Computer-implemented method for training a model, as well as method and system for predicting the behavior of a road user.

Abstract

Measures are proposed that enable the use of a conventionally trained CVAE to generate output information with a multi-modal distribution, particularly in the context of predicting the behavior of participants in a traffic scene and also for generating digital images as training data for a machine learning (ML) component. For this purpose, a label sample Z is used for at least some of the training data elements. ϕ (13), a context sample Z γ (14) and a composite sample [Z ϕ ; Z γ ] (16) generated, where the composite sample [Z ϕ ; Z γ ] (16) as a combination of the label sample Z ϕ (13) and the context sample Z γ (15) is generated. Based on a set of composite samples [Z ϕ ; Z γ ] (16) then a composite probability distribution (18) is determined in the latent composite space.

Description

State of the art

[0001] The invention is based on the concept of a Conditional Variational Auto-Encoder, as described, for example, in Ghosh et al., “From variational to deterministic autoencoder”, https: / / arxiv.org / abs / 1903.12436.

[0002] Accordingly, the invention relates to a computer-implemented method for training a model comprising a Conditional Variational Auto Encoder (CVAE) structure with a label encoder ϕ, a prior encoder y, and a decoder θ. A set of training data elements is provided for the training method, each comprising at least one piece of ground truth information Y* with associated training context information X*. The physical meaning of the ground truth information Y* and the associated context information X* depends on the respective task or application for which the model is being trained. An important application for such models is the prediction of the behavior of participants in a traffic scene as part of the behavior planning of an automated vehicle.

[0003] To train the model, the following steps are performed for each training data item: - Use of the label encoder ϕ to transform the training data element into characteristics of a multivariate label probability distribution of latent features, - Use of the prior encoder y to transform the training context information X* into parameters of a multivariate context probability distribution of latent features, - Determining a first distance between the label probability distribution and the context probability distribution and comparing the first distance with a predetermined first optimization criterion, - Generate at least one label sample Z ϕ based on the parameters of the multivariate label probability distribution, - Use of the decoder θ to generate at least one output information Y based on the label sample Z ϕ , - Determining a second distance between the ground truth information Y* and the generated output information Y and comparing the second distance with a predetermined second optimization criterion, - Modifying at least one parameter of the label encoder ϕ, the prior encoder y and / or the decoder θ if the first and / or the second optimization criterion is not met.

[0004] In order to plan safe and predictable maneuvers, automated vehicles must anticipate how the current traffic scene will develop. This is particularly important when other participants, such as other vehicles, cyclists, and pedestrians, are also present in the traffic scene. Therefore, attempts are made to predict the future behavior of as many participants in the traffic scene as possible, with the possible behavior of a participant usually being described in the form of trajectories. This prediction is then taken into account when planning the behavior of the automated vehicle. Classical prediction methods generally use simple kinematic models for the individual participants in the traffic scene and can only model interactions between participants to a limited extent.For this reason, the use of machine learning, especially deep learning (DL), has become the de facto standard for prediction in recent years.

[0005] In practice, a Conditional Variational Auto-Encoder (CVAE) model is often used to predict trajectories, since such models capture the relationship between the context of the traffic scene and the ground truth future in a probabilistic latent space and take it into account in the prediction. The relationship between the context of the traffic scene and the ground truth future is essentially learned through two measures of the training procedure. One of these measures consists in using training data elements as input for the label encoder ϕ, which include both ground truth information Y* and the corresponding training context information X*. The label encoder ϕ then maps these training data elements to parameters of a multivariate label probability distribution. And the decoder θ label samples Z ϕThis label probability distribution is used to generate output information Y. The other measure is that the multivariate context probability distribution is learned during training in such a way that it is as similar as possible to the label probability distribution. The relationship between the training context information X* and the ground truth information Y* is therefore learned primarily by the label encoder ϕ using both the ground truth information Y* and the corresponding training context information X*, thus modeling a joint probability from which label samples Z ϕ for the decoder θ. Since the context probability distribution is trained to be as similar as possible to the label probability distribution, the prior encoder y can later be used to infer output information Y from context information X*.

[0006] The idea behind a CVAE is to condition on a context and capture "nuances" in a distribution of latent features. To do this, the label probability distribution and the context probability distribution are usually modeled as a multivariate normal distribution, particularly a standard normal distribution. The latent representation in this case is therefore continuous and unimodal.

[0007] However, a uni-modal latent representation does not do justice to some use cases, especially use cases with a highly multi-modal distribution of the output information Y.

[0008] One of these use cases is behavior prediction. The future behavior of a participant in a traffic scene is always largely determined by the available route options and lanes. Thus, a participant's future behavior depends significantly on whether they decide to turn right or left at an intersection, or to continue straight ahead. These three options correspond to three completely different, but realistic, behavioral modes. However, if the samples for the prediction are generated from a continuous unimodal probability distribution in latent space, the prediction inevitably also yields predictions that lie between the different behavioral modes and thus represent nonsensical, rather than realistic, behaviors, such as crossing the intersection at an angle.This can have serious implications for downstream behavior planning for an automated vehicle. Disclosure of the invention

[0009] Measures are proposed that enable the use of a conventionally trained CVAE, as described above, to generate output information with a multimodal distribution. These measures specifically enable the use of a CVAE in the context of predicting the behavior of participants in a traffic scene and also the use of a CVAE to generate digital images as training data for a machine learning (ML) component.

[0010] For this purpose, the invention proposes that upon completion of the training process, i.e. when the first distance between the label probability distribution of the label samples Z ϕ and the context probability distribution of the context samples Z γthe first optimization criterion is met and / or if the second distance between the ground truth information Y* and the generated output information Y meets the second optimization criterion, for at least some of the training data elements, a label sample Z ϕ , a context sample Z γ and a composite sample [Z ϕ ; Z γ ] The composite sample [Z ϕ ; Z γ ] is defined according to the invention as a combination of the label sample Z ϕ and the context sample Z γ According to the invention, on the basis of a set of composite samples [Z ϕ ; Z γ ] a joint probability distribution in a latent joint space is determined.

[0011] According to the invention, it has been recognized that, for corresponding sets of training data elements, a multi-modality of the output information is captured with the aid of a joint probability distribution determined in this way.

[0012] In principle, there are various possibilities for the realization of the training method according to the invention, both with regard to the training of the label encoder ϕ, the prior encoder y and the decoder θ, as well as with regard to the determination of the joint probability distribution.

[0013] In a preferred variant of the invention, the training data element is divided into parameters of a first multivariate normal distribution N(µ ϕ , Σ ϕ ) of latent features, just as the training context information X* is transformed into parameters of a second multivariate normal distribution N(µ γ , Σ γ ) of latent characteristics. These parameters are preferably parameters from which a first expected value vector µ ϕ or a second expectation vector µ γ and a first covariance matrix Σ ϕ or a second covariance matrix Σγ For example, the square root of the respective vector and matrix values ​​can also be used as parameters.

[0014] The label samples Z ϕ and / or the context samples Z γ can be generated randomly and / or deterministically, for example using the unscented transformation method.

[0015] It is particularly advantageous if the decoder θ uses not only the generated label sample Z ϕ but also uses the training context information X* and / or the generated context sample Z γ .

[0016] In a preferred embodiment of the invention, the joint probability distribution is determined by fitting a Gaussian mixture distribution to the joint samples [Z ϕ ; Z γ] in the composite space. Fitting can be performed parametrically by specifying at least the number of Gaussian distribution components and / or their properties. This variant proves particularly advantageous for applications where the number and type of expected modes are known in advance. However, fitting can also be performed non-parametrically, for example, using a Dirichlet process to adapt a Gaussian mixture distribution (Dirichlet-Process Gaussian Mixture). This variant is preferable when the type of modality of the output information is not known in advance.

[0017] The invention proves particularly advantageous in the context of behavior prediction. In this case, the training data elements each comprise at least one ground truth behavior of at least one participant in a training traffic scene as ground truth information and training scene-specific information as associated training context information. The decoder then generates at least one behavior of the participant as output information. Both the ground truth behavior and the generated or predicted behavior of the participant are often described by trajectories, which, in addition to location and time information, can also include state information such as orientation, speed, and acceleration, etc.

[0018] A further advantageous application of the invention is the generation of digital images as training data for a machine learning (ML) component. In this case, the training data elements each comprise at least one ground truth image, such as an image of a traffic situation at an intersection in sunshine, and image attributes, such as the textual description and information "intersection in sunshine" captured with the help of corresponding sensors, as associated training context information. The decoder then generates at least one image as output information.

[0019] In addition to the claimed training method, the inference and corresponding systems are also claimed, specifically for the two use cases: predicting the behavior of at least one participant in a traffic scene and generating digital images as training data for a machine learning (ML) component. For this purpose, a model with a CVAE structure is used, which, according to the invention, has been trained with a set of training data elements for the respective use case.

[0020] The computer-implemented method for predicting the behavior of at least one participant in a traffic scene comprises the following steps: a. Aggregating scene-specific information of a traffic scene, b. Using the pre-trained prior encoder y to transform the scene-specific information into characteristics of a multivariate context probability distribution of latent features, c. Generating at least one latent representation of the traffic scene in the form of a context sample Z γ the multivariate context probability distribution, d. Conditioning the pre-determined joint probability distribution in the latent joint space on the context sample Z γ , e. Generate at least one conditioned sample Z cond-mix the conditioned joint probability distribution and f. Using the pre-trained decoder θ to generate a behavior for at least one participant in the traffic scene based on the conditioned sample Z cond-mix the joint probability distribution.

[0021] Accordingly, for inference, a representation of the traffic scene in latent space is assumed as a condition for the jointly learned joint probability distribution. This corresponds to a kind of cut through the joint probability distribution. The resulting conditional joint probability distribution is then used as the basis for behavior prediction with the help of the decoder. This largely prevents overlaps of different behavioral modes from being predicted.

[0022] In a preferred embodiment of the invention, the decoder θ uses in addition to a conditioned sample Z cond-mix the joint probability distribution the aggregated scene-specific information of the traffic scene for behavior prediction.

[0023] The computer-implemented method for generating digital images as training data for a machine learning (ML) component comprises the following steps: a. Providing image attributes, in particular in the form of a textual description of image content and / or environmental influences detected by sensors, b. Using the pre-trained prior encoder y to transform the image attributes into characteristics of a multivariate context probability distribution of latent features, c. Generating at least one latent representation of the image attributes in the form of a context sample Z γ the multivariate context probability distribution, d. Conditioning the pre-determined joint probability distribution in the latent joint space on the context sample Z γ , e. Generate at least one conditioned sample Z cond-mixthe conditioned joint probability distribution and f. Using the pre-trained decoder θ to generate at least one digital image based on the conditioned sample Z cond-mix the joint probability distribution.

[0024] This method can then be used to generate at least one new image based on a modified textual description, e.g. “intersection at dusk”.

[0025] The conditioning of the pre-determined joint probability distribution in the latent joint space on the context sample Z γ is preferably carried out by determining at least some of the degrees of freedom of the joint probability distribution by the context sample Z γ be determined so that the joint probability distribution determined for the entire latent joint space is reduced to a probability distribution in a subspace of the latent joint space.

[0026] The context sample Z γ and / or the conditioned sample Z cond-mix generated randomly and / or deterministically. Drawings

[0027] Embodiments and advantageous developments of the invention are explained in more detail below in conjunction with the figures. Fig. Figure 1 shows a training setup of a CVAE for predicting trajectories of participants in a traffic scene. Fig. 2 illustrates the measures according to the invention within the framework of the training method according to the invention using the Fig. 1 shown application example. Fig. 3 illustrates the measures according to the invention in the context of the inference for the Fig. 1 and Fig. 2 shown application example. Fig. Figure 4 illustrates the inventive conditioning of the joint probability distribution in the latent joint space as well as the resulting conditioned probability distribution in a subspace of the latent joint space. Description of implementation examples

[0028] In Fig. Figure 1 shows a training setup of a CVAE structure, which includes a label encoder ϕ designated by reference numeral 1, a prior encoder γ designated by reference numeral 2, and a decoder θ designated by reference numeral 5. For training this CVAE structure, a set of training data elements is provided, each of which includes at least one piece of ground truth information Y* 11 with associated training context information X* 12. In Fig. 1 shows such a training data element, which is used to explain the training method by way of example. Since the CVAE structure in the exemplary embodiment described here is to be trained for the behavior prediction of participants in a traffic scene, trajectory data of a vehicle 10 is used as ground truth information Y* 11. Scene-specific information describing the associated training traffic scene 20 is used as training context information X* 12. This is a four-way intersection. Vehicle 10 approaches this four-way intersection from the left and essentially has three behavior options: "Drive straight ahead," "Turn right," or "Turn left." Represented as ground truth information Y* is a trajectory of the behavior option "Drive straight ahead."

[0029] Using the label encoder ϕ 1 , the training data element, ie the ground truth information Y* 11 together with the training context information X* 12 , is converted into parameters of a multivariate label probability distribution N(µ ϕ , Σ ϕ ) 3 in the latent space. Independently of this, only the training context information X* 12 of the training data element is transformed into parameters of a multivariate context probability distribution N(µ γ ,Σ γ ) 4 in the latent space. The label probability distribution N(µ ϕ ,Σ ϕ ) 3 and the context probability distribution N(µ γ ,Σ γ ) 4 are modeled as standard normal distributions in the course of the training procedure described here, which are characterized by the parameters, expected value vector µ ϕ or µ γ and covariance matrix Σ ϕ or Σ γTo do this, a distance between the label probability distribution N(µ ϕ , Σ ϕ ) 3 and the corresponding context probability distribution N(µ γ ,Σ γ ) 4 and compared with a first optimization criterion.

[0030] At this point, it should be noted that for the determination of the first distance between the label probability distribution N(µ ϕ ,Σ ϕ ) 3 and the context probability distribution N(µ γ ,Σ γ ) 4 any distance measure can be used. Preferably, this first distance is determined based on the Kullback-Leibler (KL) divergence. The KL divergence for two multivariate unimodal Gaussian distributions N(μp,∑p) and N(μq,∑q) of dimension k is given by: DKL(p‖q)=12[log|∑q||∑p|−k+(μp−μq)T∑q−1(μp−μq)+tr{∑q−1∑p}]

[0031] Based on the label probability distribution N(µ ϕ ,Σ ϕ ) 3 or their parameters, a label sample Z designated 13 ϕ generated. In addition, based on the context probability distribution N(µ γ ,Σ γ ) 4 or their parameters a context sample Z designated 14 γ generated. Samples 13 and 14 can be drawn either randomly or deterministically, e.g., as in the Unscented Kalman Filter.

[0032] The decoder θ5 generates based on the label sample Z ϕ 13 a trajectory as output information Y, which is denoted here by 15. In the embodiment described here, the decoder θ5 also uses the training context information X* 12. In addition, the decoder θ could also use the context sample Z γ 14. If multiple label samples Z ϕ from the label probability distribution N(µ ϕ ,Σ ϕ) 3 is drawn, then the decoder θ 5 generates a family of trajectories whose distribution is denoted here by 6.

[0033] Finally, a distance between the ground truth information Y* 11 and the generated output information Y 15, i.e., between the ground truth trajectory 11 and the output trajectory 15, is determined and compared with a predefined second optimization criterion. If the first and / or second optimization criterion is not met, the parameters of the label encoder ϕ 1, the prior encoder y 2, and / or the decoder θ5 are modified. The training process continues until the first and / or second optimization criterion is met.

[0034] The Fig. The procedure shown in Figure 1 and described above essentially corresponds to the known training of a CVAE.

[0035] The inventive measures of the proposed training method, namely the determination of a joint probability distribution in a latent joint space (ex-post density estimation), are described below in connection with Fig. 2 is explained in more detail. Fig. 2 builds on the content Fig. 1. Therefore, in Fig. 2 the same reference numerals as in Fig. 1, for identical elements. For an explanation of these elements, please refer to the explanations for Fig. 1.

[0036] After reaching the first and / or the second optimization criterion, i.e. after completion of the conventional training, according to the invention a label samples Z ϕ 13, a context sample Z γ 14 and a composite sample [Z ϕ ; Z γ ] 16 in a latent composite space. A composite sample [Z ϕ ; Z γ] 16 consists here of a sequence of the label sample Z ϕ 13 and the context sample Z γ 14, but can also be formed by any other combination of these two samples 13 and 14. It is important that the composite samples [Z ϕ ; Z γ ] 16 to span a latent join space. In any case, based on the training data set, a set 17 of join samples [Z ϕ ; Z γ ] 16. According to the invention, on the basis of this set of composite samples [Z ϕ ; Z γ ] a joint probability distribution 18 is determined in the latent joint space. In the embodiment described here, a Gaussian mixture distribution is applied to the composite samples [Z ϕ ; Z γ] of the set. The fitting can be performed parametrically, for example, if the number of components and possibly also their properties, such as axis-aligned Gaussians or uniform Gaussians, are specified. However, it can also be performed nonparametrically, for example, using a Dirichlet-process Gaussian mixture.

[0037] After the explanations to Fig. 1 and Fig. 2, the training method according to the invention can be summarized as follows: 1. Standard training of a CVAE; 2. Use the trained label encoder ϕ and the trained prior encoder γ to map the training dataset into the latent space and set of joint samples [Z ϕ ; Z γ ] in a latent interconnected space; 3. Determine a model for the analytical description of the empirical distribution of the composite samples, e.g. by fitting a Gaussian mixture distribution to the composite samples [Z ϕ ; Zγ ] in the network area.

[0038] The inference is explained below using the claimed prediction method as an example in conjunction with Fig. 3 explained. Fig. 3 builds on the content Fig. 1 and Fig. 2. Therefore, the same reference numerals are used here as in Fig. 1 and Fig. 2, for the same elements and refer to the corresponding explanations to Fig. 1 and Fig. 2.

[0039] For the inference, here the prediction of the behavior of at least one participant in a traffic scene 21, the trained prior encoder γ 2, the joint probability distribution 18 and the trained decoder θ are required.

[0040] The prediction method according to the invention uses context information that describes the current traffic scene 21. A computer-implemented system for behavior prediction therefore comprises a perception layer for aggregating scene-specific information from various information sources. These can be on-board sensors, such as lidar sensors, radar sensors, and / or RGB cameras installed on the ego vehicle, or external sensors, such as lidar sensors, radar sensors, and / or RGB cameras installed in or on infrastructure elements. Other possible information sources include stored map information as well as queryable weather and road condition information, traffic situation information, etc. The information from the various information sources is aggregated by the perception layer and preprocessed into context information.With the help of the trained prior encoder γ 2 , this scene-specific context information is mapped into the latent space by converting it into parameters of a normal distribution N(µ. γ ,Σ γ ) 4 of latent features. With the help of an evaluation and calculation unit, a latent representation 14 of the traffic scene 21 is then created in the form of a context sample Z γ 14 of the normal distribution N(µ γ ,Σ γ ) 4 is generated. This conditioned the pre-determined joint probability distribution 18 in the latent joint space by at least some of the degrees of freedom of the joint probability distribution 18 being determined by the context sample Z γThis reduces the joint probability distribution 18 determined for the entire latent joint space to a probability distribution 19 in a subspace of the latent joint space, which corresponds to a kind of cut through the joint probability distribution 18. From this conditioned joint probability distribution 19, at least one conditioned sample Z is then cond-mix 23 and fed to the trained conditioned joint probability distribution and decoder θ5. Based on the conditioned sample Z cond-mix the conditioned joint probability distribution 19, the decoder θ5 then generates or predicts a behavior, here a trajectory 25, for at least one participant in the traffic scene 21. At this point, it should be expressly pointed out that with the method described above, not only the behavior of individual participants in the traffic scene can be predicted individually, but also the behavior of several participants in a traffic scene together, which is also referred to as “joint prediction”.

[0041] The conditioning of the joint probability distribution 18 on a context sample Z γ can be done as follows:

[0042] Assume that the joint probability distribution can be described by C Gaussian distributions with mean µ (c) ) and covariance matrix Σ (c) with the weights w c parameterize: p(z))∑ccwcN(z;μ(c),∑(c)), where C is predefined. The random variable z represents the composite samples z = [z1; z2], where dim(z ϕ ) = dim (z1) and dim(z γ ) = dim (z2). Then each component c can be expressed as μ(c)=[μ1(c)μ2(c)],∑(c)=[∑11(c)∑12(c)∑21(c)∑22(c)].

[0043] The aim is to determine the conditioned joint probability distribution p(z1|z2)=p(z1,z2)p(z2) The conditioned distribution of a component in a multivariate Gaussian distribution is given by N(z1|z2;μ(c),∑(c))=N(z1;μ1|2(c),∑1|2(c))μ1|2(c)=μ1(c)+∑12(c )(∑22(c))−1(z2−μ2(c))∑1|2(c)=∑11(c)−∑12(c)(∑22(c))−1∑21(c).

[0044] The marginal distribution p(z2) is simply given by ∑cCwcN(z2;μ22(c)∑22(c)) Therefore, p(z1|z2) can be calculated as p(z1|z2)=∑cCwcN(z2;μ22(c),∑22(c))p(z2)N(z1;μ1|2(c),∑1|2(c)), where the fraction provides the new component weights, which are normalized using the marginal distribution p(z2).

[0045] Fig. Figure 4 shows, in the left half of the image, two dimensions of a three-dimensional compound probability distribution in the form of a Gaussian mixture distribution 40. The four components of this Gaussian mixture distribution are represented by four ellipses 41 to 44. They represent different modes of the output information, for example, four different behavioral options in a traffic scene with a four-way intersection. A serpentine line 45 indicates an inventive conditioning of this compound probability distribution 40. The resulting two-dimensional conditioned probability distribution 46 is shown in the right half of the image and essentially comprises only two preferred behavioral options, for example, "drive straight" or "turn right."Since at least components 43 and 44 can be clearly distinguished from each other, samples from a mixed range, for example, between 2 and 4, are rather unlikely. In any case, the probability that the inference will produce trajectories that lead diagonally to the left of the road instead of trajectories for "driving straight" or "turning right" is very low. Fig. 4 illustrates, using the example of behavioral prediction, that the inventive measures can effectively prevent a mixing of mutually exclusive behavioral modes. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited non-patent literature

[0000] Ghosh et al., “From variational to deterministic autoencoder,” https: / / arxiv.org / abs / 1903.12436

[0001]

Claims

[1] Computer-implemented method for training a model comprising a Conditional Variational Auto Encoder (CVAE) structure with a label encoder ϕ (1), a prior encoder γ (2) and a decoder θ (5), wherein a set of training data elements is provided for the training method, each comprising at least one piece of ground truth information Y* (11) with associated training context information X* (12), and wherein the following steps are carried out for each training data element: - Use of the label encoder ϕ (1) to transform the training data element into characteristics of a multivariate label probability distribution (3) of latent features, - Use of the prior encoder γ (2) to transform the training context information X* (12) into parameters of a multivariate context probability distribution (4) of latent features, - determining a first distance between the label probability distribution (3) and the context probability distribution (4) and comparing the first distance with a predetermined first optimization criterion, - Generate at least one label sample Z ϕ (13) based on the parameters of the multivariate label probability distribution (3), - Use of the decoder θ (5) to generate at least one output information Y (15) based on the label sample Z ϕ (13), - determining a second distance between the ground truth information Y* (11) and the generated output information Y (15) and comparing the second distance with a predetermined second optimization criterion, - modifying at least one parameter of the label encoder ϕ (1), the prior encoder γ (2) and / or the decoder θ (5) if the first and / or the second optimization criterion is not met; characterized by that if the first and / or the second optimization criterion is met, for at least some of the training data elements a label sample Z ϕ (13), a context sample Z γ (14) and a composite sample [Z ϕ ; Z γ ] (16), where the composite sample [Z ϕ ; Z γ ] (16) as a combination of the label sample Z ϕ (13) and the context sample Z γ (15) and that based on a set of composite samples [Z ϕ ; Z γ ] (16) a joint probability distribution (18) is determined in a latent joint space. [2] Method according to claim 1, characterized by that the training data element is in parameters of a first multivariate normal distribution N(µ ϕ , Σ ϕ ) of latent features, in particular into parameters from which a first expected value vector µ ϕand a first covariance matrix Σ ϕ and / or that the training context information X* is converted into parameters of a second multivariate normal distribution N(µ γ , Σ γ ) of latent characteristics, in particular into parameters from which a second expectation vector µ γ and a second covariance matrix Σ γ can be derived. [3] Method according to one of claims 1 or 2, characterized by that the label samples Z ϕ (13) and / or the context samples Z γ (14) are generated randomly and / or deterministically. [4] Method according to one of claims 1 to 3, characterized by that the decoder θ (5) generates at least one output information Y (15) the label sample Z ϕ (13) and the training context information X* (12) and / or the associated context sample Z γ (14) uses. [5] Method according to one of claims 1 to 4, characterized by that to determine the joint probability distribution (18) a Gaussian mixture distribution is applied to the joint samples [Z ϕ ; Z γ ] (16) is fitted in the composite space. [6] Method according to one of claims 1 to 5, characterized by that the fitting is carried out parametrically, whereby at least the number of Gaussian distribution components and / or their properties are specified. [7] Method according to one of claims 1 to 5, characterized by that the fitting is non-parametric, in particular with a Dirichlet process to fit a Gaussian mixture distribution (Dirichlet-Process Gaussian Mixture). [8] Method according to one of claims 1 to 7 for training a model for predicting a behavior of at least one participant in a traffic scene, in particular for predicting trajectories of at least one participant in a traffic scene, characterized bythat the training data elements each comprise at least one basic truth behavior as ground truth information and training scene-specific information as associated training context information and that the decoder generates at least one behavior as output information. [9] Method according to one of claims 1 to 7 for training a model for generating digital images as training data for a Machine Learning (ML) component, characterized by that the training data elements each comprise at least one ground truth image and image attributes as associated training context information and that the decoder generates at least one image as output information. [10] Computer-implemented method for predicting the behavior of at least one participant in a traffic scene using a model trained according to claim 8, comprising the following steps: a. Aggregating scene-specific information of a traffic scene (21), b. Using the pre-trained prior encoder γ (2) to transform the scene-specific information into characteristics of a multivariate context probability distribution (4) of latent features, c. Generating at least one latent representation of the traffic scene in the form of a context sample Z γ (14) the multivariate context probability distribution (4), d. Conditioning the pre-determined joint probability distribution (18) in the latent joint space on the context sample Z γ (14), e. Generate at least one conditioned sample Z cond-mix (23) the conditioned joint probability distribution (19) and f. Using the pre-trained decoder θ (5) to generate a behavior for at least one participant of the traffic scene based on the conditioned sample Zcond-mix (23) the conditioned joint probability distribution. [11] Method according to claim 10, characterized by that the pre-trained decoder θ (5) additionally uses the aggregated scene-specific information of the traffic scene (21) to generate a behavior for at least one participant of the traffic scene. [12] A computer-implemented method for generating digital images as training data for a machine learning (ML) component, wherein the digital images are generated using a model trained according to claim 9 and the method comprises at least the following steps: a. Providing image attributes, in particular in the form of a textual description of image content, b. Using the pre-trained prior encoder γ to transform the image attributes into characteristics of a multivariate context probability distribution of latent features, c. Generating at least one latent representation of the image attributes in the form of a context sample Z γ the multivariate context probability distribution, d. Conditioning the pre-determined joint probability distribution in the latent joint space on the context sample Z γ , e. Generate at least one conditioned sample Z cond-mix the conditioned joint probability distribution and f. Using the pre-trained decoder θ to generate at least one digital image based on the conditioned sample Z cond-mix the conditioned joint probability distribution. [13] Method according to claim 12, characterized by that the pre-trained decoder θ additionally uses the provided image attributes to generate at least one digital image. [14] Method according to one of claims 10 or 11, characterized bythat the pre-determined joint probability distribution (18) in the latent joint space is applied to the context sample Z γ is conditioned by determining at least some of the degrees of freedom of the joint probability distribution (18) by the context sample Z γ (14) are determined so that the joint probability distribution (18) determined for the entire latent joint space is reduced to a probability distribution (19) in a subspace of the latent joint space. [15] Method according to one of claims 10 to 12, characterized by that the context sample Z γ (14) and / or the conditioned sample Z cond-mix (23) are generated randomly and / or deterministically. [16] Computer-implemented system for predicting the behavior of at least one participant in a traffic scene, the system comprising at least the following components: a. a perception layer for aggregating scene-specific information of a traffic scene, b. a model trained according to claim 8 i. with a prior encoder γ (2) to transform the scene-specific information into parameters of a multivariate context probability distribution (4) of latent features, ii. a joint probability distribution (18) and iii. a decoder θ (5) for generating a behavior for at least one participant in the traffic scene; c. an evaluation and computing unit designed to i. at least one latent representation of the traffic scene in the form of a context sample Z γ (14) to generate the multivariate context probability distribution (4), ii. the joint probability distribution (18) in the latent joint space on the context sample Z γ (14) to condition, and iii. at least one conditioned sample Z cond-mix (23) of the conditioned joint probability distribution (19) as input for the decoder θ (5). [17] Computer-implemented system for generating digital images as training data for a machine learning (ML) component, the system comprising at least the following components: a provisioning facility for image attributes, b. a model trained according to claim 9 i. with a prior encoder γ to transform the image attributes into parameters of a multivariate context probability distribution of latent features, ii. a joint probability distribution and iii. a decoder θ for generating at least one digital image; c. an evaluation and computing unit designed to i. at least one latent representation of the image attributes in the form of a context sample Z γ to generate the multivariate context probability distribution, ii. the joint probability distribution in the latent joint space on the context sample Z γ to condition, and iii. at least one conditioned sample Z cond-mix the conditioned joint probability distribution as input for the decoder θ.

Images