CN114792148A

CN114792148A - Method and device for predicting motion trail

Info

Publication number: CN114792148A
Application number: CN202110099754.7A
Authority: CN
Inventors: 蒋竺希; 张驰
Original assignee: Momenta Suzhou Technology Co Ltd
Current assignee: Momenta Suzhou Technology Co Ltd
Priority date: 2021-01-25
Filing date: 2021-01-25
Publication date: 2022-07-26
Also published as: WO2022156181A1

Abstract

The embodiment of the invention discloses a method and a device for predicting a motion trail, wherein the method comprises the following steps: obtaining historical track and motion attribute information of each traffic participating object corresponding to the target object and corresponding current map information; determining the track prediction characteristics corresponding to each traffic participant by using the characteristic extraction layer of the target track prediction model and the initial characteristics corresponding to each traffic participant; determining hidden random variable multi-modal probability distribution corresponding to each traffic participant by utilizing a feature extraction layer of a target track prediction model and track prediction features corresponding to each traffic participant; and determining the multi-modal predicted track corresponding to each traffic participating object by utilizing the characteristic regression layer of the target track prediction model, the track prediction characteristics corresponding to each traffic participating object and the hidden random variable multi-modal probability distribution corresponding to each traffic participating object so as to reduce the limitation of motion track prediction and better adapt to a more complex automatic driving scene.

Description

Method and device for predicting motion trail

Technical Field

The invention relates to the technical field of trajectory prediction, in particular to a method and a device for predicting a motion trajectory.

Background

In the field of automatic driving, an automatic driving vehicle needs to refer to a future motion track of surrounding traffic participating objects during driving to plan a driving track of the vehicle so as to ensure the safety of the vehicle and the traffic participating objects. Correspondingly, the fact that the automatic driving vehicle can accurately predict the future motion trail of the traffic participation object in time is very important.

Considering that the future behaviors of the traffic participant have obvious uncertainty, that is, the future motion trajectory of the traffic participant has obvious uncertainty, in the related art, the probability distribution of the future motion trajectory of the traffic participant is generally modeled by a mixed gaussian distribution; and the modal class and the classification rule of the future motion track are designed manually, a neural network model is obtained by training in a classification and regression mode, and the future motion track corresponding to each modal class of each traffic participant and the probability corresponding to the future motion track are obtained by predicting through the neural network model, the historical motion track of each traffic participant and other corresponding dynamic information and static information. The other dynamic information corresponding to the traffic participant may include other traffic participants except the traffic participant and the historical track of the autonomous vehicle, and the static information may include current map information corresponding to the autonomous vehicle.

In the process, the number of types and the classification rule of the modal of the future motion trajectory predicted by the neural network model need to be designed manually, so that the predicted future motion trajectory has limitations to a certain extent and is difficult to adapt to a complex automatic driving scene.

Disclosure of Invention

The invention provides a motion trail prediction method and a motion trail prediction device, which are used for reducing the limitation of motion trail prediction and better adapting to a more complex automatic driving scene. The specific technical scheme is as follows:

in a first aspect, an embodiment of the present invention provides a method for predicting a motion trajectory, where the method includes:

acquiring historical track and motion attribute information of each traffic participating object corresponding to the target object and corresponding current map information;

determining the track prediction characteristics corresponding to the traffic participation objects by utilizing the characteristic extraction layer of the target track prediction model and the initial characteristics corresponding to the traffic participation objects, wherein the initial characteristics corresponding to the traffic participation objects comprise: historical track and motion attribute information of the traffic participating objects, historical track and motion attribute information of other corresponding traffic participating objects and target objects, and the current map information;

determining hidden random variable multi-modal probability distribution corresponding to each traffic participating object by utilizing a feature extraction layer of a target track prediction model and track prediction features corresponding to each traffic participating object, wherein the hidden random variables represent behavior randomness of each traffic participating object;

and determining the multi-modal predicted track corresponding to each traffic participant by utilizing the characteristic regression layer of the target track prediction model, the track prediction characteristic corresponding to each traffic participant and the hidden random variable multi-modal probability distribution corresponding to each traffic participant.

Optionally, the initial features corresponding to the traffic participation objects are features arranged based on a time sequence, and include features of a plurality of historical moments corresponding to the traffic participation objects;

the step of determining the track prediction characteristics corresponding to each traffic participant by using the characteristic extraction layer of the target track prediction model and the initial characteristics corresponding to each traffic participant comprises the following steps:

aiming at each traffic participant, circularly executing the following steps A-C for a plurality of times on the initial characteristics corresponding to the traffic participant by using the characteristic extraction layer of the target track prediction model, and determining the intermediate prediction characteristics corresponding to the traffic participant;

for each traffic participating object, fusing intermediate prediction features corresponding to each static object in the intermediate prediction features corresponding to the traffic participating object based on a graph neural network, and determining track prediction features corresponding to the traffic participating object, wherein the static objects comprise each static object in the current map information;

step A: carrying out nonlinear mapping on the to-be-processed features corresponding to the traffic participation objects from feature dimensions to obtain mapping features corresponding to the traffic participation objects, wherein the to-be-processed features are initial features corresponding to the traffic participation objects or intermediate prediction features corresponding to the traffic participation objects generated in the previous iteration;

and B, step B: performing feature aggregation operation on the mapping features from a time dimension to obtain aggregation features corresponding to the traffic participation objects;

step C: and fusing the aggregation characteristics with the characteristics of the to-be-processed characteristics at each historical moment.

Optionally, the step of determining the hidden random variable multi-modal probability distribution corresponding to each traffic participant by using the feature extraction layer of the target trajectory prediction model and the trajectory prediction features corresponding to each traffic participant includes:

aiming at each traffic participant, determining hidden random variable monomodal probability distribution corresponding to the traffic participant by utilizing a feature extraction layer of a target track prediction model and track prediction features corresponding to the traffic participant;

aiming at each traffic participant, a normalized stream mapping algorithm and hidden random variable monomodal probability distribution corresponding to the traffic participant are utilized to obtain hidden random variable multimodal probability distribution corresponding to the traffic participant.

Optionally, before the step of determining the trajectory prediction feature corresponding to each traffic participant by using the feature extraction layer of the target trajectory prediction model and the initial feature corresponding to each traffic participant, the method further includes:

training a process of obtaining a target trajectory prediction model, wherein the process comprises:

obtaining an initial trajectory prediction model;

obtaining sample training information corresponding to each sample traffic object and a sample future track corresponding to each sample traffic object, wherein the sample training information corresponding to the sample traffic object comprises: sample historical track and sample motion attribute information of the sample traffic object, and sample historical track, sample motion attribute information and sample static object information of a corresponding sample dynamic object;

for each sample traffic object, determining a sample prediction feature corresponding to the sample traffic object by using a feature extraction layer of the initial trajectory prediction model and an initial sample feature corresponding to the sample traffic object, wherein the initial sample feature corresponding to the sample traffic object comprises: sample traffic object sample historical track and sample motion attribute information, and sample historical track, sample motion attribute information and sample static object information of a corresponding sample dynamic object;

aiming at each sample traffic object, determining hidden random variable multi-modal probability distribution corresponding to the sample traffic object by utilizing a feature extraction layer of the initial track prediction model and sample prediction features corresponding to the sample traffic object;

for each sample traffic object, determining a multi-modal predicted track corresponding to the sample traffic object by utilizing a feature regression layer of the initial track prediction model, sample predicted features corresponding to the sample traffic object and hidden random variable multi-modal probability distribution corresponding to the sample traffic object;

processing a sample future track corresponding to each sample traffic object by using a preset variation algorithm to obtain hidden random variable variation probability distribution corresponding to the sample traffic object;

aiming at each sample traffic object, determining a KL divergence value of a hidden random variable corresponding to the sample traffic object by using the multimode probability distribution of the hidden random variable corresponding to the sample traffic object and the variational probability distribution of the hidden random variable corresponding to the sample traffic object;

for each sample traffic object, determining a track reconstruction loss value corresponding to the sample traffic object by using a multi-modal predicted track, hidden random variable variation probability distribution and a sample future track corresponding to the sample traffic object;

aiming at each sample traffic object, constructing a variation lower bound of a maximized likelihood function by utilizing a KL divergence value of a hidden random variable corresponding to the sample traffic object and a track reconstruction loss value corresponding to the sample traffic object; judging whether the variation lower bound of the constructed maximum likelihood function reaches the maximum;

if the variation lower bound of the constructed maximum likelihood function does not reach the maximum, adjusting model parameters of a feature extraction layer and a feature regression layer of the initial track prediction model, returning to the step of determining sample prediction features corresponding to the sample traffic objects by using the feature extraction layer of the initial track prediction model and the initial sample features corresponding to the sample traffic objects;

and if the variation lower bound of the constructed maximum likelihood function reaches the maximum, determining the convergence of the initial trajectory prediction model to obtain the target trajectory prediction model comprising a feature extraction layer and a feature regression layer.

Optionally, the step of determining the multi-modal predicted trajectory corresponding to each traffic participant by using the feature regression layer of the target trajectory prediction model, the trajectory prediction features corresponding to each traffic participant, and the hidden random variable multi-modal probability distribution corresponding to each traffic participant includes:

sampling hidden random variable multi-mode probability distribution corresponding to the traffic participation object aiming at each traffic participation object to obtain a plurality of hidden random variable samples corresponding to the traffic participation object;

and determining a multi-modal predicted track corresponding to each traffic participant by using the characteristic regression layer of the target track prediction model, the track prediction characteristics corresponding to each traffic participant and a plurality of hidden random variable samples corresponding to each traffic participant.

In a second aspect, an embodiment of the present invention provides an apparatus for predicting a motion trajectory, where the apparatus includes:

the obtaining module is configured to obtain historical track and motion attribute information of each traffic participation object corresponding to the target object and corresponding current map information;

the first determining module is configured to determine a track prediction feature corresponding to each traffic participant by using the feature extraction layer of the target track prediction model and the initial feature corresponding to each traffic participant, wherein the initial feature corresponding to the traffic participant comprises: historical track and motion attribute information of the traffic participating objects, historical track and motion attribute information of other corresponding traffic participating objects and target objects, and the current map information;

the second determination module is configured to determine hidden random variable multi-mode probability distribution corresponding to each traffic participant by using a feature extraction layer of the target track prediction model and track prediction features corresponding to each traffic participant, wherein the hidden random variables represent behavior randomness of each traffic participant;

and the third determining module is configured to determine a multi-modal predicted track corresponding to each traffic participating object by using the characteristic regression layer of the target track prediction model, the track prediction characteristics corresponding to each traffic participating object and the hidden random variable multi-modal probability distribution corresponding to each traffic participating object.

the first determining module is specifically configured to, for each traffic participant, circularly perform the following steps a-C on the initial feature corresponding to the traffic participant for multiple times by using the feature extraction layer of the target trajectory prediction model, and determine an intermediate prediction feature corresponding to the traffic participant;

for each traffic participating object, fusing the intermediate prediction features corresponding to each static object in the intermediate prediction features corresponding to the traffic participating object based on a graph neural network, and determining the track prediction features corresponding to the traffic participating object, wherein the static objects comprise each static object in the current map information;

and C: and fusing the aggregation characteristics with the characteristics of the to-be-processed characteristics at each historical moment.

Optionally, the second determining module is specifically configured to determine, for each traffic participant, a hidden random variable monomodal probability distribution corresponding to the traffic participant by using a feature extraction layer of a target trajectory prediction model and a trajectory prediction feature corresponding to the traffic participant;

Optionally, the apparatus further comprises:

a training module configured to train to obtain a target trajectory prediction model before determining a trajectory prediction feature corresponding to each traffic participant by using the feature extraction layer of the target trajectory prediction model and the initial feature corresponding to each traffic participant, wherein the training module is specifically configured to train to obtain the target trajectory prediction model

Obtaining an initial trajectory prediction model;

obtaining sample training information corresponding to each sample traffic object and a sample future track corresponding to each sample traffic object, wherein the sample training information corresponding to the sample traffic object comprises: the sample historical track and the sample motion attribute information of the sample traffic object, and the sample historical track, the sample motion attribute information and the sample static object information of the corresponding sample dynamic object;

for each sample traffic object, determining a sample prediction feature corresponding to the sample traffic object by using a feature extraction layer of the initial trajectory prediction model and an initial sample feature corresponding to the sample traffic object, wherein the initial sample feature corresponding to the sample traffic object comprises: sample traffic object sample historical track and sample motion attribute information, and corresponding sample historical track, sample motion attribute information and sample static object information of a sample dynamic object;

aiming at each sample traffic object, determining a multi-modal predicted track corresponding to the sample traffic object by utilizing a characteristic regression layer of the initial track prediction model, sample predicted characteristics corresponding to the sample traffic object and hidden random variable multi-modal probability distribution corresponding to the sample traffic object;

aiming at each sample traffic object, determining a KL divergence value of a hidden random variable corresponding to the sample traffic object by utilizing the multi-modal probability distribution of the hidden random variable corresponding to the sample traffic object and the variation probability distribution of the hidden random variable corresponding to the sample traffic object;

for each sample traffic object, determining a track reconstruction loss value corresponding to the sample traffic object by using a multi-modal predicted track corresponding to the sample traffic object, hidden random variable variation probability distribution and a sample future track corresponding to the sample traffic object;

and if the variation lower bound of the constructed maximum likelihood function reaches the maximum, determining the convergence of the initial trajectory prediction model to obtain the target trajectory prediction model comprising the feature extraction layer and the feature regression layer.

Optionally, the second determining module is specifically configured to sample, for each traffic participant, hidden random variable multi-modal probability distribution corresponding to the traffic participant to obtain multiple hidden random variable samples corresponding to the traffic participant;

As can be seen from the above, the method and the device for predicting a movement track, provided by the embodiment of the present invention, obtain historical tracks and movement attribute information of traffic participation objects corresponding to a target object and corresponding current map information; determining the track prediction characteristics corresponding to the traffic participation objects by utilizing the characteristic extraction layer of the target track prediction model and the initial characteristics corresponding to the traffic participation objects, wherein the initial characteristics corresponding to the traffic participation objects comprise: historical track and motion attribute information of the traffic participation object, historical track and motion attribute information of other corresponding traffic participation objects and target objects and current map information; determining hidden random variable multi-modal probability distribution corresponding to each traffic participating object by utilizing a feature extraction layer of a target track prediction model and track prediction features corresponding to each traffic participating object, wherein the hidden random variables represent behavior randomness of each traffic participating object; and determining the multi-modal predicted track corresponding to each traffic participant by utilizing the characteristic regression layer of the target track prediction model, the track prediction characteristic corresponding to each traffic participant and the hidden random variable multi-modal probability distribution corresponding to each traffic participant.

By applying the embodiment of the invention, the hidden random variables of the behavior randomness of each traffic participating object, the historical track and the motion attribute information of each traffic participating object, the dynamic object information corresponding to the historical track and the motion attribute information of other traffic participating objects and target objects corresponding to the dynamic object information, and the static object information, namely the current map information in the target track prediction model are utilized to fit the hidden random variable multi-modal probability distribution, namely the hidden random variable multi-modal prior distribution of each traffic participating object, which represents various possibilities of the future tracks of each traffic participating object and the target object, so as to determine the multi-modal predicted track corresponding to each traffic participating object, thereby realizing the accurate determination of the multi-modal predicted track of each traffic participating object, and the target track prediction model comprising the hidden random variables of the behavior randomness of each traffic participating object has universality on the scene, the bottleneck restriction on algorithm design does not exist, the modeling capability of the algorithm on the future trajectory distribution can be continuously enhanced along with the enlargement of the training data of the target trajectory prediction model obtained through training, and further the trajectory prediction capability can be continuously improved. Of course, it is not necessary for any product or method to achieve all of the above-described advantages at the same time for practicing the invention.

The innovation points of the embodiment of the invention comprise that:

1. combining the hidden random variables of the behavior randomness of each traffic participating object, the historical track and the motion attribute information of each traffic participating object and the corresponding current map information which are learned in the target track prediction model to construct hidden random variable multi-mode probability distribution corresponding to each traffic participating object, further determining the multi-modal predicted track corresponding to each traffic participant so as to realize the accurate determination of the multi-modal predicted track of each traffic participant, the target track prediction model which comprises the hidden random variables for learning the behavior randomness of each traffic participant has universality on the scene, bottleneck restriction on algorithm design does not exist, the modeling capacity of the algorithm on future track distribution can be continuously strengthened along with the enlargement of the training data of the target track prediction model obtained by training, and further the track prediction capacity can be continuously improved.

2. The method comprises the steps of sequentially carrying out feature processing on to-be-processed features corresponding to traffic participation objects from feature dimensions and time dimensions, realizing aggregation extraction of the features with different feature dimensions and the features with different time dimensions of the to-be-processed features to obtain aggregation features, further fusing the aggregation features with the features of the initial features at various historical moments, repeating the above operations for multiple times to obtain deep abstract intermediate prediction features corresponding to the traffic participation objects, further fusing the intermediate prediction features with various dynamic object information and static object information by using a graph neural network, and determining track prediction features corresponding to the traffic participation objects to provide a basis for ensuring the accuracy of subsequent future track prediction results.

3. And constructing hidden random variable multi-modal probability distribution corresponding to the traffic participation objects through a normalized stream mapping algorithm and the hidden random variable single-modal probability distribution corresponding to the traffic participation objects, and providing a basis for the subsequent prediction of multi-modal trajectories of the traffic participation objects.

4. Training an initial track prediction model through a sample historical track and sample motion attribute information corresponding to each sample traffic object, a sample historical track, sample motion attribute information and static object information of a sample dynamic object corresponding to the sample traffic object, and a sample future track corresponding to each sample traffic object, so that the randomness of the behavior of each sample traffic object is obtained through learning of a hidden random variable in the initial track prediction model, and a basis is provided for accurate prediction of the future track of a subsequent traffic participating object.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below. It is to be understood that the drawings in the following description are of some embodiments of the invention only. For a person skilled in the art, without inventive effort, other figures can also be derived from these figures.

Fig. 1 is a schematic flow chart of a method for predicting a motion trajectory according to an embodiment of the present invention;

fig. 2 is a schematic diagram of mapping hidden random variable monomodal probability distribution into hidden random variable multimodal probability distribution according to an embodiment of the present invention;

FIG. 3 is a schematic flow chart of a training process of a target trajectory prediction model according to an embodiment of the present invention;

fig. 4 is a schematic structural diagram of a device for predicting a motion trajectory according to an embodiment of the present invention.

Detailed Description

The technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. It is to be understood that the described embodiments are merely a few embodiments of the invention, and not all embodiments. All other embodiments, which can be obtained by a person skilled in the art without inventive effort based on the embodiments of the present invention, are within the scope of the present invention.

It should be noted that the terms "comprising" and "having" and any variations thereof in the embodiments and drawings of the present invention are intended to cover non-exclusive inclusions. A process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those steps or elements but may alternatively include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

The invention provides a motion trail prediction method and a motion trail prediction device, which are used for reducing the limitation of motion trail prediction and better adapting to a more complex automatic driving scene. The following provides a detailed description of embodiments of the invention.

Fig. 1 is a schematic flow chart of a method for predicting a motion trajectory according to an embodiment of the present invention. The method may comprise the steps of:

s101: and obtaining historical track and motion attribute information of each traffic participating object corresponding to the target object and corresponding current map information.

The method for predicting the motion trail provided by the embodiment of the invention can be applied to any electronic equipment with computing capacity, and the electronic equipment can be a terminal or a server. In one implementation, the functional software for implementing the motion trajectory prediction method may exist in the form of separate client software, or may exist in the form of a plug-in of currently related client software.

The target object can be an automatic driving vehicle or an intelligent robot. The target object may obtain the historical track and the motion attribute information of each traffic participant object corresponding to the target object through a sensor provided by the target object, where the motion attribute information of the traffic participant object may include the motion information and the attribute information of the traffic participant object, where the motion information of the traffic participant object includes, but is not limited to: speed and acceleration of the traffic participant. The attribute information of the traffic participant object may include, but is not limited to: the type, shape, size, etc. of the traffic participant. The historical track of the traffic participant comprises: and the position information and the posture information of the traffic participating object at each historical moment in a preset time period before the current moment.

The current time may refer to a time at which the electronic device is currently about to predict the trajectory. The above-mentioned historical track and "history" in the historical time are relative to the time when the track is predicted by the electronic device at present, that is, the current time, and refer to the track generated in a time period before the current time.

The storage device local to the electronic device or connected to the electronic device may store the complete map information of the area where the target object is located in advance, the current map information may be the complete map information, or the map information in the area range corresponding to the current pose information of the target object may be determined from the complete map information based on the current pose information of the target object at the current time.

The sensors provided by the target object may include, but are not limited to: image acquisition equipment, wheel speed sensors, radar, IMU (Inertial measurement unit), GPS (Global Positioning System), GNSS (Global Navigation Satellite System), and the like.

In one case, if the target object is an autonomous vehicle, the corresponding traffic participant object may include, but is not limited to: motor vehicles, bicycles, tricycles, pedestrians, and animals. In the case where the traffic participant is a motor vehicle, the electronic device may also obtain information on a lamp of the motor vehicle, for example: and turning on and off the steering lamp.

In one case, the target object is an autonomous vehicle, and the current map information may include, but is not limited to: traffic identification information such as lane line, parking stall, pedestrian way, traffic sign, traffic instruction arrow, light pole, wherein, can call each traffic identification information that includes in the current map information as static object, can also include: stationary buildings, vegetation, and other objects in the scene.

S102: and determining the track prediction characteristics corresponding to the traffic participation objects by utilizing the characteristic extraction layer of the target track prediction model and the initial characteristics corresponding to the traffic participation objects.

The initial characteristics corresponding to the traffic participation objects comprise: historical track and motion attribute information of the traffic participating objects, historical track and motion attribute information of other corresponding traffic participating objects and target objects, and current map information.

The target track prediction model is as follows: and training the obtained model based on the sample historical track and the sample motion attribute information corresponding to the sample traffic object, the sample historical track and the sample motion attribute information of the corresponding sample dynamic object, the static object information and the sample future track corresponding to each sample traffic object. The target track prediction model is a neural network hidden variable model. For clarity of layout, the following description will be made with respect to the training process of the target trajectory prediction model.

The sample dynamic object corresponding to the sample traffic object may include: and other dynamic traffic objects in the scene of the sample traffic object. The static object information corresponding to the sample traffic object may include: and each static object in the map information corresponding to the scene where the sample traffic object is located.

The electronic device may input the initial features corresponding to the traffic participant into a feature extraction layer of the target trajectory prediction model for each traffic participant, so as to perform feature extraction on the initial features corresponding to the traffic participant through the feature extraction layer of the target trajectory prediction model, and determine the trajectory prediction features corresponding to the traffic participant, so as to determine the trajectory prediction features corresponding to each traffic participant.

S103: and determining hidden random variable multi-modal probability distribution corresponding to each traffic participating object by utilizing the feature extraction layer of the target track prediction model and the track prediction features corresponding to each traffic participating object.

The hidden random variable represents the behavior randomness of each traffic participant.

In this step, the target trajectory prediction model is a model provided with a hidden random variable, and after the electronic device obtains the trajectory prediction features corresponding to the traffic participating objects, for each traffic participating object, the hidden random variable multi-modal probability distribution corresponding to the traffic participating object is determined by using the trajectory prediction features corresponding to the traffic participating object and the hidden random variable of the feature extraction layer of the target trajectory prediction model. The randomness and the uncertainty of the track of the traffic participant are shown through the hidden random variable table.

In one implementation manner of the present invention, the step S103 may include the following steps 011-:

011: and aiming at each traffic participant, determining the monomodal probability distribution of the hidden random variables corresponding to the traffic participant by using the feature extraction layer of the target track prediction model and the track prediction features corresponding to the traffic participant.

012: aiming at each traffic participant, a multi-modal trajectory distribution corresponding to the hidden random variable corresponding to the traffic participant is obtained by utilizing a normalized stream mapping algorithm and the hidden random variable single-modal probability distribution corresponding to the traffic participant.

In the implementation mode, the electronic equipment firstly determines the monomodal probability distribution of the hidden random variables corresponding to the traffic participant by using a feature extraction layer of a target track prediction model and track prediction features corresponding to the traffic participant; and mapping the hidden random variable single-mode probability distribution corresponding to the traffic participating object into multi-mode probability distribution through a normalized Flow (normalized Flow) mapping algorithm to obtain the hidden random variable multi-mode probability distribution corresponding to the traffic participating object.

In one case, assuming that the unimodal probability distribution of the hidden random variable corresponding to the traffic participant is a multivariate gaussian distribution, the feature extraction layer of the target trajectory prediction model may output a mean value and a variance corresponding thereto, and the unimodal probability distribution of the hidden random variable corresponding to the traffic participant may be constructed through the mean value and the variance corresponding thereto. Subsequently, the hidden random variable single-mode probability distribution corresponding to the traffic participant is mapped into multi-mode track distribution through a normalized stream mapping algorithm to obtain the hidden random variable multi-mode probability distribution corresponding to the traffic participant, so that the difficulty of mapping the randomness of the track, namely the randomness of the hidden random variable, into track space multi-mode probability distribution by a subsequent target track prediction model is simplified, and a better multi-mode future track modeling effect is achieved. The effect diagram is shown in fig. 2, where "monomodal distribution" represents a monomodal probability distribution of the hidden random variables corresponding to the traffic participant, and "multimodal distribution" represents a multimodal trajectory distribution of the hidden random variables corresponding to the traffic participant.

S104: and determining the multi-modal predicted track corresponding to each traffic participating object by utilizing the characteristic regression layer of the target track prediction model, the track prediction characteristics corresponding to each traffic participating object and the hidden random variable multi-modal probability distribution corresponding to each traffic participating object.

The electronic equipment determines hidden random variable multi-modal probability distribution corresponding to each traffic participating object, and for each traffic participating object, the characteristic regression layer of the target track prediction model is utilized to fuse the track prediction characteristics corresponding to the traffic participating object and the hidden random variable multi-modal probability distribution corresponding to the traffic participating object so as to determine the multi-modal prediction track corresponding to the traffic participating object.

In an implementation manner of the present invention, the S104 may include the following steps 021-:

021: and sampling the hidden random variable multi-mode probability distribution corresponding to the traffic participating object aiming at each traffic participating object to obtain a plurality of hidden random variable samples corresponding to the traffic participating object.

022: and determining a multi-modal predicted track corresponding to each traffic participant by using the characteristic regression layer of the target track prediction model, the track prediction characteristics corresponding to each traffic participant and a plurality of hidden random variable samples corresponding to each traffic participant.

In this implementation manner, the electronic device samples, for each traffic participant, multi-modal probability distribution of hidden random variables corresponding to the traffic participant to obtain multiple hidden random variable samples corresponding to the traffic participant, and further maps, by using a feature regression layer of a target trajectory prediction model, a trajectory prediction feature corresponding to the traffic participant and multiple hidden random variable samples corresponding to the traffic participant to a trajectory space, that is, fuses the trajectory prediction feature corresponding to the traffic participant and each hidden random variable sample corresponding to the traffic participant to obtain a multi-modal prediction trajectory corresponding to each traffic participant.

By applying the embodiment of the invention, the conditional probability distribution of the future track of the participating objects, namely the multi-modal probability distribution of the hidden random variables corresponding to each traffic participating object, can be fitted by utilizing the hidden random variables of the behavior randomness of each traffic participating object, the historical track and the motion attribute information of each traffic participating object, the dynamic object information corresponding to each traffic participating object, namely the historical track and the motion attribute information of other traffic participating objects and the static object information corresponding to the dynamic object information, namely the current map information, which are learned in the target track prediction model, so as to realize the accurate determination of the multi-modal prediction track of each traffic participating object, and the target track prediction model comprising the hidden random variables of the behavior randomness of each traffic participating object has universality on the scene, the bottleneck restriction on algorithm design does not exist, the modeling capability of the algorithm on the future trajectory distribution can be continuously enhanced along with the enlargement of the training data of the target trajectory prediction model obtained through training, and further the trajectory prediction capability can be continuously improved.

In another embodiment of the invention, the initial characteristics corresponding to the traffic participation objects are characteristics arranged based on time sequence, and the characteristics comprise characteristics of a plurality of historical moments corresponding to the traffic participation objects;

the step S102 may include the following steps 031-:

031: and aiming at each traffic participant, circularly executing the following steps A-C for a plurality of times on the initial characteristics corresponding to the traffic participant by using the characteristic extraction layer of the target track prediction model, and determining the intermediate prediction characteristics corresponding to the traffic participant.

032: and for each traffic participating object, fusing the intermediate prediction features corresponding to the static objects in the intermediate prediction features corresponding to the traffic participating object based on the graph neural network, and determining the track prediction features corresponding to the traffic participating object.

The static objects comprise all static objects in the current map information.

Step A: and carrying out nonlinear mapping on the to-be-processed features corresponding to the traffic participation objects from the feature dimensions to obtain the mapping features corresponding to the traffic participation objects.

The feature to be processed is an initial feature corresponding to the traffic participation object or an intermediate prediction feature generated in the previous iteration and corresponding to the traffic participation object.

And B, step B: and carrying out feature aggregation operation on the mapping features from the time dimension to obtain aggregation features corresponding to the traffic participation objects.

And C: and fusing the aggregation characteristic with the characteristic of each historical moment in the characteristic to be processed.

In this implementation, the initial features corresponding to the traffic participation objects are features arranged based on a time sequence, and include features of a plurality of historical moments corresponding to the traffic participation objects. It is understood that each historical time may correspond to a plurality of characteristics corresponding to the traffic participant, such as: the position information, the posture information, such as the orientation angle, the speed, the shape, the type and the size of the traffic participating object, the position information, the posture information, such as the orientation angle, the speed, the shape, the type and the size of other traffic participating objects and target objects corresponding to the traffic participating object, the relative position information and the type of each static object in the current map information corresponding to the traffic participating object, and the like. The corresponding characteristics of the historical time are arranged according to the sequence of the information of the historical time.

The electronic equipment can firstly perform nonlinear mapping on the initial features corresponding to the traffic participation objects from feature dimensions by utilizing a feature extraction layer of a target track prediction model aiming at each traffic participation object, namely perform nonlinear mapping on the features corresponding to the historical moments from the feature dimensions aiming at each historical moment to obtain the mapping features corresponding to the traffic participation objects; performing feature aggregation operation on the mapping features from the time dimension, namely performing feature aggregation operation on the mapping features corresponding to each historical moment to obtain aggregation features corresponding to the traffic participation objects; fusing the aggregation feature corresponding to the traffic participation object with the feature of each historical moment in the initial features, taking the fused feature as a new feature to be processed corresponding to the traffic participation object, and re-executing the step of carrying out nonlinear mapping on the initial feature corresponding to the traffic participation object from the feature dimension by using the feature extraction layer of the target track prediction model aiming at each traffic participation object to obtain the mapping feature corresponding to the traffic participation object until the step is repeatedly executed for multiple times to obtain the intermediate prediction feature corresponding to the traffic participation object and containing the deep abstract feature.

Furthermore, for each traffic participant, the electronic device fuses the intermediate prediction features corresponding to the static objects in the intermediate prediction features corresponding to the traffic participant based on the graph neural network, and determines the track prediction feature corresponding to the traffic participant.

In another embodiment of the present invention, before the S102, the method may further include:

a process of training to obtain a target trajectory prediction model, wherein, as shown in fig. 3, the process includes:

s301: an initial trajectory prediction model is obtained.

S302: and obtaining sample training information corresponding to each sample traffic object and a sample future track corresponding to each sample traffic object.

The sample training information corresponding to the sample traffic object comprises: the sample traffic object sample historical track and sample motion attribute information, and the corresponding sample dynamic object sample historical track, sample motion attribute information and sample static object information. The sample motion attribute information may include motion information and attribute information of sample traffic objects, wherein the motion information of sample traffic objects includes, but is not limited to: and the speed, the acceleration and other information of the sample traffic object. The attribute information of the sample traffic object may include, but is not limited to: the type, shape, and size of the sample traffic object, etc. The sample historical trajectories of sample traffic objects include: and the position information and the posture information of the sample traffic object at each historical moment in a preset time period before the sample track acquisition moment. The sample future trajectory of the sample traffic object is: the real moving track of the sample traffic object in the sample track acquisition time and the first time period after the sample track acquisition time comprises real position information and attitude information of the sample traffic object at each time in the sample track acquisition time and the first time period after the sample track acquisition time.

The sample dynamic object corresponding to the sample traffic object is a dynamic object around the sample traffic object in the environment, and can include vehicles, pedestrians, animals and the like; the sample static object information corresponding to the sample traffic object includes each static object in the map information of the environment in which the sample traffic object is located.

In one case, the target object is an autonomous vehicle, accordingly, the sample vehicle may collect corresponding information for each object in its environment during driving, and sample training information and sample future trajectories corresponding to a certain sample traffic object may be determined based on sensor data collected by the sample vehicle through a sensor provided in the sample vehicle.

In one case, in the case that the sample traffic object is a motor vehicle, the sample training information corresponding to the sample traffic object may further include car light information of the motor vehicle, for example: and turning on and off the steering lamp.

S303: and for each sample traffic object, determining a sample prediction characteristic corresponding to the sample traffic object by using the characteristic extraction layer of the initial track prediction model and the initial sample characteristic corresponding to the sample traffic object.

The initial sample characteristics corresponding to the sample traffic object comprise: and the sample traffic object sample historical track and the sample motion attribute information correspond to the sample traffic object sample historical track, the sample motion attribute information and the sample static object information of the sample dynamic object.

S304: and aiming at each sample traffic object, determining hidden random variable multi-modal probability distribution corresponding to the sample traffic object by utilizing a feature extraction layer of the initial track prediction model and sample prediction features corresponding to the sample traffic object.

S305: and for each sample traffic object, determining a multi-modal predicted track corresponding to the sample traffic object by utilizing the characteristic regression layer of the initial track prediction model, the sample predicted characteristic corresponding to the sample traffic object and the hidden random variable multi-modal probability distribution corresponding to the sample traffic object.

S306: and aiming at each sample traffic object, processing the sample future track corresponding to the sample traffic object by using a preset variation algorithm to obtain the hidden random variable variation probability distribution corresponding to the sample traffic object.

S307: and aiming at each sample traffic object, determining a KL divergence value of the hidden random variable corresponding to the sample traffic object by using the multimode probability distribution of the hidden random variable corresponding to the sample traffic object and the variation probability distribution of the hidden random variable corresponding to the sample traffic object.

S308: and for each sample traffic object, determining a track reconstruction loss value corresponding to the sample traffic object by using the multi-modal predicted track corresponding to the sample traffic object and the sample future track corresponding to the sample traffic object.

S309: aiming at each sample traffic object, constructing a variation lower bound of a maximum likelihood function by utilizing a KL divergence value of an implicit random variable corresponding to the sample traffic object and a track reconstruction loss value corresponding to the sample traffic object; and judging whether the variation lower bound of the constructed maximum likelihood function reaches the maximum.

S310: if the variation lower bound of the constructed maximum likelihood function does not reach the maximum, adjusting the model parameters of the feature extraction layer and the feature regression layer of the initial trajectory prediction model, and returning to the step S303.

S311: and if the variation lower bound of the constructed maximum likelihood function reaches the maximum, determining the convergence of the initial trajectory prediction model to obtain a target trajectory prediction model comprising a feature extraction layer and a feature regression layer.

In order to ensure the accuracy of the future track prediction of each traffic participant, the embodiment of the invention also comprises a training process of a target track prediction model. Correspondingly, the electronic device may first obtain an initial trajectory prediction model, where the initial trajectory prediction model may be a neural network hidden variable model; and obtaining sample training information corresponding to each sample traffic object and a sample future track corresponding to each sample traffic object.

Furthermore, for each sample traffic object, the electronic device inputs the initial sample features corresponding to the sample traffic object into the feature extraction layer of the initial trajectory prediction model, performs feature extraction and fusion on the initial sample features corresponding to the sample traffic object by using the feature extraction layer of the initial trajectory prediction model, and determines the sample prediction features corresponding to the sample traffic object, where the process of performing feature extraction and fusion on the initial sample features corresponding to the sample traffic object by using the feature extraction layer of the initial trajectory prediction model can be referred to the process of performing feature extraction and fusion on the initial features of the traffic participant by using the feature extraction layer of the target trajectory prediction model, and details are not repeated here.

And aiming at each sample traffic object, obtaining hidden random variable single-mode probability distribution corresponding to the sample traffic object by utilizing a feature extraction layer of the initial track prediction model and sample prediction features corresponding to the sample traffic object, and mapping the hidden random variable single-mode probability distribution corresponding to the sample traffic object into hidden random variable multi-mode probability distribution corresponding to the sample traffic object through a normalized stream mapping algorithm.

And for each sample traffic object, inputting the sample prediction characteristics corresponding to the sample traffic object and the hidden random variable multi-modal probability distribution corresponding to the sample traffic object into a characteristic regression layer of the initial trajectory prediction model, and fusing the sample prediction characteristics corresponding to the sample traffic object and the hidden random variable multi-modal probability distribution corresponding to the sample traffic object through the characteristic regression layer of the initial trajectory prediction model to obtain the multi-modal prediction trajectory corresponding to the sample traffic object.

Subsequently, in order to ensure the accuracy of the prediction result of the constructed target track prediction model, a variation lower bound of the maximum likelihood function can be constructed by using the multi-modal prediction track corresponding to the sample traffic object and the sample future track corresponding to the sample traffic object, and the model parameters of the initial track prediction model are adjusted through the constructed variation lower bound of the maximum likelihood function, so as to obtain the final target track prediction model. In order to construct a variation lower bound of a maximum likelihood function, the electronic device processes a sample future track corresponding to each sample traffic object by using a preset variation algorithm aiming at each sample traffic object to obtain a hidden random variable variation probability distribution corresponding to the sample traffic object, wherein the preset variation algorithm can be a variation algorithm constructed on the basis of a variation Bayesian principle.

And determining a KL divergence value of the hidden random variable corresponding to the sample traffic object by utilizing the multimode probability distribution of the hidden random variable corresponding to the sample traffic object and the variational probability distribution of the hidden random variable corresponding to the sample traffic object based on a KL divergence algorithm. And determining a track reconstruction loss value corresponding to the sample traffic object by using the multi-modal predicted track corresponding to the sample traffic object and the sample future track corresponding to the sample traffic object. Constructing a variation lower bound of a maximized likelihood function by utilizing a track reconstruction loss value corresponding to the sample traffic object and a KL divergence value of a hidden random variable corresponding to the sample traffic object; calculating a function value corresponding to the variation lower bound of the maximum likelihood function, judging whether the variation lower bound of the constructed maximum likelihood function reaches the maximum, namely whether the function value corresponding to the variation lower bound of the maximum likelihood function reaches the maximum, if the variation lower bound of the constructed maximum likelihood function does not reach the maximum, adjusting model parameters of a feature extraction layer and a feature regression layer of the initial trajectory prediction model by using a preset optimization algorithm, and returning to execute S203; and if the variation lower bound of the constructed maximum likelihood function reaches the maximum, determining that the initial trajectory prediction model is converged to obtain a target trajectory prediction model comprising a feature extraction layer and a feature regression layer.

In one case, a probability distribution corresponding to the obtained multi-modal predicted trajectory corresponding to the sample traffic object is constructed, and the multi-modal predicted trajectory corresponding to the sample traffic object can be constructed by using the probability distribution, which can be represented by the following formula (1):

p(x _f |x _p ,Φ)＝∫p(x _f |z,x _p ,Φ)p(z|x _p ,Φ)dz； (1)

wherein x is _p Representing the sample historical track, x, corresponding to the sample traffic object _f Representing a multi-modal predicted trajectory corresponding to the sample traffic object; phi represents other information except the sample historical track corresponding to the sample traffic object in the initial sample characteristic corresponding to the sample traffic object; p (x) _f |x _p Phi) represents the probability distribution corresponding to the multi-modal predicted trajectory corresponding to the sample traffic object; z represents a hidden random variable;p(z|x _p phi) represents the multi-modal probability distribution of the hidden random variables corresponding to the sample traffic object, and represents the randomness of the future track of the sample traffic object according to the historical track of the sample traffic object and the surrounding map, namely the static object information of the sample and the overall consideration of the dynamic object of the sample, for the prior distribution of the hidden random variables z under the condition of giving the historical track of the sample and other information in the initial sample characteristic except the historical track of the sample corresponding to the sample traffic object; p (x) _f |z,x _p Phi), namely the probability distribution of the future track under the condition of giving additional information such as hidden random variables, sample historical tracks and maps, namely, the prediction result of the future track is output by comprehensively considering all deterministic and stochastic information. The modeling mode can represent the behavior randomness of a sample traffic object or a traffic participant through a hidden random variable z, and the randomness is mapped to an original track data space by using a neural network model, namely an initial track prediction model or a target track prediction model, so that any future track distribution can be fitted theoretically, and the modeling mode has high universality and effect.

Accordingly, the lower variation bound of the constructed maximum likelihood function can be expressed by the following formula;

logp(x _f |x _p ,Φ)≥E _q [(x _f |z,x _p ,Φ)]-KL(q(z|x _f ,x _p ,Φ)||p(z|x _p ,Φ))；

wherein logp (x) _f |x _p Φ) represents the maximum likelihood function constructed, E _q [(x _f |z,x _p ,Φ)]Represents the track reconstruction loss value corresponding to the sample traffic object, KL (q (z | x) _f ,x _p ,Φ)||p(z|x _p Phi) represents the KL divergence value of the hidden random variable corresponding to the sample traffic object, E _q [(x _f |z,x _p ,Φ)]-KL(q(z|x _f ,x _p ,Φ)||p(z|x _p Φ)), represents the lower bound of the variation of the maximum likelihood function.

In the implementation mode, in the process of constructing the target track prediction model, the historical tracks and the running attribute information of all traffic participating objects and the information of surrounding static objects are fully considered, the information presents high-dimensional multi-source heterogeneous characteristics, and the characteristics of all directions, namely characteristic dimensions and time dimensions, among all the characteristics are extracted and fused through the characteristic extraction layer of the initial track prediction model, so that all the characteristics corresponding to the traffic participating objects are fully extracted and fused, and the future tracks are predicted by a subsequent model.

Corresponding to the above method embodiment, an embodiment of the present invention provides an apparatus for predicting a motion trajectory, as shown in fig. 4, the apparatus may include:

an obtaining module 410 configured to obtain historical trajectory and motion attribute information of each traffic participation object corresponding to the target object and corresponding current map information;

a first determining module 420 configured to determine a track prediction feature corresponding to each traffic participant by using the feature extraction layer of the target track prediction model and the initial feature corresponding to each traffic participant, where the initial feature corresponding to the traffic participant includes: historical track and motion attribute information of the traffic participation object, historical track and motion attribute information of other corresponding traffic participation objects and target objects, and the current map information;

the second determining module 430 is configured to determine a multi-modal probability distribution of a hidden random variable corresponding to each traffic participating object by using the feature extraction layer of the target trajectory prediction model and the trajectory prediction features corresponding to each traffic participating object, where the hidden random variable represents behavior randomness of each traffic participating object;

a third determining module 440, configured to determine a multi-modal predicted trajectory corresponding to each traffic participant by using the feature regression layer of the target trajectory prediction model, the trajectory prediction feature corresponding to each traffic participant, and the hidden random variable multi-modal probability distribution corresponding to each traffic participant.

By applying the embodiment of the invention, the conditional probability distribution of the future track of the participating objects, namely the multi-modal probability distribution of the hidden random variables corresponding to each traffic participating object, can be fitted by utilizing the hidden random variables of the behavior randomness of each traffic participating object, the historical track and the motion attribute information of each traffic participating object, the dynamic object information corresponding to each traffic participating object, namely the historical track and the motion attribute information of other traffic participating objects and the static object information corresponding to the dynamic object information, namely the current map information, which are learned in the target track prediction model, so as to realize the accurate determination of the multi-modal prediction track of each traffic participating object, and the target track prediction model comprising the hidden random variables of the behavior randomness of each traffic participating object has universality on the scene, the bottleneck restriction on algorithm design does not exist, the modeling capability of the algorithm on future track distribution can be continuously enhanced along with the enlargement of the training data of the target track prediction model obtained by training, and further the track prediction capability can be continuously improved.

the first determining module 420 is specifically configured to, for each traffic participant, perform the following steps a to C on the initial feature cycle corresponding to the traffic participant by using the feature extraction layer of the target trajectory prediction model for multiple times, and determine an intermediate prediction feature corresponding to the traffic participant;

step A: carrying out nonlinear mapping on the to-be-processed features corresponding to the traffic participating objects from feature dimensions to obtain mapping features corresponding to the traffic participating objects, wherein the to-be-processed features are initial features corresponding to the traffic participating objects or intermediate prediction features corresponding to the traffic participating objects generated in previous iteration;

and B: performing feature aggregation operation on the mapping features from a time dimension to obtain aggregation features corresponding to the traffic participation objects;

In another embodiment of the present invention, the second determining module 430 is specifically configured to, for each traffic participant, determine a hidden random variable monomodal probability distribution corresponding to the traffic participant by using the feature extraction layer of the target trajectory prediction model and the trajectory prediction feature corresponding to the traffic participant;

In another embodiment of the present invention, the apparatus further comprises:

a training module (not shown in the figure) configured to train to obtain a target trajectory prediction model before determining a trajectory prediction feature corresponding to each traffic participant by using the feature extraction layer of the target trajectory prediction model and the initial feature corresponding to each traffic participant, wherein the training module is specifically configured to train to obtain the target trajectory prediction model

Obtaining an initial trajectory prediction model;

for each sample traffic object, carrying out variation processing on a sample future track corresponding to the sample traffic object by using a preset variation algorithm to obtain a variation distribution probability corresponding to the sample traffic object;

if the variation lower bound of the constructed maximum likelihood function does not reach the maximum, adjusting model parameters of a feature extraction layer and a feature regression layer of the initial trajectory prediction model, returning to the step of determining sample prediction features corresponding to the sample traffic objects by using the feature extraction layer of the initial trajectory prediction model and the initial sample features corresponding to the sample traffic objects;

In another embodiment of the present invention, the second determining module 430 is specifically configured to, for each traffic participating object, sample a hidden random variable multi-modal probability distribution corresponding to the traffic participating object, to obtain a plurality of hidden random variable samples corresponding to the traffic participating object;

The system and apparatus embodiments correspond to the system embodiments, and have the same technical effects as the method embodiments, and for the specific description, refer to the method embodiments. The device embodiment is obtained based on the method embodiment, and for specific description, reference may be made to the method embodiment section, which is not described herein again. Those of ordinary skill in the art will understand that: the figures are schematic representations of one embodiment, and the blocks or processes shown in the figures are not necessarily required to practice the present invention.

Those of ordinary skill in the art will understand that: modules in the devices in the embodiments may be distributed in the devices in the embodiments according to the description of the embodiments, or may be located in one or more devices different from the embodiments with corresponding changes. The modules of the above embodiments may be combined into one module, or further split into multiple sub-modules.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, and not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting a motion trajectory, the method comprising:

obtaining historical track and motion attribute information of each traffic participating object corresponding to the target object and corresponding current map information;

determining hidden random variable multi-mode probability distribution corresponding to each traffic participant by using a feature extraction layer of a target track prediction model and track prediction features corresponding to each traffic participant, wherein the hidden random variables represent behavior randomness of each traffic participant;

and determining the multi-modal predicted track corresponding to each traffic participating object by utilizing the characteristic regression layer of the target track prediction model, the track prediction characteristics corresponding to each traffic participating object and the hidden random variable multi-modal probability distribution corresponding to each traffic participating object.

2. The method of claim 1, wherein the initial features corresponding to the traffic participant are features arranged based on time sequence, and the initial features comprise features of a plurality of historical moments corresponding to the traffic participant;

the step of determining the track prediction characteristics corresponding to the traffic participation objects by using the characteristic extraction layer of the target track prediction model and the initial characteristics corresponding to the traffic participation objects comprises the following steps:

3. The method according to claim 1, wherein the step of determining the hidden random variable multi-modal probability distribution corresponding to each traffic participant by using the feature extraction layer of the target trajectory prediction model and the trajectory prediction features corresponding to each traffic participant comprises:

aiming at each traffic participant, determining hidden random variable monomodal probability distribution corresponding to the traffic participant by using a feature extraction layer of a target track prediction model and track prediction features corresponding to the traffic participant;

4. The method according to any one of claims 1 to 3, wherein before the step of determining the trajectory prediction feature corresponding to each traffic participant by using the feature extraction layer of the target trajectory prediction model and the initial feature corresponding to each traffic participant, the method further comprises:

obtaining an initial trajectory prediction model;

for each sample traffic object, determining a track reconstruction loss value corresponding to the sample traffic object by utilizing a multi-modal predicted track corresponding to the sample traffic object and a sample future track corresponding to the sample traffic object;

aiming at each sample traffic object, constructing a variation lower bound of a maximum likelihood function by utilizing a KL divergence value of an implicit random variable corresponding to the sample traffic object and a track reconstruction loss value corresponding to the sample traffic object; judging whether the variation lower bound of the constructed maximum likelihood function reaches the maximum;

5. The method according to any one of claims 1 to 4, wherein the step of determining the multi-modal predicted trajectory corresponding to each traffic participant by using the feature regression layer of the target trajectory prediction model, the trajectory prediction feature corresponding to each traffic participant and the hidden random variable multi-modal probability distribution corresponding to each traffic participant comprises:

6. An apparatus for predicting a motion trajectory, the apparatus comprising:

the second determining module is configured to determine multi-modal probability distribution of hidden random variables corresponding to the traffic participating objects by using the feature extraction layer of the target track prediction model and the track prediction features corresponding to the traffic participating objects, wherein the hidden random variables represent behavior randomness of the traffic participating objects;

7. The apparatus of claim 6, wherein the initial feature corresponding to the traffic participant object is a feature arranged based on a time sequence, and comprises a plurality of historical time features corresponding to the traffic participant object;

and B, step B: performing feature aggregation operation on the mapping features from the time dimension to obtain aggregation features corresponding to the traffic participation objects;

8. The apparatus of claim 6, wherein the second determining module is specifically configured to determine, for each traffic participant, a hidden random variable monomodal probability distribution corresponding to the traffic participant by using a feature extraction layer of a target trajectory prediction model and a trajectory prediction feature corresponding to the traffic participant;

9. The apparatus of claim 6, wherein the apparatus further comprises:

Obtaining an initial trajectory prediction model;

10. The apparatus according to any one of claims 6 to 9, wherein the second determining module is specifically configured to, for each traffic participant, sample a hidden random variable multi-modal probability distribution corresponding to the traffic participant to obtain a plurality of hidden random variable samples corresponding to the traffic participant;