DE102020129451A1 - Method for predicting driving interventions, method for training an algorithm and motor vehicle - Google Patents

Abstract

Method for predicting driving interventions (21) for a road user (2) and/or at least one other road user (5, 6), with an algorithm (20) being applied in a respective prediction step in order for the and/or the respective other road user (2, 5, 6) to predict several different driving interventions (21), wherein a method of machine learning for training the algorithm comprises several training steps, in each of which the algorithm (20) with an initial parameterization present at the beginning of the respective training step on respective Input data (26) from training data sets (25) is used to determine a number of different driving interventions (21), - a respective error measure (31) for the determined driving interventions (21) is determined for the respective training data set (25), - for each training data set (25) the smallest error measure (31) is selected, and- the parameters (28) of the algorithm (20) in dependence be fitted with a cost function (34) which depends solely on the selected error measures (32).

Description

• - wobei das Umfeld des Verkehrsteilnehmers betreffende Umgebungsdaten, die den wenigstens einen weiteren Verkehrsteilnehmer betreffen, erfasst werden,
• - wobei in einem ersten Prädiktionsschritt ein Algorithmus auf Eingangsdaten angewandt wird, die von den Umgebungsdaten abhängen, um als Ausgangsdaten für den und/oder den jeweiligen weiteren Verkehrsteilnehmer jeweils mehrere verschiedene Fahreingriffe zu prädizieren, aus denen verschiedene Verkehrssituationen resultieren,
• - wobei in nachfolgenden Prädiktionsschritten der Algorithmus auf jeweilige Eingangsdaten angewandt wird, die von einer jeweiligen der im vorangehenden Prädiktionsschritt ermittelten Verkehrssituationen abhängen, um die verschiedenen Fahreingriffe als Ausgangsdaten zu prädizieren, aus denen wiederum verschiedene Verkehrssituationen resultieren.

The invention relates to a method for predicting driving interventions for a road user, in particular a motor vehicle, and/or at least one other road user,

• - wherein the surroundings of the road user relating to environmental data that relate to the at least one other road user are recorded,
• - In a first prediction step, an algorithm is applied to input data that depends on the environmental data in order to predict several different driving interventions as output data for the and/or the respective other road user, from which different traffic situations result,
• - In subsequent prediction steps, the algorithm is applied to respective input data that depend on a respective traffic situation determined in the preceding prediction step, in order to predict the various driver interventions as output data, from which in turn different traffic situations result.

In addition, the invention relates to a method for training an algorithm that is used to predict driving interventions for a particular road user, and a motor vehicle.

For the behavior planning of automated vehicles and also for the implementation of other complex driver assistance systems, for example for collision avoidance or collision mitigation and systems for assistance in certain driving situations, for example with overtaking assistants, a prediction of the intentions and actions of the surrounding vehicles is necessary in addition to the pure perception. In addition, it is advantageous to plan one's own trajectory or, in the case of at least partially manual driving, to predict it.

Predicting the behavior of other drivers or the driver of one's own vehicle is relatively complex, since numerous factors, such as a future course of the roadway, the current state of the vehicle and possible interactions with other vehicles, have to be taken into account. Since different drivers can react differently to the same traffic situation and the reaction of automated systems can also differ depending on the specific implementation of the respective system, uncertainties in the prediction must be taken into account, especially with regard to interaction with other vehicles. For example, if a vehicle is approaching a roundabout where another motor vehicle is already located, some drivers will accelerate in the same driving situation in order to enter the roundabout before the other vehicle reaches their own junction, and some drivers will brake to yield the right of way to the other vehicle. At this point, a prediction would have to recognize the modality of the situation and predict both possible behaviors or provide a bundle of trajectories, part of which stops and another part enters the roundabout. Similar divergent trajectories can occur, for example, at junctions and traffic lights. In addition, differences can also occur, for example, which distances are maintained by different drivers or in which driving situations vehicles in front are overtaken and the like.

A possible approach for such a prediction of driving interventions and thus ultimately also of trajectories is to assume that current driving interventions depend exclusively on the current driving situation. This corresponds to the Markov assumption. As a result, an iterative prediction can take place in which future driving interventions are predicted for the individual road users in a recorded traffic situation. Based on these driving interventions, a future traffic situation is predicted and the prediction of the driving interventions is repeated for this, etc.

An example of this is given, for example, in the publication J. Schulz et al., "Learning Interaction-Aware Probabilistic Driver Behavior Models from Urban Scenarios", 30th Intelligent Vehicles Symposium (IV), 1326 - 1333, 2019. To predict driving interventions, this publication proposes learning a regression function with the help of neural networks, which determines a subsequent action for the current vehicle and environment condition, specifically an acceleration and steering position used. A kinematic model is then used to update the traffic situation depending on the determined action, and the action prediction can be repeated for the updated model. When predicting actions, it is assumed here that the determined parameters, namely the acceleration and the steering angle, are both normally distributed and independent of one another. Uncertainties are thus taken into account by providing mean values and standard deviations for both variables through the regression function. The procedure explained is therefore not very suitable for making longer-term predictions in driving situations in which possible trajectories diverge greatly from one another and/or a There is a strong coupling between steering angle and acceleration, since the assumed probability distribution significantly overestimates the probability of certain actions, which is actually close to zero, which means that a relatively large amount of computing time is required to predict trajectories that hardly ever occur in real life. In addition, the quality of the prediction also decreases. Although this problem can be mitigated by assuming a sum of several normal distributions instead of a normal distribution, ie using a so-called Gaussian Mixture Model, for example, real probability distributions are typically not optimally represented by this either.

Another stepwise prediction approach is from the reference US2020249674A1 famous. There, the environment in cells and the driver's options for action are discretized and a decision tree is built with the help of a prediction function that can be trained through machine learning. In order to avoid a decision tree that is too broad, on the one hand a driver type is classified in order to restrict the options for action, and on the other hand branches with a low probability are subsequently discarded. The disadvantage here is that a relatively large amount of prior information, for example with regard to the driver type, is required and due to the discretization and the discarding of certain branches, the accuracy of the prediction necessarily decreases, which means that use in complex urban driving situations is only possible to a limited extent.

The invention is therefore based on the object of specifying a method for predicting driving interventions for a road user, which is improved in comparison and in particular enables a robust prediction even in complex traffic situations without additional prior information.

• - der Algorithmus mit einer zu Beginn des jeweiligen Trainingsschritts vorliegenden Anfangsparametrisierung auf die jeweiligen Eingangsdaten der Trainingsdatensätze angewandt wird, um für jeden der Trainingsdatensätze mehrere verschiedene Fahreingriffe zu ermitteln,
• - für den jeweiligen Trainingsdatensatz ein jeweiliges Fehlermaß für die Abweichung jedes der ermittelten Fahreingriffe von dem jeweiligen Sollfahreingriff ermittelt wird,
• - für jeden Trainingsdatensatz das kleinste der für diesen Trainingsdatensatz in diesem Trainingsschritt bestimmten Fehlermaße ausgewählt wird, und
• - die Parameter des Algorithmus in Abhängigkeit einer Kostenfunktion angepasst werden, wobei die Kostenfunktion ausschließlich von den ausgewählten Fehlermaßen abhängt oder wobei die ausgewählten Fehlermaße in der Kostenfunktion stärker gewichtet werden als die anderen Fehlermaße.

The object is achieved by a method of the type mentioned at the outset, wherein an algorithm is used to predict the driving interventions for the respective road user, which is parameterized by a number of parameters that are determined by a machine learning method in which a number of training data are used each relate to input data for the algorithm and a target driving intervention, the method of machine learning comprising several training steps, in each of which

• - the algorithm with an initial parameterization present at the beginning of the respective training step is applied to the respective input data of the training data sets in order to determine several different driving interventions for each of the training data sets,
• - a respective error measure for the deviation of each of the determined driving interventions from the respective target driving intervention is determined for the respective training data set,
• - the smallest of the error measures determined for this training data set in this training step is selected for each training data set, and
• - the parameters of the algorithm are adapted as a function of a cost function, the cost function depending exclusively on the selected error measures or the selected error measures being weighted more heavily in the cost function than the other error measures.

Similar to the cited publication by J. Schulz et al., a trained algorithm is used in the method according to the invention in order to predict actions by drivers depending on the traffic situation. However, instead of predicting individual parameters of a fixed probability distribution, an algorithm is used that predicts a specified number of specific driving interventions that can be understood as random samples or representatives for a conditional probability distribution. The predicted driving interventions can also be referred to as hypotheses and can be understood as representative drawings from an underlying, unknown density of possible behaviors in this situation.

At the same time, however, it is avoided that the possible actions are restricted to a relatively small number of discrete actions, since the plurality of predicted actions specifically depend on the traffic situation at hand. In the individual prediction step, a predetermined number of different driving interventions is thus predicted, but which driving interventions are involved is predetermined by the respective traffic situation. Thus, the advantages of the P. Schulz et al. Proposed procedure achieved, in contrast to this teaching, no assumptions about the form of the underlying probability density must be made, whereby the disadvantages explained above are avoided or at least their effects are significantly reduced.

In addition, the prediction efficiency can be further increased, since for the in Schulz et al. used probability density for an iterative prediction certain samples must be selected from these probability densities and such a sampling in the method according to the invention can be omitted since specific driving interventions are already being predicted.

The procedure described can be viewed as a multi-hypothesis prediction in which a number of hypotheses are provided for a particular traffic situation. Training algorithms to carry out such a multi-hypothesis prediction is known per se from the publication by C. Rupprecht et al., "Learning in an Uncertain World: Representing Ambiguity Through Multiple Hypotheses", The IEEE International Conference on Computer Vision (ICCV), 2017 and is used there, for example, to classify human poses. Within the scope of the invention, it was found that the procedure described there, with certain modifications, is well suited for the iterative prediction of driving interventions, for example in the context of trajectory planning.

However, it was also found that if the procedure proposed there were used unchanged, the problem arises that the distribution of the hypothesis generated there does not correspond to the actual probability distribution that is specified by the training data. For example, for normally distributed results, the hypotheses will have a significantly greater variance than the training data used in the training. As a result, numerous hypotheses are generated which are actually not relevant or are very improbable, with the result that more and more irrelevant or extremely improbable traffic situations are determined as the number of iterations increases. On the one hand, this leads to a very high computational effort and memory requirement due to the large number of traffic situations taken into account that are irrelevant or only slightly relevant, and on the other hand to the fact that ultimately an additional discarding of irrelevant options would be necessary. However, this is not trivially possible.

Within the scope of the invention, however, it was recognized that, surprisingly, it is possible to avoid or at least significantly alleviate the problems mentioned by slightly modifying the known training for a multi-hypothesis prediction, namely in particular by modifying the cost function used in the training of the algorithm that the trained algorithm for multi-hypothesis prediction is also well suited for an iterative prediction of driving interventions. Possibilities for such a modification will be explained in more detail later.

The method according to the invention can be used, for example, to determine a trajectory for one's own vehicle and/or other vehicles. Such a trajectory is already described by the time sequence of the predicted driving interventions. However, in order to obtain a representation of the trajectory in spatial space, a kinematic model, for example a one- or two-lane model for motor vehicles, can be used to determine the trajectory depending on the driving interventions, in particular on the steering angles and/or accelerations used. In addition or as an alternative, the positions and/or speeds of the road users can also be predicted at different times in the future, for example in one, two or three seconds. As explained at the beginning, such information can be highly relevant for planning automated driving operations or for other assistance functions.

The method can be used to predict both the behavior of the road user and the behavior of all other road users. This can be useful, for example, if the behavior of the driver of the host vehicle is to be predicted as part of assisted driving or to provide additional information. Alternatively, however, a fixed trajectory can also be assumed for the road user, for example, which has been planned using a separate method, e.g. as part of highly or fully automated driving. The position, orientation and/or speed of the road user can thus be predetermined by this trajectory in all predicted traffic situations, so that only the driving interventions of the at least one other road user are predicted and the predicted traffic situations thus change, for example, only with regard to the positions, orientations and/or Distinguish between the speeds of other road users. This can be used, for example, to check whether a planned trajectory can be driven without any problems or whether this results in problematic driving situations, e.g. situations with a high risk of collision. If problematic driving situations are detected, a replanning of the trajectory for the own vehicle can be initiated in this case, for example. In principle, the driving behavior of other road users that are taken into account can also be specified, e.g. if a planned trajectory of another road user is received. In this case, the driving interventions can be predicted, e.g. only for the own vehicle and/or for other road users for whom no planned trajectory is known.

A neural network, for example, can be used as the algorithm. The provision of different driving interventions can be realized in this case, for example, that in a first or several first shifts a common Same processing of the input data for all driving interventions to be determined takes place, the processing result of this first layer or layers being further processed by one or more further layers, which are parameterized differently in order to predict different driving interventions.

The algorithm can also be referred to as a trained algorithm. The training of the algorithm can be understood as a learning process or as a type of regression process.

The training, ie in particular the implementation of the training steps or the machine learning method, can be regarded as a preparatory step which is not part of the claimed method. The method according to the invention thus relates to the use of an algorithm that has been trained or parameterized in a specific way. In a development of the method, however, the training of the algorithm can also be included as an additional method step in the claimed method.

The environmental data can, for example, be detected by sensors and/or received via a communication device. The environmental data for predicting the driving interventions can be processed, for example, by the road user, but also by an external device, for example by a backend server. The algorithm can be trained in particular on a separate device and also at a time interval from the implementation of the method. It can, for example, be carried out by a vehicle manufacturer or another service provider. However, it is also possible for further training or something similar to take place in the motor vehicle itself during ongoing operation of a motor vehicle.

The training data sets can include the input data directly or the input data can be determined from the training data set by an upstream algorithm. For example, the respective training data set can describe the entire traffic situation and the input data can form a feature vector that relates to the traffic situation from the perspective of the respective road user, i.e., for example, properties of the route ahead, the distance to a preceding and/or following motor vehicle, the occupancy of neighboring lanes , a distance from specific features, for example from a junction or a roundabout, etc. Approaches to extracting feature vectors from an environment model are known per se in the prior art and will not be explained in detail. The target driving intervention can specify which driving intervention should ideally be determined by the algorithm for the input data described by the respective training data record. This can in particular be a driving intervention by a road user that actually took place or was observed in a traffic situation described by the input data. The target driving intervention can thus map an observed behavior or a behavior that is to be predicted.

The respective error measure can be determined by applying a processing function to a distance measure for the distance between the respective determined driving intervention and the target driving intervention specified by the respective training data set, with the slope of the processing function decreasing at least in a value range of the processing function as the value of the distance measure increases. As already mentioned, when using standard error measures, for example a squared error, as described in the cited publication by C. Rupprecht et al. is proposed, a distribution of hypotheses is achieved that deviates significantly from the probability density in the training data. It was recognized that this effect can be avoided or reduced in particular by using an error measure whose slope decreases with increasing errors.

In a specific implementation, for example, the driving intervention and the target driving intervention can be represented as vectors and the distance can be calculated as the sum of powers, in particular squares, of the differences between the vector entries or of amounts of these powers. Typically, corresponding distance measures are used directly as summands of the cost function. Deviating from this, it is proposed to additionally apply a processing function to a distance measure determined in this way in order to achieve a different weighting of distances and thereby achieve that hypotheses are generated at least approximately with the same probability, so that each of the hypotheses is representative of the same part of the probability measure stands.

The respective error measure can be determined as an algorithm of the or a distance measure for the distance between the respective determined driving intervention and the respective target driving intervention. In particular, a logarithm can thus be used as the processing function explained above. In the context of the invention, it was found that this means that there is a particularly small deviation in the distribution of the hypotheses from the truth probability distribution can be achieved in the training data. With a suitable choice of the distance measure, it can even be achieved that in the theoretical case of generating an infinite number of hypotheses, i.e. an infinite number of possible driving interventions, the probability distribution that is specified by the training data sets can be reproduced more precisely. It can be advantageous if the distance is selected in such a way that a certain constant, in particular smaller, offset is used, which can prevent the distance from reaching zero and thus the negative logarithm becoming infinite.

Other examples of processing functions that decrease in slope as the value of the distance measure increases are higher-order roots or the tanh function.

The driving intervention and the target driving intervention can each be defined by a vector, with the distance being proportional to the 2-norm of the difference vector between driving intervention and target driving intervention or to any power of this 2-norm. It was recognized that this choice of distance measure allows probability distributions to be mapped particularly precisely in the training data. For example, if a logarithm is used as the processing function, a change in the power corresponds to a change in the prefactor of the error measure, which is typically insignificant when minimizing the cost function. For example, the square of the 2-norm can be used as a distance measure.

In at least one of the prediction steps, a clustering method can be used to find a cluster center that represents a cluster of multiple driving interventions and/or multiple traffic situations and/or multiple input data, and this cluster center can be used instead of the individual driving interventions and/or traffic situations and/or input data of the respective clusters are further processed. This can be advantageous since otherwise a very large number of possible traffic situations must be taken into account in late prediction steps. Without clustering, the number of traffic situations to be considered increases with each iteration step by a factor that corresponds to the product of the number of road users considered and the number of different driving interventions predicted for each road user. By using clustering, the computing effort for the prediction can thus be significantly reduced or, with the same computing effort, the length of the prediction horizon can be increased and/or the time resolution of the prediction can be improved.

If, as explained above, it is at least approximately achieved that the distribution of the hypotheses generated corresponds to the actual conditional probability distribution specified by the training data sets, such clustering is possible without any problems, since it ultimately corresponds to a resampling of the probability distribution defined by the individual hypotheses or determined driving interventions with a lower resolution . Approaches to clustering are well known in the art and will not be discussed in detail. Essentially, cluster centers are selected during clustering so that the distances between the points in the respective cluster and these cluster centers are minimal. It is particularly advantageous here if, as explained above for the error measure, a processing function is also applied to a standard distance measure to determine the distances between cluster centers, the slope of which decreases at least in a value range as the value of the distance measure increases. For example, as explained above, a logarithm can be used.

From the traffic situation, a feature vector can be determined for the individual road user, which contains at least one position and/or a relative speed with respect to at least one other road user and/or with respect to the roadway and/or the at least one feature of a preceding route section as entries, with the Input data to the algorithm consists of or includes the feature vector. For example, the feature vector can describe the distance to the vehicle in front, to the edge of the road, to route features such as a curve, a junction, an intersection or a roundabout, or the like. In addition, a respective relative speed can be taken into account. In addition or as an alternative, information such as the time until a collision with another road user with unchanged driving, a direction of travel, for example with regard to the center line and/or the edge of the road, a transverse position of your own vehicle or that of a vehicle in front of you in the respective lane and the like can be included as part of the feature vector are taken into account. The traffic situation can also be represented as an image or the feature vector can include the image data of such an image.

The respective traffic situation can be determined by parameterizing a kinematic model, in particular a one-lane or two-lane model, for the respective road user according to the driving intervention. For example, based on the acceleration described by the driving intervention, a current speed for the ver road users are identified. Based on this speed and a current steering angle specified by the driving intervention, known kinematic models, for example using the single-lane model, can be used to determine the position of the road user with which orientation after a certain elapsed time interval.

Depending on at least one of the determined driving interventions and/or a respective trajectory defined by the chronological sequence of the driving interventions for the road user and/or the other road user, at least one actuator of the road user can be controlled to carry out a driving intervention and/or an information device of the road user for output be controlled by a notice. Information can be given in particular to the occupants of the road user, for example a driver of a vehicle, or to persons outside the road user, for example other road users.

• - der Algorithmus mit einer zu Beginn des jeweiligen Trainingsschritts vorliegenden Anfangsparametrisierung auf die jeweiligen Eingangsdaten der Trainingsdatensätze angewandt wird, um für jeden der Trainingsdatensätze mehrere verschiedene Fahreingriffe zu ermitteln,
• - für den jeweiligen Trainingsdatensatz ein jeweiliges Fehlermaß für die Abweichung jedes der ermittelten Fahreingriffe von dem jeweiligen Sollfahreingriff ermittelt wird,
• - für jeden Trainingsdatensatz das kleinste der für diesen Trainingsdatensatz in diesem Trainingsschritt bestimmten Fehlermaße ausgewählt wird, und
• - die Parameter des Algorithmus in Abhängigkeit einer Kostenfunktion angepasst werden, wobei die Kostenfunktion ausschließlich von den ausgewählten Fehlermaßen abhängt oder wobei die ausgewählten Fehlermaße in der Kostenfunktion stärker gewichtet werden als die anderen Fehlermaße.

In addition to the method for predicting driver interventions, the invention relates to a method for training an algorithm, which is used to predict driver interventions for a respective road user and which is parameterized by a number of parameters, by a machine learning method, using a number of training data sets, each of which is input data relate to the algorithm and a target driving intervention, the method of machine learning comprising a plurality of training steps, wherein in the respective training step

• - the algorithm with an initial parameterization present at the beginning of the respective training step is applied to the respective input data of the training data sets in order to determine several different driving interventions for each of the training data sets,
• - a respective error measure for the deviation of each of the determined driving interventions from the respective target driving intervention is determined for the respective training data set,
• - the smallest of the error measures determined for this training data set in this training step is selected for each training data set, and
• - the parameters of the algorithm are adapted as a function of a cost function, the cost function depending exclusively on the selected error measures or the selected error measures being weighted more heavily in the cost function than the other error measures.

The further possible configurations already explained above with regard to the training of the algorithm can also be transferred to the method according to the invention for training an algorithm. In particular, as explained above, the error measure can be determined by applying a processing function to a distance measure, it being possible in particular to use a logarithm as the processing function.

The invention also relates to a motor vehicle, comprising at least one sensor and/or at least one communication device for detecting environmental data relating to the area surrounding the motor vehicle and a processing device, the processing device being set up to carry out the method according to the invention for predicting driving interventions for a road user if the Motor vehicle is used as the road user. Suitable sensors and communication devices for acquiring environmental data are known in the prior art and will not be explained in detail.

• 1 ein Ausführungsbeispiel des erfindungsgemäßen Kraftfahrzeugs sowie eine Verkehrssituation, in der ein Ausführungsbeispiel des erfindungsgemäßen Verfahrens zur Prädiktion von Fahreingriffen durchgeführt wird,
• 2 ein Ablaufdiagramm eines Ausführungsbeispiels des erfindungsgemäßen Verfahrens zur Prädiktion von Fahreingriffen, und
• 3 ein Ablaufdiagramm eines Ausführungsbeispiels des erfindungsgemäßen Verfahrens zum Training eines Algorithmus beziehungsweise das Training des im erfindungsgemäßen Verfahren zur Prädiktion von Fahreingriffen genutzten Algorithmus.

Further advantages and details of the invention are shown in the following exemplary embodiments and in the associated drawings. Here show schematically:

• 1 an exemplary embodiment of the motor vehicle according to the invention and a traffic situation in which an exemplary embodiment of the method according to the invention for predicting driving interventions is carried out,
• 2 a flowchart of an exemplary embodiment of the method according to the invention for predicting driving interventions, and
• 3 a flow chart of an exemplary embodiment of the method according to the invention for training an algorithm or the training of the algorithm used in the method according to the invention for predicting driver interventions.

1 shows a traffic situation 1 in which a method for predicting driving interventions is used in order to predict future driving interventions and thus ultimately also trajectories of road users 2, 5, 6. The method can be carried out by a processing device 7 of the road user 2 configured as a motor vehicle 3, but it would also be possible to carry out the method, for example, by an external device.

A possible implementation of the method is described below with reference to 2 illustrated flowchart explained. Here, in step S1, first the area surrounding the traffic section Participant 2 and thus the traffic situation 1 relevant environmental data recorded, which relate to other road users 5, 6, among other things. The detection can be carried out by at least one vehicle-side sensor 11 and/or by receiving corresponding information via the communication device 12, for example from the other road user 6 via his communication device 35. In the traffic situation 1 shown, the road users 2, 5 are driving one behind the other on the roadway 4 and the road user 6 is moving on a turning road in the direction of the roadway 4.

In step S2, input data 13 for an algorithm 20, which is used to predict driving interventions 21, is determined on the basis of environmental data 10. The input data can be determined in the form of a feature vector 14 which describes relevant features of the traffic situation from the point of view of road user 2 . For example, individual entries in this feature vector can describe the relative position 15 or the distance between the road user 2 and the road user 5 driving ahead and/or a relative speed to this road user. Further entries of the feature vector 14 can, for example, describe corresponding information regarding the relative position or speed to the road user 6 and/or to features 16, 17 of the road section, i.e. for example to a curve or a junction or parameters of corresponding features, for example a curve radius. It can also be expedient to take into account the relative position of the road user 2 and the preceding road user 5 with respect to the lane 18 traveled on. The entries of the feature vector 14 mentioned are to be understood as examples and feature vectors with a relatively large number of entries, for example with 30 parameters to be taken into account, are typically used.

In 2 several separate sets of input data 13 are shown schematically in step S2. This is the case because in later iterations of step S2, several predicted driving situations for road user 2 are to be taken into account, as will be explained later. The use of a cluster formation method in step S3 to find a cluster center 19 is also irrelevant in the first run of steps S2 to S9, ie in the first prediction step of the algorithm shown, and will therefore only be explained later.

In step S4, an algorithm 20 is applied to the input data 13, which on the basis of this input data 13 predicts several different driving interventions 21, by means of which, for example, different possible driver reactions in the traffic situation 1 can be taken into account. As will be explained in more detail later, algorithm 20 is trained using a machine learning method, a special training method being used to ensure that the distribution of the various driving interventions determined corresponds at least largely to a conditional probability of driving interventions in similar traffic situations 1, which is Training data sets 25 used in the training are specified. Such an algorithm can be implemented, for example, as a neural network, with an input layer or several first layers being able to be used to determine all driving interventions, and a differentiation between the different driving interventions to be generated being able to take place in that the last layer of the network or the last Layers of the network exist several times and are parameterized differently in order to generate the different driving interventions.

Since the behavior of the other road users 5, 6 is also to be taken into account in the method explained, the steps explained are repeated for the other road users 5, with the exception that the input data determined here relate to road users 5 instead of to road users 2, the same procedure is used. Thus, the procedure in step S5 corresponds to the procedure in step S2, the procedure in step S6 to the procedure in step S3 and the procedure in step S7 to the procedure in step S4. As symbolically represented by the points 24, a corresponding repetition of the determination of the driving interventions also takes place for all other relevant road users, for example also for road user 6.

In step S8, based on the driving interventions 21 predicted in steps S4, S7, etc., future traffic situations 15 are predicted for the various road users 2, 5, 6 by a kinetic model 23 according to the respective driving intervention, i.e., for example, with a predetermined acceleration and a predetermined steering angle, is parameterized in order to determine, for example, a future position and orientation of the individual road users 2, 5, 6. A respective traffic situation is thus determined for the different combinations of the determined driving interventions. In this case, the first prediction step is based on the traffic situation 1 that is described by the environmental data 10 . In later prediction steps, the traffic situation 22 predicted in the respective preceding step is modified according to the kinematic model in order to find the current traffic situation 15 .

Since, as explained, several different driving interventions 21 are determined for each of the road users 2, 5, 6, a separate traffic situation 15 is predicted for each of the combinations of driving interventions. Since a number of previously predicted traffic situations 22 are thus assumed in subsequent prediction runs, the number of predicted traffic situations 15 increases relatively quickly with an increasing number of prediction steps. By suitably selecting the time resolution of the prediction, that is to say the respective time interval for which a precalculation takes place in the respective prediction step, and the time horizon of the prediction, a prediction with an acceptable computing and memory requirement can nevertheless be achieved.

In step S9 it is checked whether further prediction steps should be carried out. In particular, the number of prediction steps already carried out can be compared with a limit value. If further prediction steps are to be carried out, the method can be repeated from step S2 or S5.

The other prediction steps essentially correspond to the procedure explained above with reference to steps S2 to S9, with the difference that the input data 13 or the feature vectors 14 are based on the predicted instead of on the basis of the environmental data 10 and thus on the basis of the actual traffic situation 1 Traffic situations 15 are determined. Since, as explained, multiple traffic situations 15 are always predicted in step S8, multiple sets of input data 13 or multiple feature vectors 14 are generated in step S2 in the subsequent prediction steps, namely one set of input data or one feature vector 14 for each of the traffic situations determined in the previous prediction step 15.

As already explained above for determining the traffic situations 15, this leads to a rapid increase in the sets of input data 13 to be taken into account as the number of prediction steps increases. One way of reducing the resulting memory and computing effort is to carry out cluster formation in the optional additional step S3 and thus, for example, to determine a fixed number of cluster centers 19 which each represent a cluster of feature vectors 14 . Corresponding methods for cluster formation are known per se and will therefore not be explained in detail. Since the training of the algorithm 20 explained below, as already explained, can ensure that the distribution of the determined driving interventions 21 and thus also the determined traffic situations 15 corresponds to the conditional probability for the development of the traffic situation specified by the training data sets 25, a Such cluster formation involves resampling of the conditional probability distribution with a lower resolution, which is possible with a sufficiently large selected number of clusters essentially without a loss of accuracy of the prediction.

If it was determined in step S9 that a sufficient number of prediction steps have already been carried out, the determined driving interventions or the trajectories resulting from sequences of these driving interventions can be evaluated in step S10 and used, for example, to activate an actuator 8 of the motor vehicle to carry out a driving intervention, for example for Carrying out an automatic acceleration or braking and/or a steering intervention, and/or to control a notification device 9 for outputting a notification to a vehicle occupant. For example, the information can be collision warnings, information about a possible or impossible overtaking of road user 5 or the like. The driving interventions or trajectories can also be used to plan an automated journey of the motor vehicle 2, during which the actuators of the vehicle are then activated.

The preceding explanation is based on the assumption that driving interventions are predicted both for road user 2 and for other road users 5, 6. However, it would also be possible, for example, to specify a trajectory for road user 2 or e.g. for one of the other road users 5, 6 and to predict driving interventions only for the remaining road users, which means that the predicted traffic situations only change with regard to the positions, orientations and speeds of these can distinguish remaining road users. As already explained in the general part of the description, this can be used, for example, to validate a planned trajectory for highly or fully automated driving or to carry out a consequence assessment, in particular with regard to the reactions of other road users to a planned trajectory, as part of trajectory planning.

A possible training of the algorithm 20 is described below with reference to FIG 3 explained in more detail. In this case, a large number of training data sets 25 are initially provided, each of which describes input data 26 for algorithm 20 and a target intervention 27 in the respective driving situation. Corresponding training data records 25 can be generated, for example, by recording the behavior of real road users and the traffic situation in which they are moving will. For example, image data can be captured by drones or infrastructure cameras or cameras of individual motor vehicles, by evaluating which a model for the traffic situation and thus also input data 26 relating to the traffic situation from the perspective of a respective road user can be determined. The target driving intervention 27 can be the respective driving intervention that was actually carried out by a driver or an automated system of the respective road user in the respective driving situation. In other words, the training data 25 depicts the behavior of a large number of real road users in a large number of traffic situations.

The algorithm 20 is parameterized by a large number of parameters 28, which can be input weights of artificial neurons of a neural network, for example. Initial values are first determined for these parameters, for example randomly, and the algorithm 20 is applied with this initial parameterization to the respective input data 26 of the training data sets in order to determine the multiple driving interventions 21 .

An error measure 31 is determined for each of the determined driving interventions 21, which relates to the deviation of the respective determined driving intervention 21 from the target driving intervention 27 of the training data set 25, for whose input data 26 the respective driving intervention 21 was determined. It was recognized here that it is advantageous first to determine a distance measure 29 which, for example, can correspond to a standard square error or another power of the error, and then to modify this distance measure 29 by applying a processing function 30 to the distance measure 29 , to determine the respective error measure 31. The processing function 30 is selected in such a way that its slope decreases as the distance increases. The use of a logarithm as processing function 30 has proven to be particularly advantageous, since the conditional probability distribution for the target driving interventions 27 specified by the training data records 25 can be mapped particularly well, i.e. training is achieved in which the various determined driving interventions 21 are training are essentially equally probable and thus cover equally large areas of the probability space.

In order to achieve such a distribution, it has also turned out to be particularly advantageous to use a "winner takes it all" approach to determine the cost function, in which for the respective training data set only the smallest error measure 32, which for this training data set is 25 im current training step was determined, is selected from the error measures 31 and taken into account. The respective smallest error measures 32 for the different training data records 25 are combined, for example added, to form the cost function 34, as symbolically represented by the curly brackets 33. The cost function 34 can then, as usual, be minimized, for example by a gradient descent method, in which the parameters 28 are varied.

In other words, the choice of the smallest error measure 32 for the cost function 34 corresponds to an exclusive consideration of that hypothesis of the algorithm 20, ie that driving intervention 21 that comes closest to the target driving intervention 27. The result of this is that the conditional probabilities for the target driving interventions 27 under the condition of the input data 26, which are specified by the training data sets 25, are modeled particularly well by the trained algorithm.

In individual cases it may be possible that the initial values of the parameters 28 lead to one or more of the generated hypotheses, that is to say the determined driving interventions 21, never or only very rarely leading to the smallest degree of error. This can result in only very slow training taking place for those parameters 28 on which this hypothesis exclusively depends. It can therefore be advantageous, at least in some of the training steps, not only to consider the selected error measure 32 in the cost function 34, but also to consider the other error measures 31, with a significantly lower weight for these, for example a factor of between 10 and 100 less weight than chosen for the selected error measure 32.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• US2020249674A1 [0007]

Claims (10)

Method for predicting driving interventions (21) for a road user (2), in particular a motor vehicle (3), and/or at least one other road user (5, 6), - environmental data (10) relating to the environment of the road user (2), which relate to the at least one other road user (5, 6) are detected, - in a first prediction step, an algorithm (20) is applied to input data (13) which depend on the environmental data (10) in order to be used as output data for and /or to predict the respective other road users (2, 5, 6) in each case several different driving interventions (21), from which different traffic situations (15) result, - the algorithm (20) being applied to the respective input data (13) in subsequent prediction steps , which depend on one of the traffic situations (15) determined in the preceding prediction step, in order to pr the various driving interventions (21) as output data which in turn result in different traffic situations (15), characterized in that an algorithm (20) which is parameterized by a number of parameters (28) is used to predict the driving interventions (21) for the respective road user (2, 5, 6). is determined by a method of machine learning, in which a plurality of training data sets (25) are used, each relating to input data (26) for the algorithm (20) and a target driving intervention (27), the method of machine learning comprising a plurality of training steps, in which - the algorithm (20) with an initial parameterization present at the beginning of the respective training step is applied to the respective input data (26) of the training data sets (25) in order to determine several different driving interventions (21) for each of the training data sets (25), - For the respective training data set (25), a respective error measure (31) for the deviation of each of the ascertained lten driving interventions (21) is determined from the respective target driving intervention (27), - the smallest of the error measures (31) determined for this training data set (25) in this training step is selected for each training data set (25), and - the parameters (28) of the Algorithm (20) depending on a cost function (34) are adapted, the cost function (34) depends exclusively on the selected error measures (32) or the selected error measures (32) in the cost function (34) are weighted more heavily than the other error measures (31). procedure after claim 1 , characterized in that the respective error measure (31) is determined by applying a processing function (30) to a distance measure (29) for the distance between the respective determined driving intervention (21) and the target driving intervention (27) specified by the respective training data set (25). the gradient of the processing function (30) decreasing at least in a value range of the processing function (30) as the value of the distance measure (29) increases. procedure after claim 1 or 2 , characterized in that the respective error measure (31) is determined as the logarithm of the or a distance measure (29) for the distance between the respective determined driving intervention (21) and the respective target driving intervention (27). procedure after claim 2 or 3 , characterized in that the driving intervention (21) and the target driving intervention (27) are each defined by a vector, the distance (29) being proportional to the 2-norm of the difference vector between driving intervention (21) and target driving intervention (25) or any power of this 2-norm. Method according to one of the preceding claims, characterized in that a cluster formation method is used in at least one of the prediction steps in order to find a cluster center (19) in each case, that a cluster of several driving interventions (21) and/or several traffic situations (15) and/or several Represents input data (13) and further processes this cluster center (19) instead of the individual driving interventions (21) and/or traffic situations (15) and/or input data (13) of the respective cluster. Method according to one of the preceding claims, characterized in that a feature vector (14) is determined from the traffic situation (15) for the individual road user (2, 5, 6), which contains at least one position (15) and/or a relative speed as entries with regard to at least one other road user (5, 6) and/or with regard to the roadway (4) and/or at least one feature (16, 17) of a preceding route section, the input data (13) of the algorithm from the feature vector (14) exist or include the feature vector (14). Method according to one of the preceding claims, characterized in that the respective traffic situation (15) is determined by a kinematic model (23), in particular a one-lane or two-lane model, for the respective road user (2, 5, 6) according to the driving intervention (21 ) is parameterized. Method according to one of the preceding claims, characterized in that depending on at least one of the determined driving interventions (21) and/or one due to the time sequence of the driving interventions (21) for the road user (2) and/or the other road user (5, 6 ) defined respective trajectory at least one actuator (8) of the road user (2) is controlled to carry out a driving intervention and / or a notification device (9) of the road user (2) is controlled to output a message. Method for training an algorithm (20), which is used to predict driving interventions (21) for a respective road user (2, 5, 6) and which is parameterized by a number of parameters (28), using a machine learning method, with a number of training data sets ( 25) are used, which each relate to input data (26) for the algorithm (20) and a target driving intervention (27), the method of machine learning comprising a number of training steps, characterized in that in the respective training step - the algorithm (20) with a the initial parameterization present at the beginning of the respective training step is applied to the respective input data (26) of the training data sets (25) in order to determine several different driving interventions (21) for each of the training data sets (25), - a respective error measure for the respective training data set (25). (31) for the deviation of each of the determined driving interventions (21) from the respective target driving input riff (27) is determined, - the smallest of the error measures (31) determined for this training data set (25) in this training step is selected for each training data set (25), and - the parameters (28) of the algorithm (20) as a function of a cost function (34) are adjusted, wherein the cost function (34) depends exclusively on the selected error measures (32) or wherein the selected error measures (32) are weighted more heavily in the cost function (34) than the other error measures (31). Motor vehicle comprising at least one sensor (11) and/or at least one communication device (12) for detecting environmental data relating to the area surrounding the motor vehicle (3) and a processing device (7), characterized in that the processing device (7) for carrying out the method after one of Claims 1 until 8th is set up when the motor vehicle (3) is used as the road user (2).

Images