Method for training a neural network to predict a trajectory of a motor vehicle along a road having a predetermined course

Abstract

The invention relates to a method for training a neural network to predict a trajectory (T*) of a motor vehicle (1), according to which a deviation (Δ_T) of the trajectory (T*) predicted by the neural network from the actual trajectory (T) is characterized by computing a loss function (L2). The loss function (L2) is computed in local coordinate systems (K) associated with the actual trajectory (T). The invention also relates to a deep learning system comprising a neural network that is configured / programmed to carry out the method. The invention also relates to a computer program product and a data carrier for carrying out the method.

Description

[0001]Method for training the prediction of a trajectory of a motor vehicle along a roadway having a predetermined course using a neural network. The present invention relates to a method for training the prediction of a trajectory of a motor vehicle along a roadway having a predetermined course using a neural network. The invention further relates to a deep learning system having such a neural network, which is configured / programmed to carry out this method. Furthermore, the invention relates to a computer program product and a data carrier for executing the method. Modern vehicle assistance systems in motor vehicles enable at least partially automated driving.Corresponding driving strategies, which can be defined with the help of said assistance systems, involve calculating a trajectory along which the motor vehicle is to travel, in particular following the course of a roadway to be traveled. In order to obtain such driving strategies – both for simulation and for real driving – and to continuously improve them, it is known to use so-called deep learning systems with neural networks that can train said driving strategies. For training, a trajectory of the motor vehicle predicted by the neural network is compared with a simulated trajectory or with a trajectory actually traveled by the motor vehicle, and the resulting deviation is determined in the form of a so-called loss function. Conventional methods use the so-called L2 loss as the loss function.The deviation of the predicted trajectory from the simulated / real trajectory is determined by separately determining the X-deviation and the Y-deviation in a uniform, stationary, typically Cartesian coordinate system with predefined X- and Y-axes. This allows for different weighting of the X- and Y-deviations. The X-deviation and the Y-deviation are determined for all trajectory points of the predicted trajectory and the simulated / real trajectory in the uniform coordinate system, the so-called "world coordinate system." A disadvantage of this approach is that the course of the roadway is not taken into account when determining the loss function.It is therefore an object of the present invention, against the above background, to provide an improved method for training a neural network, in which the above-mentioned disadvantage is at least partially, preferably largely, and particularly preferably completely, eliminated. This object is achieved by the subject matter of the independent patent claims. Preferred embodiments are the subject matter of the dependent patent claims. The basic idea of ​​the invention is therefore to calculate a loss function, which characterizes the deviation of a trajectory of a motor vehicle predicted by a neural network from an actual trajectory, not in a uniform world coordinate system, but in local coordinate systems dependent on the course of the trajectory.This makes it possible to consider the course of the trajectory and thus, for example, the course of a roadway to be traveled, especially the direction of the trajectory, which generally does not extend exclusively in a straight line, when calculating the loss function. In particular, the use of local coordinate systems makes it easy to calculate the loss function component by component, with an X component in a local longitudinal direction and a Y component in a local transverse direction. The individual local coordinate systems can be aligned differently and, in particular, rotated relative to one another, depending on the course of the trajectory.In particular, the individual local coordinate systems can be positioned and aligned such that, at multiple trajectory points, particularly at each trajectory point, the loss function can be split into a loss component in the longitudinal direction of travel and a loss component in the transverse direction of travel perpendicular to the longitudinal direction of travel. This allows a different weight to be assigned to the loss component in the longitudinal direction of travel for each trajectory point than to the loss component in the transverse direction of travel. This proves to be particularly significant and advantageous because, for vehicle assistance systems, for example, knowledge of the deviation in the transverse direction of travel is of great importance when driving on roads with two or more lanes, especially if the motor vehicle is to follow a selected lane of the road or if the vehicle assistance system is to support or even automate a lane change.In this respect, it can be crucially advantageous when calculating the loss function if the Y component of the loss function in the transverse direction of travel can be assigned a higher weight than the X component in the longitudinal direction of travel. Likewise, in the case of a planned stop of the motor vehicle, for example, at a traffic light, a deviation in the X component such that the predicted trajectory is offset toward the stop position or exceeds the stop position can be assigned a higher weight than a deviation with a distance from the stop position. The actual trajectory in this sense is preferably the trajectory according to "ground truth." This can be a trajectory actually traveled by a motor vehicle or a trajectory simulated by simulation.In the course of the method according to the invention, the corresponding trajectory point of the actual trajectory can be selected as the origin for the respective local coordinate system. According to this variant, it is also proposed to define an abscissa axis or X-coordinate axis of the respective local coordinate system such that it runs along a direction of travel, i.e., along a longitudinal direction of travel at the corresponding trajectory point of the actual trajectory. Accordingly, an ordinate axis or Y-coordinate axis of the respective local coordinate system should be selected such that it runs along a transverse direction of travel at the corresponding trajectory point, i.e., perpendicular to the longitudinal direction of travel of the corresponding trajectory point and thus perpendicular to the corresponding X-coordinate axis.This allows the loss function to be calculated at the respective trajectory point with an X component along the associated local X coordinate axis, i.e., along the longitudinal direction of travel, and a Y component along the associated local Y coordinate axis, i.e., perpendicular to the longitudinal direction of travel. The method according to the invention is used to train a neural network to predict the trajectory of a motor vehicle, usually along a roadway with a predetermined course. According to the method, a loss function is calculated to characterize a deviation of the trajectory predicted by the neural network from the actual trajectory.According to the invention, the loss function is not calculated in a single, uniform world coordinate system, but for at least one trajectory point, preferably for several trajectory points, particularly preferably for all trajectory points, of the actual trajectory in local coordinate systems that depend on the course of the actual trajectory. According to a preferred embodiment of the method according to the invention, it is therefore proposed to define a local and thus individual coordinate system for calculating the deviation or loss function for each trajectory point of the actual trajectory for which the deviation from the corresponding trajectory point is to be calculated. Within the scope of the present invention, the loss function is the component-wise representation of the deviation. The loss function can be weighted component-wise as described here.In the latter case, the term "weighted loss function" is also used below. Particularly preferably, the origin of the respective local coordinate system can be placed at the associated trajectory point of the actual trajectory. Thus, the coordinate system to be used to determine the deviation or the loss function is preferably individually defined for at least one trajectory point, preferably for several trajectory points, particularly preferably for all of the trajectory points, and corresponds in each case to the local coordinate system of the actual trajectory point. Particularly advantageously, to calculate the deviation or loss function in the local coordinate system, a deviation of the associated actual trajectory point from the assigned predicted trajectory point is determined. It is understood that both the actual and the predicted trajectory are each defined as a function of time.Thus, in the method presented here, or in the calculation of the deviation or loss function, trajectory points of the predicted trajectory are compared with trajectory points of the actual trajectory points at the same time each time. In a preferred embodiment, the actual trajectory comprises a plurality of trajectory points defined in a, preferably Cartesian, world coordinate system. In this embodiment, the desired loss function can be calculated for at least one trajectory point, preferably several trajectory points, of the actual trajectory by calculating the loss function for each of these trajectory points in a local coordinate system associated with the respective trajectory point of the actual trajectory. The local coordinate system is determined by the course of the actual trajectory at the trajectory point.In a preferred embodiment, a local X component and a local Y component are calculated for the respective trajectory point to calculate the loss function. For this purpose, the X component is calculated as the deviation of the predicted trajectory point from the associated actual trajectory point along the longitudinal direction of travel, i.e., along the X coordinate axis of the local coordinate system at this trajectory point, or at least calculated as a function of this deviation. Furthermore, the Y component is calculated as the deviation of the predicted trajectory point from the associated actual trajectory point along the transverse direction of travel, i.e., along the Y coordinate axis of the local coordinate system, at this trajectory point, or at least calculated as a function of this deviation.If necessary, as explained above, for training purposes, the loss function along the local X-coordinate axis, i.e. the deviation of the X-component, can be weighted differently than along the local Y-coordinate axis, i.e. as the Y-component. In this embodiment, a transformation of the coordinates of the predicted trajectory point and the associated actual trajectory point, i.e. present at the same time, preferably takes place beforehand from the world coordinate system to the associated local coordinate system. This transformation corresponds to a rotational transformation of the world coordinate system into the local coordinate system. In another preferred embodiment, in order to calculate the X-component and the Y-component in the local coordinate system, the loss function is first calculated in the world coordinate system. This loss function is also referred to below as the global loss function.The global loss function has as its elements the deviation of the actual trajectory point from the predicted trajectory point along an x-axis of the world coordinate system and the deviation of the actual trajectory point from the predicted trajectory point along a y-axis of the world coordinate system. In this embodiment, the global loss function calculated in the world coordinate system is converted into the corresponding local coordinate system of the trajectory point of the actual trajectory using a rotation transformation, which corresponds in particular to the rotation transformation described above. Thus, the loss function is first determined as a vector with the elements in the world coordinate system, and then this "vector" is transformed into the corresponding local coordinate system of the trajectory point of the actual trajectory.The thus determined X component and the Y component of the transformed, local loss function can, as explained above, be weighted differently if necessary. Preferably, the rotational transformation can be carried out by a rotation about a rotation axis that extends perpendicular to both the X axis and the Y axis of the world coordinate system. The rotational transformation is then carried out by a rotation angle that is an intermediate angle between the X axis of the world coordinate system and the X axis of the local coordinate system at the respective trajectory point of the actual trajectory. Analogously, the rotation angle is the intermediate angle between the Y axis of the world coordinate system and the Y axis of the local coordinate system at the respective trajectory point. The rotational transformation is preferably carried out using the transformation matrix. described, where θ is the intermediate angle. The transformation matrix is ​​also referred to below as the rotation matrix. In an advantageous embodiment, the direction or course of a coordinate axis of the local coordinate system and thus the intermediate angle θ is determined by using coordinate points in the world coordinate system for the trajectory point belonging to the local coordinate axis and the time belonging to this trajectory point and a further coordinate point at a further time different from this time. The difference between these coordinate points corresponds at least approximately to the direction. In this way, the X-coordinate axis and the Y-coordinate axis of the respective local coordinate system can be determined easily and without additional information. The coordinate points can each be coordinate points of the actual trajectory or of the roadway in the world coordinate system.In this embodiment, the coordinate point at the time associated with the local coordinate system, hereinafter also referred to as the first time, has the coordinates x1, y1 in the world coordinate system. This coordinate point is hereinafter also referred to as the first coordinate point. At a time different from the first time, hereinafter also referred to as the second time, the coordinate point associated with the second time, hereinafter also referred to as the second coordinate point, has the coordinates x2, y2 in the world coordinate system. As explained above, the first and second coordinate points are advantageously either those of the actual trajectory or the roadway. dx and dy define the direction and course of the X-coordinate axis, where dx = x1 - x2; dy = y1 - y2. The intermediate angle θ thus corresponds to ^^ ^ = arctan ( ).^^ Coordinate points of the actual trajectory are particularly preferably used. This avoids incorrectly determined intermediate angles, for example when the actual trajectory is not locally parallel to the roadway, such as when changing lanes. On the other hand, this leads to simplified and therefore more efficient determination of the direction or course of the coordinate axes and consequently of the intermediate angle. The result is more precise and, at the same time, resource-efficient training of the trajectory prediction. The first coordinate point preferably corresponds to the current trajectory point of the actual trajectory in the world coordinate system, and the second coordinate point to one of the actual trajectory at the second point in time. The first point in time preferably follows the second point in time.This means that the first coordinate point, in particular the current trajectory point of the actual trajectory, follows the second coordinate point. The difference Δt between the first point in time and the second point in time is expediently sufficiently small to achieve a sufficiently accurate determination of the course of the coordinate axes of the local coordinate system and thus a sufficiently accurate determination of the intermediate angle θ. In particularly preferred embodiments, instead of determining the intermediate angle θ and the subsequent rotational transformation, the rotational transformation is carried out directly with the standardized difference between the first and the second coordinate point. The corresponding rotation matrix preferably corresponds to 1 ^^ −^^ ^ ^^ ^^ ^, ^^^^ where. Instead of a multiple trigonometric calculation, for example with arc functions of the type cos(arctan Scalar values ​​are used here. This leads to a resource-saving execution of the method and thus to increased efficiency. In addition, compared to the use of arc functions, which usually require access to tables with a finite number of values, this results in increased precision of the rotation transformation and, as a result, increased precision in the determination of the loss function. This leads to an improved training result. This improvement is more pronounced when different weightings of the loss function are applied along the X-coordinate axis and the Y-coordinate axis, i.e. different weightings of the X component and the Y component, since this results in a correspondingly increased accuracy in the assignment of the different weightings along the X-coordinate axis and the Y-coordinate axis.In another preferred embodiment, the actual trajectory can be a trajectory actually traveled by the motor vehicle or a trajectory simulated by means of simulation. The invention further relates to a deep learning system comprising at least one neural network, which in turn is configured / programmed to carry out the above-described method according to the invention. The above-explained advantages of the method according to the invention are therefore transferred to the deep learning system according to the invention. Furthermore, the invention relates to a computer program product designed to carry out the method, in particular by means of the deep learning system. The computer program product contains instructions which, when the computer program product is executed by a computer system and / or by the deep learning system, cause the computer system and / or the deep learning system to carry out the method.The computer program product is preferably stored on a memory comprising at least one non-volatile memory. The invention also comprises a computer-readable, non-volatile data carrier for executing the method. The data carrier comprises instructions which, when executed, cause a computer system and / or the deep learning system to carry out the described method. Further important features and advantages of the invention emerge from the subclaims, from the drawings, and from the associated description of the figures with reference to the drawings. It is understood that the features mentioned above and those to be explained below can be used not only in the respective combination specified, but also in other combinations or on their own, without departing from the scope of the present invention.Preferred embodiments of the invention are illustrated in the drawings and are explained in more detail in the following description, wherein like reference numerals refer to like or similar or functionally identical components. They show, in each case schematically: Fig. 1 a roadway traveled by a motor vehicle with the predicted and actual trajectory drawn therein Fig. 2 a detailed illustration of Figure 1 in the region of a specific trajectory point of the actual trajectory of the motor vehicle. The method according to the invention serves to train the prediction of a trajectory T* of a motor vehicle 1, shown as an example in Figure 1. The predicted trajectory T* usually leads, as shown in Figure 1, along a course 3 of a roadway 2. The method serves to train the prediction of the trajectory T* by means of a neural network, in particular in a deep learning system.A computer program product containing corresponding instructions and / or a computer-readable data carrier containing corresponding instructions can be used. The instructions of the computer program product and / or the data carrier, when executed, cause a computer system and / or the deep learning system to carry out the method. According to the method, a loss function L2 is calculated for various trajectory points P of the actual trajectory T in order to characterize or determine a deviation Δ_T of a trajectory T* of the motor vehicle 1 predicted by the neural network and indicated by a solid line in Figure 1 from the trajectory T actually traveled, indicated by a dashed line in Figure 1.However, the calculation is not performed – as in conventional methods – in a uniform world coordinate system W, but in local coordinate systems K, which depend on one of the trajectories T, T*, in the illustrated embodiments, on the actual trajectory T. The trajectory referred to as the "actual trajectory T" in the context of the present invention can be a trajectory actually traveled by the motor vehicle 1 or a trajectory simulated by means of simulation and expediently corresponds to the so-called "ground truth." As Figure 1 shows, the actual trajectory T of the motor vehicle 1 can therefore have a plurality of trajectory points P, of which three trajectory points P1, P2, P3 are shown as examples in the example in Figure 1, in each of which the loss function L2 is to be calculated.The trajectory points P can be defined with respect to the uniform Cartesian world coordinate system W in the example scenario. In the method according to the invention, the desired loss function L2 is calculated individually for each of the trajectory points P of the actual trajectory T. For the calculation, the deviation of the trajectory T* predicted by the neural network from the actual trajectory T is calculated for each of these trajectory points P by determining the distance a of the respective trajectory point P of the actual trajectory T from the associated trajectory point P* of the predicted trajectory T*.In the example of the three actual trajectory points P1, P2, and P3 shown, a distance a1 to the associated predicted trajectory point P1* is calculated for the actual trajectory point P1, a distance a2 to the associated predicted trajectory point P2* is calculated for the actual trajectory point P2, and a distance a3 to the associated predicted trajectory point P3* is calculated for the actual trajectory point P3. Here, both the actual trajectory T and the predicted trajectory T* are defined as a function of time, i.e., T=T(t) and T*=T*(t), respectively. Thus, when calculating the deviation or loss function L2, trajectory points P* of the predicted trajectory T* are compared with trajectory points P of the actual trajectory T at the same time, i.e. P1 = T(t1) with P1* = T*(t1), P2 = T(t2) with P2* = T*(t2), P3 = T(t3) with P3* = T*(t3).The loss function L2 is calculated in the method according to the invention using a local coordinate system K assigned to the respective trajectory point P. In the example shown, the trajectory point P1 is assigned the local coordinate system K1, the trajectory point P2 is assigned the local coordinate system K2, and the trajectory point P3 is assigned the local coordinate system K3. The local coordinate system K is defined by the course 3 of the actual trajectory T at the respective trajectory point P. Thus, as illustrated in Figure 1, the individual local coordinate systems K can differ from one another in terms of their position and orientation and can also be different from the world coordinate system W. To define the individual local coordinate systems K, the respective trajectory point P of the actual trajectory T is selected as the respective coordinate origin U of the relevant local coordinate system K in the exemplary embodiment shown.In the exemplary embodiment shown, an X-coordinate axis XK of the respective local coordinate system K is selected such that it extends from the associated trajectory point P along the direction of travel FLR of the actual trajectory T at the associated trajectory point P. In the exemplary embodiment shown, in which the actual trajectory T runs parallel to the roadway 2, the direction of travel FLR corresponds to the longitudinal direction of the roadway. In other words, the X-coordinate axis XK1 of the local coordinate system K1 extends from the trajectory point P1, which corresponds to the origin U1 of the coordinate system K1, away along a direction of travel FLR1 or the longitudinal direction of the roadway at the trajectory point P1. The X-coordinate axis XK2 of the local coordinate system K2 extends from the trajectory point P2, which corresponds to the origin U2 of the coordinate system K2, away along a direction of travel FLR2 or the longitudinal direction of the roadway at the trajectory point P2.The X-coordinate axis XK3 of the local coordinate system K3 extends from the trajectory point P3, which corresponds to the origin U3 of the coordinate system K3, along a longitudinal direction of travel FLR3 or the longitudinal direction of the road at the trajectory point P3. Accordingly, the Y-coordinate axis YK of the respective local coordinate system K is selected such that it extends from the associated trajectory point P along a direction FQR in this trajectory point P that is transverse to the associated X-coordinate axis XK and thus transverse to the associated direction of travel FLR, wherein the direction FQR is also referred to below as the transverse direction of travel FQR. In the exemplary embodiment shown, in which the actual trajectory T runs parallel to the roadway 2, the transverse direction of travel FQR runs along a transverse direction of the road that is orthogonal to the longitudinal direction of the road at the associated trajectory point P.In other words, the Y-coordinate axis YK1 of the local coordinate system K1 extends from the trajectory point P1 along a transverse direction of travel FQR1, which at the trajectory point P1 is orthogonal to the direction of travel FLR1. The Y-coordinate axis YK2 of the local coordinate system K2 extends from the trajectory point P2 along a transverse direction of travel FQR2, which at the trajectory point P2 is orthogonal to the longitudinal direction of travel FLR2. The Y-coordinate axis YK3 of the local coordinate system K3 extends from the trajectory point P3 along a transverse direction of travel FQR3, which at the trajectory point P3 is orthogonal to the longitudinal direction of travel FLR3. To calculate the loss function L2 at the respective trajectory point P, a local X-component L2_X and a local Y-component L2_Y of the loss function L2 are determined. This is explained below using Figure 2 as an example for the trajectory point P3.Figure 2 is therefore a detailed representation of Figure 1 in the area of ​​trajectory point P3. The following explanations regarding trajectory point P3 also apply mutatis mutandis to all trajectory points P for which the inventive method is to be implemented as described here. According to Figure 2, the X component L2_X is determined as the deviation ΔT_FLR of the predicted trajectory point P*3 of the predicted trajectory T* from the associated actual trajectory point P3 of the actual trajectory T along the direction of travel FLR3, i.e., along the X coordinate axis XK3 of the local coordinate system K3 at trajectory point P3.Accordingly, the Y component L2_Y is determined as the deviation ΔT_FQR of the predicted trajectory point P*3 of the predicted trajectory T* from the associated actual trajectory point P3 of the actual trajectory T along the transverse direction of travel FQR3, i.e. along the Y coordinate axis YK3 of the local coordinate system K3 at the trajectory point P3. For the concrete calculation of the X component L2_X and also the Y component L2_Y, the loss function can be calculated in the form of a vector in the world coordinate system W with the X axis XW and the Y axis YW and transferred to the local coordinate system K. In the example discussed here, a deviation ΔT_XW and ΔT_YW of the predicted trajectory T* from the actual trajectory T along the X axis XW and the Y axis YW of the world coordinate system W are first calculated at the trajectory point P3.The deviations ΔT_XW, ΔT_YW calculated in the world coordinate system W are converted into the corresponding local coordinate system K of the corresponding trajectory point P, in the example discussed here specifically, the trajectory point P3, and thus into the local coordinate system K3, using a rotation transformation DT. In the illustrated embodiment, the rotation transformation DT is performed by a rotation about an axis of rotation that extends perpendicular to both the X-axis XW and the Y-axis YW of the world coordinate system W. The rotational transformation DT is then carried out by a rotation angle which is an intermediate angle θ (see Figure 2) between the X-axis XW of the world coordinate system W and the X-coordinate axis XK of the local coordinate system K3 in the respective trajectory point P, or equivalently the intermediate angle θ between the Y-axis YW of the world coordinate system W and the Y-coordinate axis XY of the local coordinate system K in the respective trajectory point P.In Figure 2, the world coordinate system W is shown in dotted form at the origin U of the local coordinate system U visible there for a better understanding of the intermediate angle θ. The rotation transformation DT is described by the following rotation matrix: cos(θ) −sin(θ) ^ sin(θ) cos(θ) ^ This gives rise to the loss function L2 required and thus to the deviations ΔT_FLR, ΔT_QLR or the X components L2_X, L2_Y in the coordinate system K3: ΔT_FL cos(θ) −sin(θ) ^L2_X R ^ = ^ ^ ΔT_XW L2_Y ΔT_FQR = ^ sin(θ) cos(θ) ^ ^ ΔT_YW^ The distance a3 of the trajectory point P3 of the actual trajectory T to the assigned trajectory point P3* of the predicted trajectory T* is therefore calculated as ^3 = ^(ΔT_FLR. ^ + ΔT_FQR ^) The loss function L2 can thus be calculated from the two components L2_X, L2_Y, i.e. L2 = f (L2_X, L2_Y). In particular, an individual weighting G_X, G_Y of the two components L2_X, L2_Y is conceivable, for example in the form ^2 = (^ ^ ^2 ^ + ^ ^ ^2 ^ ) or, alternatively, in the form To determine the intermediate angle θ between the respective local coordinate system K and the world coordinate system W, the difference between a coordinate point WP at a point in time belonging to the associated trajectory point P, hereinafter also referred to as the first point in time, and a coordinate point WP at a further point in time different from this point in time, hereinafter also referred to as the second point in time, in the world coordinate system W is preferably used. The coordinate point WP at the first point in time is hereinafter also referred to as the first coordinate point WP1 and the coordinate point WP at the second point in time is also referred to as the second coordinate point WP2. To determine the intermediate angle θ, the difference between the first coordinate point WP1 and the second coordinate point WP2 is used, whereby the coordinate points WP are in the world coordinate system W.Preferably, the named coordinate points WP are those of the actual trajectory T. Expediently, the first coordinate point WP1 corresponds to the current / associated trajectory point P. In the example for the trajectory point P3 described above with reference to Figure 2, the first coordinate point WP1 corresponds to the trajectory point P3 and thus to the coordinate origin point U3 in the world coordinate system W. In the exemplary embodiment shown, the second point in time is preferably before the first point in time, so that the second coordinate point WP2 lies before the first coordinate point WP1 along the actual trajectory T. In the example described in more detail in Figure 3, the second coordinate point WP2 lies before the trajectory point P3 along the actual trajectory T.If the first coordinate point WP1 has the coordinates x1 along the X-axis XW and y1 along the Y-axis YW and the second coordinate point WP2 has the coordinates x2 along the X-axis XW and y2 along the Y-axis YW in the world coordinate system W, the intermediate angle θ is given by ^^ ^ = arctan ^ ^, ^^ where ^^ = ^1 − ^2 and ^^ = ^1 − ^2. Particularly preferably, a rotation matrix is ​​used for the rotation transformation DT, which uses the standardized difference between the coordinate points WP1 and WP2, where the following applies to the rotation matrix: 1 ^^ ^ ^^^^ ^^. with The deviations ΔT_FLR, ΔT_QLR described above or the X components L2_X, L2_Y in the local coordinate system K can thus be determined in a simplified manner with the corresponding coordinate points WP1 and WP2 by: Alternatively, to calculate the loss function L2 in the local coordinate system K, the coordinates of the actual trajectory point P and the associated trajectory point P* can first be transferred from the world coordinate system W to the corresponding local coordinate system K. This transfer is again carried out using the rotation transformation DT described above. The loss function L2 can then be calculated directly in the local coordinate system K with the X component L2_X as the difference between the transformed coordinates along the corresponding X coordinate axis XK and the Y component L2_Y as the difference between the transformed coordinates along the corresponding Y coordinate axis YK. In the respective example, the individual weighting of the loss function L2 in the trajectory point P in the longitudinal direction of travel FLR or perpendicular to this in the transverse direction of travel FQR is possible and, if required, also implemented.For example, when driving on multi-lane roads 2, the Y component of the loss function, i.e., L2_Y, can be assigned a higher weight G_Y than the X component in the longitudinal direction of the road, i.e., L2_X. Likewise, in the case of a planned stop of motor vehicle 1, for example, at a traffic light or when encountering an obstacle, the X component of the loss function L2_X can be assigned a higher weight G_X than the Y component L2_Y. *******.

Claims

1. A method for training the prediction of a trajectory (T*) of a motor vehicle (1) by a neural network, - according to which a loss function (L2) is determined to characterize a deviation (Δ_T) of the trajectory (T*) predicted by the neural network from the actual trajectory (T), - wherein the loss function (L2) is determined for at least one trajectory point (P), preferably for several trajectory points (P), particularly preferably for all trajectory points (P), of the actual trajectory (T) in a local coordinate system (K) associated with the at least one trajectory point (P), which depends on the course of the actual trajectory (T). 2.Method according to claim 1, characterized in that the actual trajectory point (P) and an associated trajectory point (P*) of the predicted trajectory (T*) are each transformed from a world coordinate system (W) into the local coordinate system (K1, K2, K3) by an intermediate angle θ between an X-axis (XW) of the world coordinate system (W) and an X-coordinate axis (XK) of the local coordinate system (K), wherein the loss function (L2) is determined using the actual trajectory point (P3) of the actual trajectory (T) transformed into the local coordinate system (K) and the associated trajectory point (P*) of the predicted trajectory (T*).

3. Method according to claim 1,. characterized in that, to determine the loss function (L2), it is determined in a world coordinate system (W) and then transformed into the local coordinate system (K3) by an intermediate angle θ between an X-axis (XW) of the world coordinate system (W) and an X-coordinate axis (XK) of the local coordinate system (K).

4. Method according to claim 2 or 3, characterized in that the intermediate angle θ is determined by using a first coordinate point (WP, WP1) in the world coordinate system (W) at a first time point associated with a trajectory point (P) associated with the local coordinate system (K) and a second coordinate point (WP, WP2) at a second time point deviating from the first time point, and the difference between the coordinate points (WP) is assumed to be the direction of an X-coordinate axis (XK) of the local coordinate system (K) associated with the trajectory point (P). 5.Method according to claim 4, characterized in that the rotation about the intermediate angle θ is carried out by means of a rotation transformation (DT), wherein the rotation transformation (DT) is carried out with the standardized difference between the coordinate points (WP), so that the rotation transformation (DT) is determined by the rotation matrix 1 ^^ −^^ ^ ^^ ^ ^, ^^^^ ^. where the first coordinate point (WP1) has the coordinates x1, y1 in the world coordinate system (W) and the second coordinate point (WP2) has the coordinates x2, y2 in the world coordinate system (W), and where: ^^ = ^1 − ^2, ^^ = ^1 − ^2, 6. Method according to claim 4 or 5, characterized in that the coordinate points (WP) selected are those of the actual trajectory (T), wherein the first coordinate point (WP, WP1) is the trajectory point (P) of the actual trajectory (T) associated with the local coordinate system (K).

7. Method according to one of claims 4 to 6, characterized in that the coordinate points (WP) are selected such that the second time lies before the first time.

8. Method according to one of claims 1 to 7, characterized in that, for calculating the deviation (Δ_T) or loss function (L2) in the local coordinate system (K), a distance (a) of the actual trajectory point (P3) of the actual trajectory (T) from the associated trajectory point (P*) of the predicted trajectory (T*) is determined.

9. Method according to one of claims 1 to 8, characterized in that - a coordinate origin (U) of the respective local coordinate system (K) is the trajectory point (P) to which this local coordinate system (K) is assigned. - an X-coordinate axis (XK) of the respective local coordinate system (K) extends away from the assigned trajectory point (P) in a direction of travel (FLR) of the actual trajectory (T) in the associated trajectory point (P).

10. Method according to one of claims 1 to 9, characterized in that - to calculate the loss function (L2) in the respective trajectory point (P), a local X-component (L2_X) and a local Y-component (L2_Y) of the loss function (L2) are calculated,- the X component (L2_X) is calculated as the deviation (ΔT_FLR) of the trajectory point (P) of the actual trajectory (T) from the associated trajectory point (P*) of the predicted trajectory (T) along a travel direction (FLR) of the actual trajectory (T) in the associated trajectory point (P) or is calculated as a function of this deviation (ΔT_FLR), - the Y component (L2_Y) is calculated as the deviation (ΔT_FQR) of the trajectory point (P) of the actual trajectory (T) from the associated trajectory point (P*) of the predicted trajectory (T) along a travel transverse direction (FQR) of the actual trajectory (T) in the associated trajectory point (P) or is calculated as a function of this deviation (ΔT_FQR).

11. Method according to claim 10, characterized in that for training the X component (L2_X) with an associated weighting (G_X) and the Y component (L2_Y) with an associated, Weighting (G_Y) is weighted, wherein the weightings (G_X, G_Y) are individually selected depending on a driving situation associated with the actual trajectory point (P).

12. Method according to one of the preceding claims, characterized in that the actual trajectory (T) is a trajectory actually traveled by the motor vehicle (1) or a trajectory simulated by means of simulation.

13. Deep learning system comprising at least one neural network which is configured / programmed to carry out the method according to one of the preceding claims.

14. Computer program product which contains instructions which, when the computer program product is executed by a computer system and / or by the deep learning system according to claim 13, cause the computer system and / or the deep learning system according to claim 13 to carry out the method according to one of claims 1 to 12. 15.A data carrier containing instructions which, when executed by a computer system and / or by the deep learning system according to claim 13, cause the system to execute the method according to one of claims 1 to 12. *******.

Images