, and the exported scene data are all simulation data, which can be implemented by the commercial software VTD; the original real vehicle full-scene data is real vehicle data, which can be collected by engineers through real vehicles. For the total sample (i.e., simulation data or real vehicle data), this embodiment is divided into training set train80%, validation set val10%, and test set test10%. In order to improve the trajectory prediction accuracy of the model in multiple scenarios, while not changing the existing algorithm and basically maintaining the original full-scene evaluation accuracy of the deep learning model, this embodiment designs the above three training strategies and obtains the following experimental results, as shown in Tables 1, 2 and 3 respectively.

[0093] Table 1 Experimental results corresponding to the first training strategy

[0094]

[0095] Table 2 Experimental results corresponding to the second training strategy

[0096]

[0097] Table 3 Experimental results corresponding to the third training strategy

[0098]

[0099] Among them, the mixed test in Tables 1, 2 and 3 refers to: the total sample data set of the second scene mixed training data in the second fine-tuning, which is then divided into 80% for training set, 10% for validation set and 10% for test set. The real car model is a pre-trained model trained with the original real car full scene data. The entry model is the result of the first fine-tuning: the first scene model, and the import model is the result of the second fine-tuning: the second scene model. In the prediction evaluation, ade is the variance, fde is the gap between the last trajectory points, and acc is the prediction accuracy; the lower ade and fde, the higher acc, which means the more accurate the trajectory prediction.

[0100] By comparing and analyzing the three training strategies, we can draw the following conclusions: 1. Since the three training strategies are the same in the first fine-tuning, the model indicators (entry model) obtained are consistent; starting from the second fine-tuning, the three training strategies are different, so the model indicators (input model) obtained are inconsistent, but the model indicators of the three training strategies are not much different.

[0101] The test results of the 2.3 models on the real car test show that the model indicators have all decreased slightly, but the difference is very small, indicating that fine-tuning the original full-scene model using simulation data can maintain the original full-scene evaluation accuracy.

[0102] 3. The test results of the three models on the cut-in test show that the imported model obtained by the second fine-tuning of the third training strategy has not learned the simulated cut-in scene data, so there is a "forgetting" phenomenon; the first training strategy and the second training strategy can better maintain the evaluation accuracy.

[0103] 4. The test results of the three models on the imported test can also prove the above conclusion 3, and in comparison, the evaluation accuracy of the first training strategy is better.

[0104] 5. According to Conclusion 3 and Conclusion 4, and by comparing the test results of the first training strategy and the second training strategy on the mixed test, although the proportion of original real vehicle full-scene data in the total sample volume decreases with the increase of simulation data volume for the first training strategy, it does not affect its evaluation accuracy. Therefore, it is recommended to select the first training strategy and the second training strategy.

[0105] In an optional example, the first training strategy is then selected to continue the third fine-tuning to obtain the export model obtained by the third fine-tuning (ie, the third scene model, the trajectory prediction model). The experimental results are shown in Table 4 below:

[0106] Table 4 Experimental results of the third fine-tuning corresponding to the first training strategy

[0107]

[0108] As can be seen from Table 4, the export model obtained after three fine-tunings of the first training strategy has no significant decrease in test accuracy on the real vehicle test, and can maintain the trajectory prediction accuracy of the first two fine-tunings, and significantly improve the trajectory prediction accuracy of the model in the cut-in scene, import scene, and export scene. Therefore, the experimental results prove that this embodiment can optimize and improve the trajectory prediction accuracy of multiple scenes on the basis of the original full-scene model when using deep learning for trajectory prediction, and provide an effective training model.

[0109] In combination with the above embodiments, in one implementation, an embodiment of the present invention further provides a trajectory prediction model training method. In this method, in the first training strategy, the scene mixed training data used in the previous fine-tuning accounts for the thirteenth percentage of the scene mixed training data used in the current fine-tuning, and the simulation trajectory data under the newly added scene in the current fine-tuning accounts for the fourteenth percentage of the scene mixed training data used in the current fine-tuning; in the second training strategy, the total amount of scene mixed training data used for each fine-tuning remains unchanged.

[0110] The following is a complete implementation and an example of the specific sample size:

[0111] Step 1: Process the full-scene data of the real vehicle collected by the engineers to obtain a total of 185,776 valid samples. The 185,776 total samples are divided into 80% for the training set, 10% for the validation set, and 10% for the test set, resulting in 150,396 samples for the real vehicle training set, 17,690 samples for the real vehicle validation set, and 17,690 samples for the real vehicle test set. Use the training set and validation set to train the existing deep learning algorithm, save the optimal weight of the model, and obtain the pre-trained model, i.e., the real vehicle model.

[0112] Step 2: Build the first scene (entry scene), the second scene (import scene) and the third scene (export scene), and use the commercial software VTD to obtain batch simulation data in different scenes. Perform data cleaning on the simulation data to obtain valid samples. Divide and mix the samples according to the training set train80%, the validation set val10%, and the test set test10%. The specific data volume of each training strategy is shown in Tables 5, 6 and 7 below:

[0113] Table 5. Partition results of the first training strategy

[0114] Real car Cut in Mix 1 Import Mix 2 Export Mix 3 Train 150396 60781 211177 89739 300916 128207 429123 val 17690 7590 25280 11217 36497 16020 52517 test 17690 7590 25280 11217 36497 16020 52517

[0115] Table 6 Partition results of the second training strategy

[0116] Real car Cut in Mix 1 Import Mix 2 Export Mix 3 Train 150396 60781 211177 40512 211177 20256 211177 val 17690 7590 25280 5064 25280 2532 25280 test 17690 7590 25280 5064 25280 2532 25280

[0117] Table 7. Partition results of the third training strategy

[0118] Real car Cut in Mix 1 Import Mix 2 Export Mix 3 Train 150396 60781 211177 60781 211177 60781 211177 val 17690 7590 25280 7590 25280 7590 25280 test 17690 7590 25280 7590 25280 7590 25280

[0119] Among them, it can be seen from Table 5 that in the first training strategy, the scene mixed training data used in the previous fine-tuning accounts for the thirteenth proportion (70%) of the scene mixed training data used in the current fine-tuning, and the simulation trajectory data under the newly added scene in the current fine-tuning accounts for the fourteenth proportion (30%) of the scene mixed training data used in the current fine-tuning. And, it can be seen from Table 6 that in the second training strategy, the total amount of scene mixed training data used in each fine-tuning remains unchanged (all 211177), and the proportion of simulation trajectory data for different scenes is adjusted when the total amount of data remains unchanged.

[0120] Step three, the train and val of the mixed 1 data of the three training strategies in step two are fine-tuned for the first time based on the real vehicle model to obtain the entry model (i.e., the first scenario model), and then the train and val of the mixed 2 data are fine-tuned for the second time based on the entry model to obtain the input model (i.e., the second scenario model), and so on. Of course, more scenarios can be added, which will not be elaborated here.

[0121] Step 4: After obtaining the cut-in model and import model of each training strategy, we tested them on the real vehicle test, cut-in test, import test and mixed 2test, and obtained a total of 48 groups of test indicators. By comparing the test indicators of the three training strategies, we concluded that the first training strategy and the second training strategy have better trajectory prediction accuracy when fine-tuning multiple scenarios.

[0122] Step 5, select the first training strategy, and use the train and val of the mixed 3 data to perform the third fine-tuning on the import model based on the first training strategy to obtain the export model (i.e., the third scene model, i.e., the trajectory prediction model). Then use the export model to test on the real vehicle test, cut-in test, import test, export test, and mixed 2test. At the same time, the real vehicle model, cut-in model, and import model are tested on the export test. Through indicator analysis, the export model (i.e., trajectory prediction model) obtained after three fine-tunings has no significant decrease in test accuracy on the real vehicle test, and can maintain the trajectory prediction accuracy of the first two fine-tunings, and significantly improve the trajectory prediction accuracy of the model in the cut-in scene, import scene, and export scene. This proves that on the basis of the original full-scene model, when using deep learning for trajectory prediction, it is effective to optimize and improve the trajectory prediction accuracy of multiple scenes by fine-tuning the simulation data without changing the original algorithm. Through the trajectory prediction model training method provided in this embodiment, multiple scene fine-tuning can be performed on the basis of any full-scene deep learning trajectory prediction algorithm. The advantage is that there is no need to change the existing algorithm, and the evaluation accuracy of the algorithm model in the whole scene will not be reduced. At the same time, it can effectively improve the trajectory prediction accuracy of multiple scenes.

[0123] It should be noted that, for the sake of simplicity, the method embodiments are described as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present invention are not limited by the order of the actions described, because according to the embodiments of the present invention, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present invention.

[0124] Based on the same inventive concept, an embodiment of the present invention provides a trajectory prediction model training device 200. Figure 2 , Figure 2 is a structural block diagram of a trajectory prediction model training device provided by an embodiment of the present invention. Figure 2 As shown, the device 200 includes:

[0125] The intermediate model acquisition module 201 is used to fine-tune the pre-trained model multiple times using mixed training data of multiple scenarios to obtain multiple intermediate models, and each mixed training data of each scenario at least includes: simulation trajectory data in one scenario and real trajectory data of the real vehicle in multiple scenarios;

[0126] An intermediate model selection module 202, configured to select a target intermediate model from the plurality of intermediate models;

[0127] The trajectory model determination module 203 is used to fine-tune the target intermediate model using target scene mixed training data to obtain a trajectory prediction model; the target scene mixed training data includes: simulated trajectory data in a newly added scenario, simulated trajectory data in multiple scenarios used to train the target intermediate model, and real trajectory data of the actual vehicle in multiple scenarios.

[0128] Optionally, the intermediate model obtaining module 201 includes:

[0129] A first scenario model training module is used to fine-tune the pre-trained model using first scenario mixed training data to obtain a first scenario model, wherein the first scenario mixed training data includes: simulated trajectory data in the first scenario and real trajectory data of the real vehicle in multiple scenarios;

[0130] The second scene model training module is used to use multiple second scene mixed training data to fine-tune the first scene model to obtain multiple second scene models; each second scene mixed training data at least includes: simulated trajectory data under the second scene and real trajectory data of the real vehicle under multiple scenes.

[0131] Optionally, the intermediate model selection module 202 and the trajectory model determination module 203 include:

[0132] The third scene model training module is used to determine the target intermediate model from the multiple second scene models, and use the third scene mixed training data to fine-tune the target intermediate model to obtain the trajectory prediction model; the third scene mixed training data includes: the first scene, the second scene, the third scene under the simulated trajectory data and the real trajectory data of the real vehicle in multiple scenes.

[0133] Optionally, the plurality of second scene mixed training data correspond to at least the first training strategy and the second training strategy; the third scene model training module includes:

[0134] a first strategy selection module, configured to determine a target intermediate model corresponding to the first training strategy from the plurality of second scene models, and fine-tune the target intermediate model corresponding to the first training strategy using the third scene mixed training data corresponding to the first training strategy to obtain the trajectory prediction model;

[0135] In the third-scenario mixed training data corresponding to the first training strategy, the simulated trajectory data under the third scenario corresponds to a first proportion, and the sum of the real trajectory data of the actual vehicle in multiple scenarios and the simulated trajectory data under the first scenario and the second scenario corresponds to a second proportion.

[0136] Optionally, the plurality of second scene mixed training data correspond to at least the first training strategy and the second training strategy; the third scene model training module includes:

[0137] a second strategy selection module, configured to determine a target intermediate model corresponding to the second training strategy from a plurality of the second scene models, and fine-tune the target intermediate model corresponding to the second training strategy using the third scene mixed training data corresponding to the second training strategy to obtain the trajectory prediction model;

[0138] In the third-scenario mixed training data corresponding to the second training strategy, the simulated trajectory data under the first scenario, the second scenario, and the third scenario each correspond to a third proportion, and the actual trajectory data of the actual vehicle under multiple scenarios corresponds to a fourth proportion.

[0139] Optionally, the plurality of second scene mixed training data correspond to the first training strategy, the second training strategy and the third training strategy; the second scene model training module includes:

[0140] A first strategy fine-tuning module is used to fine-tune the first scenario model using the second scenario mixed training data corresponding to the first training strategy to obtain the second scenario model corresponding to the first training strategy; in the second scenario mixed training data corresponding to the first training strategy, the simulated trajectory data under the second scenario corresponds to the fifth proportion, and the sum of the real trajectory data of the real vehicle under multiple scenarios and the simulated trajectory data under the first scenario corresponds to the sixth proportion;

[0141] A second strategy fine-tuning module is used to fine-tune the first scenario model using the second scenario mixed training data corresponding to the second training strategy to obtain the second scenario model corresponding to the second training strategy; in the second scenario mixed training data corresponding to the second training strategy, the simulated trajectory data under the first scenario and the second scenario respectively correspond to the seventh proportion, and the real trajectory data of the real vehicle under multiple scenarios corresponds to the eighth proportion;

[0142] The third strategy fine-tuning module is used to use the second scene mixed training data corresponding to the third training strategy to fine-tune the first scene model to obtain the second scene model corresponding to the third training strategy; in the second scene mixed training data corresponding to the third training strategy, the simulation trajectory data under the second scene corresponds to the ninth proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the tenth proportion.

[0143] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps in the trajectory prediction model training method described in any of the above embodiments of the present invention are implemented.

[0144] Based on the same inventive concept, another embodiment of the present invention provides an electronic device 300, such as Figure 3 shown. Figure 3 30 is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device includes a memory 302, a processor 301, and a computer program stored in the memory and executable on the processor, and when executed by the processor, the steps in the trajectory prediction model training method according to any of the above embodiments of the present invention are implemented.

[0145] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0146] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0147] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, embodiments of the present invention may take the form of complete hardware embodiments, complete software embodiments, or embodiments combining software and hardware. Moreover, embodiments of the present invention may take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0148] The embodiments of the present invention are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0149] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0150] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0151] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0152] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0153] The above is a detailed introduction to a trajectory prediction model training method, device, equipment and medium provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A trajectory prediction model training method, characterized in that: The method comprises: The pre-trained model is fine-tuned multiple times using mixed training data of multiple scenarios to obtain multiple intermediate models, wherein the mixed training data of each scenario includes at least: simulated trajectory data of one scenario and real trajectory data of the real vehicle in multiple scenarios; selecting a target intermediate model from the plurality of intermediate models; The target intermediate model is fine-tuned using target scene mixed training data to obtain a trajectory prediction model; the target scene mixed training data includes: simulated trajectory data in newly added scenarios, simulated trajectory data in multiple scenarios used to train the target intermediate model, and real trajectory data of the actual vehicle in multiple scenarios.

2. The method according to claim 1, characterized in that The pre-trained model is fine-tuned multiple times using mixed training data of multiple scenarios to obtain multiple intermediate models, including: Using first-scenario mixed training data, fine-tuning the pre-trained model to obtain a first-scenario model, wherein the first-scenario mixed training data includes: simulated trajectory data in the first scenario and real trajectory data of the real vehicle in multiple scenarios; The first scene model is fine-tuned using a plurality of second scene mixed training data to obtain a plurality of second scene models; each second scene mixed training data includes at least: simulated trajectory data under the second scene and real trajectory data of the real vehicle under a plurality of scenes.

3. The method according to claim 2, characterized in that Selecting a target intermediate model from the multiple intermediate models, and fine-tuning the target intermediate model using target scene mixed training data to obtain a trajectory prediction model, including: The target intermediate model is determined from the multiple second scene models, and the target intermediate model is fine-tuned using the third scene mixed training data to obtain the trajectory prediction model; the third scene mixed training data includes: the first scene, the second scene, the simulated trajectory data under the third scene and the real trajectory data of the actual vehicle under multiple scenes.

4. The method according to claim 3, characterized in that The plurality of second-scenario mixed training data correspond to at least the first training strategy and the second training strategy; determining the target intermediate model from the plurality of second-scenario models, and fine-tuning the target intermediate model using the third-scenario mixed training data to obtain the trajectory prediction model, including: Determine a target intermediate model corresponding to the first training strategy from the multiple second scene models, and use the third scene mixed training data corresponding to the first training strategy to fine-tune the target intermediate model corresponding to the first training strategy to obtain the trajectory prediction model; In the third-scenario mixed training data corresponding to the first training strategy, the simulated trajectory data under the third scenario corresponds to a first proportion, and the sum of the real trajectory data of the actual vehicle in multiple scenarios and the simulated trajectory data under the first scenario and the second scenario corresponds to a second proportion.

5. The method according to claim 3, characterized in that: The plurality of second-scenario mixed training data correspond to at least the first training strategy and the second training strategy; determining the target intermediate model from the plurality of second-scenario models, and fine-tuning the target intermediate model using the third-scenario mixed training data to obtain the trajectory prediction model, including: Determine a target intermediate model corresponding to the second training strategy from a plurality of the second scene models, and use the third scene mixed training data corresponding to the second training strategy to fine-tune the target intermediate model corresponding to the second training strategy to obtain the trajectory prediction model; In the third-scenario mixed training data corresponding to the second training strategy, the simulated trajectory data under the first scenario, the second scenario, and the third scenario each correspond to a third proportion, and the actual trajectory data of the actual vehicle under multiple scenarios corresponds to a fourth proportion.

6. The method according to claim 4 or 5, characterized in that: The plurality of second scenario mixed training data correspond to the first training strategy, the second training strategy and the third training strategy; The method of fine-tuning the first scene model using a plurality of second scene mixed training data to obtain a plurality of second scene models includes: Using the second scene mixed training data corresponding to the first training strategy, fine-tuning the first scene model to obtain the second scene model corresponding to the first training strategy; in the second scene mixed training data corresponding to the first training strategy, the simulated trajectory data under the second scene corresponds to the fifth proportion, and the sum of the real trajectory data of the real vehicle under multiple scenes and the simulated trajectory data under the first scene corresponds to the sixth proportion; Using the second scene mixed training data corresponding to the second training strategy, fine-tuning the first scene model to obtain the second scene model corresponding to the second training strategy; in the second scene mixed training data corresponding to the second training strategy, the simulation trajectory data in the first scene and the second scene respectively correspond to the seventh proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the eighth proportion; Using the second scene mixed training data corresponding to the third training strategy, the first scene model is fine-tuned to obtain the second scene model corresponding to the third training strategy; in the second scene mixed training data corresponding to the third training strategy, the simulation trajectory data under the second scene corresponds to the ninth proportion, and the real trajectory data of the real vehicle in multiple scenes corresponds to the tenth proportion.

7. A trajectory prediction model training device, characterized in that: The device comprises: An intermediate model acquisition module is used to fine-tune the pre-trained model multiple times using mixed training data of multiple scenarios to obtain multiple intermediate models, where each mixed training data of each scenario includes at least: simulated trajectory data in one scenario and real trajectory data of the real vehicle in multiple scenarios; An intermediate model selection module, used to select a target intermediate model from the multiple intermediate models; The trajectory model determination module is used to fine-tune the target intermediate model using target scene mixed training data to obtain a trajectory prediction model; the target scene mixed training data includes: simulated trajectory data in a newly added scenario, simulated trajectory data in multiple scenarios used to train the target intermediate model, and real trajectory data of the actual vehicle in multiple scenarios.

8. The device according to claim 7, characterized in that The intermediate model acquisition module includes: A first scenario model training module is used to fine-tune the pre-trained model using first scenario mixed training data to obtain a first scenario model, wherein the first scenario mixed training data includes: simulated trajectory data in the first scenario and real trajectory data of the real vehicle in multiple scenarios; The second scene model training module is used to use multiple second scene mixed training data to fine-tune the first scene model to obtain multiple second scene models; each second scene mixed training data at least includes: simulated trajectory data under the second scene and real trajectory data of the real vehicle under multiple scenes.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is executed by the processor, the trajectory prediction model training method according to any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the trajectory prediction model training method as described in any one of claims 1 to 6 is implemented.

Images