More realistic simulation of physical measurement data

Abstract

A method for simulating a point cloud P in a domain A A The simulated measurement data (2) is applied to at least one learning point cloud P contained in the domain B obtained by physical measurement L The method (100) for determining the influence of the physical properties (1a) shared by the measured data (1) in the point cloud P L and P A The measurement data (1, 2) in the simulation point cloud P represent coordinates respectively. The method comprises the steps of: A Transformed into the density distribution ρ in domain A A (110); • Use the transformation to transform the density distribution ρ A Transformed into the density distribution ρ in domain B B (120), wherein the transformation is implemented such that the density distribution ρ given directly in domain B is indistinguishable from that in domain B as the learning point cloud P L The density distribution ρ L is obtained as the density distribution ρ A The transformation result pe is obtained; • In domain B, a result point cloud P is generated that is statistically consistent with the density distribution pe B (130); Result point cloud P B It is evaluated as the A Result (140) of the effect of the simulated measured data (2) exerting the desired characteristic (1a). Method (200) for training. Data record obtained using the method (100). Trained KI module or corresponding data record. Method (300) for identifying objects (5a) and situations (5b). Related computer program.

Description

Technical Field

[0001] The present invention relates to a method for a more realistic simulation of physical measurement data, which method can be used, for example, to generate training data for a KI system in the field of autonomous driving. Background Art

[0002] In order for a vehicle to be able to move at least partially automatically in road traffic, it is necessary to detect the vehicle's surroundings and initiate countermeasures if a collision with an object in the vehicle's surroundings is imminent. For safe automated driving, it is also necessary to create a surrounding image and localization.

[0003] The detection of objects by means of radar is independent of lighting conditions and is possible, for example, even at night and at greater distances, without the oncoming traffic being blinded by the high beams. Furthermore, radar, as the only sensor, can "see through" the vehicle ahead by directing the radar waves around the vehicle due to reflections from guardrails or other objects. This makes it possible, for example, to detect the emergency braking of the vehicle ahead even before the vehicle ahead slows down.

[0004] US 8,682,821 B2 discloses that radar signals are classified by machine learning according to whether they are caused by the movement of certain objects or non-human animals. This knowledge can be used to avoid false alarms when monitoring whether there are human intrusions in the area, or can also be used to select the correct action to avoid collisions during at least partially automated driving.

[0005] Machine learning requires learning radar data of learning scenes whose content in terms of objects or animals, respectively, is known. Summary of the invention

[0006] It has been recognized that the required learning data is a scarce resource, especially in the case of object recognition from radar signals by means of machine learning. The learning data for object recognition from camera images usually include learning camera images that have been annotated (labeled) by humans according to which objects are contained at which positions in these learning camera images. Visual recognition of objects is particularly intuitive for humans, so that the requirements for assistants for the annotation of camera images are relatively low. Recognizing objects from radar signals, on the other hand, requires expert knowledge. In addition, the radar signal generated by the same object also depends on the characteristics of the antenna and sensor used, for example on the modulation mode of the sensor or on the position and manner in which the sensor is installed on the vehicle. The signal may be altered due to multipath propagation, for example in that the signal has been reflected multiple times on different surfaces (such as lanes, guardrails and / or walls). Finally, the radar signal also depends on the material. Some materials reflect radar waves with different backscatter coefficients that depend on the material and shape, while other materials are penetrated by the radar waves, which in turn may cause objects that were actually obscured to suddenly appear in the radar signal.

[0007] As a result, learning data for object recognition from radar signals is more difficult to obtain and more learning data is required than for object recognition from camera images. In order to alleviate this shortage, a method has been developed which is particularly capable of generating more realistic simulated (synthetic) radar data.

[0008] Since the propagation, absorption and reflection of radar waves follow known physical laws, the radar signal to be expected at the receiver can be calculated based on a predetermined arrangement of the transmitter, the receiver and the scene to be observed. In this way, learning radar data for scenes whose contents in terms of objects are known in advance can be generated in an automated manner.

[0009] However, these synthetic learning radar data are usually of much higher quality than radar data that can be obtained using real physical measurements. Real sensors also always impose special sensor characteristics on the signal. Thus, the radar reflections of real measurements show characteristic noise, which causes the reflections to vary in their position or not appear at all in individual measurements. This noise depends in particular on the distance of the object from the sensor. Thus, in particular near the vehicle, there is stronger noise, which may also have a characteristic spatial structure due to the layout of the radar antenna. Another characteristic radar phenomenon is ghosting, which means that objects are detected at locations where these objects do not exist at all, for example due to reflections of radar radiation on guardrails.

[0010] If radar data-based object recognition is therefore trained using such idealized synthetic learning radar data, there is a residual uncertainty as to the extent to which the mentioned disturbances and distortions, which are not a problem during learning in real driving operation, influence the result of the object recognition.

[0011] In order to reduce the residual uncertainty, a method for simulating the point cloud P in domain A is proposed. A The simulated measurement data are applied to at least one learning point cloud P contained in the domain B obtained by physical measurement. L Method to examine the effects of physical properties common to the measurement data.

[0012] In this context, the term "point cloud" quite generally includes a set of points in a vector space, which is described by the spatial coordinates of the points contained in the set. In such a point cloud there is no organized spatial structure, as is predefined, for example, in digital images by neighbor relationships between pixels. In contrast to images of fixed size, in a point cloud a variable number of points can also be predefined.

[0013] For example, radar measurement data can advantageously be presented as a point cloud. Thus, for example, the angular position relative to the receiver can be detected for a single radar reflection in combination with the determined distance to the reflecting object. Such a single measurement result can be regarded as a point in space. If, for example, one coordinate of the point is specified by the angular position and the other coordinate is specified by the distance, the point is located in two-dimensional space. In addition, newer radar or lidar devices detect a second angular coordinate for the height of the reflection. The point representing the reflection is then located in three-dimensional space. A complete radar measurement of the vehicle surroundings now includes a large number of reflections at different angular positions and, if necessary, at different heights, and the points formed in each case can be presented as a point cloud.

[0014] If other measurement variables, such as intensity or other spatial coordinates, are additionally detected during the measurement, the space in which the individual points of the point cloud are located can also include more dimensions. Regardless of the number of dimensions, the real measurement data or simulated measurement data are located in the point cloud P. L or P A where represent coordinates, respectively, as described in the two-dimensional or three-dimensional examples.

[0015] In this context, the term "domain" refers quite generally to a collection of elements, here point clouds, that have a common predefined property. That is, the domain A includes all simulated point clouds P A , these simulated point clouds contain the simulated measurement data, while domain B includes all the learned point clouds P L , these learning point clouds contain measurement data obtained through physical measurement. Here, the point cloud PA and P L The points in the θ may in particular be located in the same space, ie each individual measurement result represented as a point may comprise the same set of measurement variables, regardless of whether it is a real measurement result or a simulated measurement result.

[0016] According to what has been described previously, its influence is to be applied to the simulated point cloud P of the simulated measurement data. A The physical property can be particularly advantageously at least one disturbance or an artifact that occurs during the physical measurement of the measurement data. However, the method is not restricted to these special physical properties.

[0017] In this method, the simulation point cloud P A Transformed into the density distribution ρ in domain A A . Optionally, the density distribution can be normalized. For example, for processing by a neural network, normalization to the interval [0; 1] is advantageous. Another reasonable possibility is to normalize the density distribution in such a way that it can be interpreted directly as a probability density. To this end, it is then additionally required that the integral within the distribution is equal to 1.

[0018] Now, we transform the density distribution ρ A Transformed into the density distribution ρ in domain B B The transformation is implemented so that the density distribution ρ given directly in domain B is indistinguishable from the learned point cloud P L The density distribution ρ L is obtained as the density distribution ρ A The transformation result ρ B As will be described later, such a transformation can be obtained, for example, by means of machine learning, for example by training a “Generative Adversarial Network” GAN. However, the effect of the transformation within the framework of the method does not depend on the way in which the transformation is obtained.

[0019] In domain B, the density distribution ρ B Statistically consistent results point cloud P B This can be achieved, for example, by randomly determining the coordinates of the candidate points in the domain B and distributing them in a manner that is consistent with the density distribution ρ. B Assign a probability to each candidate point's coordinates to accept the candidate point.

[0020] The resulting point cloud P B It is evaluated as the A The result of applying the desired characteristic influence on the simulated measurement data.

[0021] An important advantage of this method is that it does not require an exact description of the physical properties whose influence should be exerted on the simulated measurement data. It is sufficient when there is a sufficiently large set of real measurement data with this physical property.

[0022] Furthermore, it is not necessary that the simulated measurement data and the measurement data obtained by physical measurement relate to the same scene. L The learning data in the form of can be any real data, such as those generated, for example, during normal driving operation of the vehicle. For learning, it is sufficient to know independent examples from the two domains A and B.

[0023] Finally, we can also use the density distribution ρ A and ρ B Statistical effects can be described at the level of, for example, additional radar reflections that appear in some individual measurements and then disappear again.

[0024] In an advantageous embodiment, the density distribution ρ in the domain A is determined A , by discretizing the domain A into voxels, where the simulated point cloud P A Each point of is explicitly assigned to exactly one voxel. In this context, voxel is a general term for "pixel", which is also to be interpreted for more than two dimensions. The resolution used to discretize the domain A does not have to be the same for all coordinate directions of the point. If, for example, angle and distance are combined in one point, the domain A can be discretized more finely with respect to angle than with respect to distance, for example. For each voxel, the density distribution p formed in this way is A Specifies the number of points that lie within this voxel.

[0025] The resolution used to discretize domain A can be selected independently of the resolution of the measurement data obtained by physical measurements available in domain B. Thus, for example, high-resolution simulation data can be mapped onto lower-resolution real data during the transformation, or vice versa.

[0026] Particularly advantageously, the density distribution is interpreted as an image in which each voxel represents a pixel with an intensity value that depends on the number of points assigned to the voxel. In this way, transformations provided for processing images can be repurposed or adapted for the method. For example, a deep neural network used for domain conversion of images (e.g. between summer images and winter images, between images of apples and oranges, or between photographs and images drawn by a certain artist) can be trained to perform transformations within the framework of the method.

[0027] Advantageously, belonging to different point clouds PA The density distribution ρ of the same measurement object A As different color channels are combined in one image. In this way, these point clouds P A can be processed simultaneously. If a deep neural network with convolution filters is used for the transformation, the correlation between the individual channels can also be taken into account by correspondingly selecting the dimensions of the convolution filters. There is no fundamental limit to the number of channels.

[0028] In this case, the different simulated point clouds PA can advantageously relate in particular to the simulation of different physical contrasts caused by the measured object. For example, simultaneous measurements using radar, lidar and ultrasound can be simulated, and the corresponding density distributions ρ A can be encoded into the red, green and blue color channels of an image in the RGB color space. Similarly, for example, the density distribution p belonging to four simultaneously simulated measurements can be A The color channels of cyan, magenta, yellow and black of an image are encoded in the CMYK color space. However, by analogy, the individual channels do not necessarily correspond to real colors. That is, any number of k measurement parameters can be encoded and processed in an artificial k-dimensional "color space".

[0029] According to what has been described above, it is particularly advantageous to select as measurement data the location of an object at least in the form of an angle and an associated distance by a radar sensor, a lidar sensor and / or an ultrasonic sensor. In the case of these types of measurements, interferences and artifacts often occur in reality that are difficult to describe analytically, so that there is a need to exert corresponding influences on the simulated measurement data. At the same time, these measurement data are also not present as ordered images, but as point clouds without ordered neighbor relationships, so that the method described with regard to the density distribution is particularly advantageous.

[0030] In a particularly advantageous embodiment, the transformation is performed using a KI module having an internal processing chain parameterized using a plurality of parameters, wherein the parameters are set such that it is not possible to distinguish in the domain B whether a given density distribution ρ was generated using the KI module or directly in the domain B as a learning point cloud P L The density distribution ρ L As described later, such a KI module can be trained, for example, by competing with a discriminator in a “generative adversarial network” GAN, which tries to compare the density distribution generated by the KI module with the density distribution directly in the domain B according to the learning point cloud P L The obtained density distribution ρ L Distinguish.

[0031] As described above, the KI module may in particular include at least one deep neural network based on convolutional filters.

[0032] The invention also relates to a method comprising a plurality of point clouds P with simulated measurement data. B A data record is formed, which is obtained using the described method, and which imposes on these simulated measurement data physically observed properties that are not contained in the simulation. As explained at the outset, the simulated measurement data contained in the data record can be used directly as training data for identifying objects or situations in the surroundings of the vehicle, without the need for manual "labeling" of the objects or situations contained. The influence of the imposed physical properties has the effect that these properties are also taken into account when training the recognition.

[0033] In this respect, the data record is a separately marketable product, the customer benefit of which is that the training of the recognition system is sufficient with only a small amount of manual work for marking the training data and includes a more comprehensive scenario concept. In particular, in the data record, for example, the weighting of the frequency of occurrence of certain scenarios can be adapted. Thus, for example, more simulated measurement data can be added to the data record in a targeted manner for rare and simultaneously dangerous traffic situations. In this way, the tendency of scenarios that occur frequently during driving to dominate the training of the recognition system and to "replace" rarely occurring scenarios can be counteracted.

[0034] As a further product, a KI module is conceivable which is used for this transformation in a trained state. The trained state of the KI module can likewise be embodied in a data record which contains the parameters of the internal processing chain. The KI module trained according to a specific transformation can then be used by the user to convert his own simulations into the domain of real measurement data and thus facilitate the use of these own simulations for training the recognition of objects or situations.

[0035] According to what has been described above, the invention also relates to a method for training a deep neural network for use in the described method. In this case, the deep neural network is incorporated as a generator into a generative adversarial network GAN. Alternating with the generator, a discriminator operating in domain B is trained, wherein the discriminator evaluates to what extent it is not possible to distinguish in domain B whether a given density distribution ρ was generated using the generator or directly in domain B as a learning point cloud P L The density distribution ρ L to be obtained.

[0036] Here, the learning point cloud P given for training L In particular, for example, it can be used in the same way as before for the simulation point cloud PA The same method as described is transformed into the density distribution ρ L , i.e., is transformed into a density distribution ρ by discretizing the domain B into voxels and analyzing the number of points per voxel. L .

[0037] Here, the obtained density distribution ρ can optionally be analyzed before training. L In this way, the fuzzy L For cases with very few points (these learned point clouds lead to a correspondingly strongly localized density distribution ρ L ) Density distribution ρ A Transformed to the empty density distribution ρ B From the perspective of the loss function used during training, the empty density distribution ρ B It can be a distribution consisting of a small number of points ρ L Nearby local minimum.

[0038] The training of the Generative Adversarial Network (GAN) is carried out in a cat-and-mouse game between the generator and the discriminator. At the beginning, the generator produces a density distribution ρ B , these density distributions are consistent with those belonging to the learning point cloud P L The density distribution ρ L It doesn’t matter much. As training continues, the density distribution ρ produced by the generator B becomes the density distribution ρ L However, at the same time, the requirement for similarity is increased by training the discriminator.

[0039] In another particularly advantageous embodiment, a method for converting the density distribution ρ in the domain B is additionally provided. B Transformed into the density distribution ρ in domain A A The second generator and the second discriminator working in domain A are trained. The second discriminator evaluates to what extent it cannot distinguish in domain A whether a given density distribution ρ is generated using the second generator or directly in domain A as a simulated point cloud P A The density distribution ρ A to be obtained.

[0040] A second generator, also trained, can be used to check the density distribution ρ provided by the first generator for plausibility and consistency. B .

[0041] For example, in another particularly advantageous embodiment, the first generator and the second generator can be additionally optimized as follows: the density distribution p in the domain A is Ais reproduced as identically as possible after applying the first generator and then the second generator; and / or the density distribution p in the domain B B After applying the second generator and then the first generator, the results are reproduced as identically as possible. This means that with the transformation from domain A to domain B, only the influence of the physical properties to be applied is added, and when transforming back from domain B to domain A, this influence is eliminated again, while the effective content reflected in the measured data does not change.

[0042] According to what has been described above, the ultimate goal pursued in terms of generating more realistic simulated measurement data is an improved recognition of objects and situations in the surroundings of a vehicle. Thus, the invention also relates to a method for recognizing at least one object and / or at least one situation in the surroundings of a vehicle, wherein the vehicle comprises at least one sensor for physically detecting measurement data from at least a part of the surroundings of the vehicle, and wherein the physically detected measurement data are classified by a classifier module according to which objects or situations are present in the surroundings of the vehicle. A KI module is selected as the classifier module and is trained or trained using training data, which training data include the data records with simulated measurement data described above.

[0043] Advantageously, in response to the recognition of at least one object or at least one situation, a physical warning device, a drive system, a steering system and / or a braking system of the vehicle, which is perceptible to the driver of the vehicle, is actuated in order to avoid a collision between the vehicle and the object and / or to adapt the speed and / or trajectory of the vehicle. Training in the manner described increases the probability that objects and situations represented in the simulated measurement data are correctly classified according to the measurement data obtained by physical measurements even during real driving operation and that a prescribed reaction is triggered.

[0044] The methods described can in particular be implemented in software, which can be used not only in conjunction with new hardware but can also be used, for example, as an add-on, update or upgrade for existing hardware for continued use and in this respect is a separately salable product with customer benefits. The invention therefore also relates to a computer program, which contains machine-readable instructions, which, when executed on a computer and / or on a control device, cause the computer and / or the control device to carry out one of the methods described. The invention also relates to a machine-readable data carrier or a download product having the computer program. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In the following, further measures for improving the invention are presented in detail together with the description of preferred embodiments of the invention based on the drawings.

[0046] in:

[0047] Figure 1 An embodiment of a method 100 is shown;

[0048] Figure 2 shows the density distribution ρ with A and ρ B An exemplary point cloud P A and P B ;

[0049] Figure 3 It shows that the point cloud P A Converted into density distribution ρ A images;

[0050] Figure 4 An embodiment of a method 200 is shown;

[0051] Figure 5 An embodiment of a method 300 is shown. DETAILED DESCRIPTION

[0052] Figure 1 An embodiment of the method 100 is shown. Figure 1 In the example shown in FIG. 1 , according to step 105, object localization by a radar sensor has been selected as measurement data 1, 2, and according to step 108, interference and artifacts formed when physically recording radar data 1 have been selected as characteristics 1a of physical measurement data 1, which should be applied to simulated measurement data 2. The physical measurement measurement data 1 from which the characteristics 1a can be learned form the learning point cloud P L and is located in domain B. The simulated measurement data 2 forms a simulated point cloud P A and is located in domain A.

[0053] According to step 110, the simulation point cloud P A Transformed into the density distribution ρ in domain A A To this end, according to step 112, the domain A is discretized into voxels, and according to step 114, the simulated point cloud P A Each point of is explicitly assigned to exactly one voxel. Optionally, according to step 116, the density distribution ρ A The image can be interpreted as an image, wherein in particular according to step 118, the plurality of simulated point clouds P can be A The density distribution ρ A are combined in one image. To illustrate, Figure 1 Density distribution of lieutenant generals ρ ADepicted as an image.

[0054] According to step 120, the density distribution ρ A Transformed into the density distribution ρ in domain B B According to step 122, the transformation is performed using a KI module that is trained to produce the following density distribution p B , the density distribution ρ B It cannot be compared with the learned point cloud P L The obtained density distribution ρ L The density distribution in domain B is ρ B The same is illustrated as an image.

[0055] According to the density distribution ρ B To obtain the resulting point cloud P in domain B B The original simulated measurement data 2 are now present in a modified version 2'. The modification is due to the influence of the characteristic 1a, which is now present in the modified measurement data 2'. Thus, according to step 140, the resulting point cloud P B It is evaluated as the A The simulated measurement data 2 shows the result of the influence of characteristic 1a.

[0056] Figure 2 The point cloud P composed of simulated radar data is shown. A Converted into the result point cloud P B These reflections are plotted relative to Figure 2 The angle of the direction of travel of the vehicle not drawn in A function of the distance in the direction of that angle.

[0057] Figure 2 a shows the simulated radar reflection. Figure 2 b shows the density distribution ρ of the domain A A . Figure 2 c shows the density distribution p transformed into domain B B .

[0058] according to Figure 2 The density distribution in c B The diagram of , already gives us an idea of ​​the changes that the measured data undergo in real physical measurements. In general, the radar reflections appear significantly more blurred. In addition, especially in the vicinity of the vehicle, i.e. at shorter distances, the density distribution ρ B Appears in the density distribution ρ A However, when the distance is large, it still exists in the density distribution ρ A This component in the density distribution ρ B disappeared.

[0059] Especially in the density distribution ρ B The additional components appearing in the simulation are based only on the original point cloud P A It is difficult or impossible to simulate them at all, because these components originally come from the radar sensor and not from the real conditions in the vehicle's surroundings, which are part of the simulation.

[0060] Even more obviously, in Figure 2 The resulting point cloud P in d B The differences mentioned appear in the illustrations of Figure 2 Compared to a, not only does each radar reflection appear blurred in itself, but the associated larger reflection is also split into a number of smaller reflections. In addition, additional reflections occur at many locations, with this behavior depending largely on the distance d from the vehicle. The smaller this distance d, the more additional reflections occur.

[0061] Figure 3 The simulation point cloud P A Used to distribute the density ρ A Ways to interpret as images. Figure 3 a shows the simulated point cloud P in the two-dimensional domain A A . Simulation point cloud P A The points are defined by the values ​​of two coordinates x1 and x2.

[0062] According to step 112, the domain A is discretized into voxels, and according to step 114, the simulated point cloud P A Each point of is assigned to exactly one voxel. Figure 3 b shows the result, where the voxels are presented here as squares.

[0063] Figure 3 c shows the discretized density distribution ρ at the voxel level A Here, for each voxel, the number of points assigned to that voxel is added.

[0064] According to step 116, the density distribution ρ A is interpreted as an image, where the number of points assigned to each voxel is added with the intensity value of the pixel corresponding to it. A The result is a grayscale image that can be processed through the deep neural network originally intended for images. Figure 3 In FIG. , only four different grayscales are shown by way of example by means of different shadings.

[0065] Figure 4An embodiment of the method 200 is shown. According to step 210, a deep neural network for transforming from domain A to domain B is incorporated into a generative adversarial network GAN as a generator. In step 220, the generator is trained. Furthermore, in step 230, a discriminator operating in domain B is trained. The discriminator evaluates the density distribution ρ generated by the generator B Compared with directly learning point cloud P in domain B L The resulting density distribution ρ L The training 220 of the generator and the training 230 of the discriminator are performed in a continuous alternation, that is, each iteration step contributes not only to the training 220 of the generator but also to the training 230 of the discriminator.

[0066] According to step 250, a second generator is trained for inverse transformation into domain A. Additionally, according to step 260, a second discriminator operating in domain A is trained. According to step 270, the second discriminator evaluates whether the density distribution in domain A generated by the second generator is different from the density distribution directly in domain A according to the simulated point cloud P A The resulting density distribution ρ A The training 250 of the second generator and the training 260 of the second discriminator are also performed in a continuous alternation, that is, each iteration step related thereto contributes not only to the training 250 of the second generator but also to the training 260 of the second discriminator.

[0067] According to step 280, the two generators are additionally optimized in such a way that the given density distribution is reproduced as identically as possible after transformation into the respective other domain and after inverse transformation into the original domain. If this is not the case to a sufficient extent (true value 0), then branching is carried out back to one of the previous trainings 220, 230, 250 or 260. If, however, the self-consistency condition is fully satisfied (true value 1), then the training is terminated.

[0068] Of course, after each training 220, 230 or 250, further checks can also be provided as to whether the training has brought sufficient success. For the sake of clarity, these further checks are described in Figure 4 Not drawn in.

[0069] Figure 5An exemplary embodiment of a method 300 is shown. The vehicle 4 on which the method 300 is applied has a sensor for physically detecting measurement data 42a from at least a portion 41a of the vehicle surroundings 41. In step 330, a classification of which objects 5a or situations 5b are present in the measurement data 42a is performed by a classifier 43. According to step 320, the KI module previously trained in step 310 with training data 43a has been selected as the classifier 43.

[0070] In response to the identification of an object 5a or a situation 5b, in accordance with step 340, the warning device 44a, drive system 44b, steering system 44c and / or braking system 44d of the vehicle 4 are operated to avoid a collision with the object 5a or to avoid disadvantages to the driver of the vehicle 4 or to other traffic members in the situation 5b.

Claims

1. A method for simulating a point cloud P in a domain A A The simulated measurement data (2) is applied to at least one learning point cloud P contained in the domain B obtained by physical measurement L The method (100) for determining the influence of the physical properties (1a) shared by the measured data (1) in the point cloud P L and P A The measurement data (1, 2) in represent coordinates respectively, and the method has the following steps: • The simulation point cloud P A Transformed into the density distribution ρ in the domain A A , by discretizing the domain A into voxels, where the simulated point cloud P A Each point of is explicitly assigned to exactly one voxel; • Use the transformation to convert the density distribution ρ A Transformed into the density distribution ρ in the domain B B , wherein the transformation is performed using a KI module, and wherein the transformation is implemented so that the density distribution ρ given cannot be distinguished in the domain B as a learning point cloud P directly in the domain B. L The density distribution ρ L is obtained as the density distribution ρ A The transformation result ρ B to be obtained; • In the domain B according to the density distribution ρ B Get the result point cloud P B ; • The resulting point cloud P B is evaluated as the point cloud P in the simulation A The simulated measurement data (2) are the result of applying the influence of the physical characteristics (1a).

2. The method (100) according to claim 1, wherein the density distribution ρ A is interpreted as an image in which each voxel represents a pixel with an intensity value that depends on the number of points assigned to the voxel.

3. The method (100) according to claim 2, wherein the points belonging to different point clouds P A The density distribution ρ of the same measurement object A As different color channels are combined in the image.

4. The method (100) according to claim 3, wherein different simulated point clouds P A It involves the simulation of different physical contrasts caused by the measurement object.

5. The method (100) according to claim 1, wherein the object positioning performed by a radar sensor, by a lidar sensor and / or by an ultrasonic sensor at least in the form of an angle and an associated distance is selected as the measurement data (1, 2).

6. The method (100) according to claim 1, wherein at least one disturbance or artifact which occurs during the physical measurement of the measurement data (1) is selected as a characteristic (1a) which is to be applied to the simulated point cloud P of the simulated measurement data (2). A .

7. The method (100) according to any one of claims 1 to 4, wherein the KI module has an internal processing chain parameterized with a plurality of parameters, wherein the parameters are set such that it is not possible to distinguish in the domain B whether a given density distribution ρ is generated with the KI module or directly in the domain B as a learning point cloud P L The density distribution ρ L to be obtained.

8. The method (100) according to claim 7, wherein a KI module (124) is selected, the KI module comprising at least one deep neural network based on convolutional filters.

9. A point cloud P having a plurality of simulated measurement data (2) B A data record is formed, which imparts to the simulated measurement data physically observed properties (1a) which are not contained in the simulation, the data record being obtained using a method (100) according to any one of claims 1 to 8.

10. A KI module trained to perform the method according to claim 7 or 8.

11. A data record containing parameters of an internal processing chain of a KI module according to claim 10.

12. A method (200) for training a deep neural network for use in a method according to claim 8, wherein the deep neural network is incorporated as a generator in a generative adversarial network (GAN), wherein alternately with the generator, a discriminator operating in a domain B is trained, and wherein the discriminator assesses to what extent it is not possible to distinguish in the domain B whether a given density distribution ρ was generated using the generator or directly in the domain B as a learning point cloud P L The density distribution ρ L to be obtained.

13. The method (200) according to claim 12, wherein additionally a method for converting the density distribution ρ in the domain B B Transformed into the density distribution ρ in domain A A A second generator and a second discriminator operating in the domain A are trained, wherein the second discriminator evaluates to what extent it is not possible to distinguish in the domain A whether a given density distribution ρ is generated using the second generator or directly in the domain A as a simulated point cloud P A The density distribution ρ A to be obtained.

14. The method (200) according to claim 13, wherein the first generator and the second generator are additionally optimized as follows: the density distribution p in the domain A A is reproduced as identically as possible after applying the first generator and immediately after applying the second generator; and / or the density distribution p in the domain B B After applying the second generator and immediately after applying the first generator, it is reproduced as identically as possible.

15. A method (300) for identifying at least one object (5a) and / or at least one situation (5b) in the surroundings (41) of a vehicle (4), wherein the vehicle (4) comprises at least one sensor (42) for physically detecting measurement data (42a) from at least a part (41a) of the surroundings (41) of the vehicle (4), and wherein the physically detected measurement data (42a) are classified by a classifier module (43) according to which objects (5a) or situations (5b) are present in the surroundings (41) of the vehicle (4), wherein a KI module is selected as the classifier module (43) and is trained or trained using training data (43a), the training data comprising a data record according to claim 9.

16. The method (300) according to claim 15, wherein, in response to the recognition of at least one object (5a) or at least one situation (5b), in order to avoid a collision between the vehicle (4) and the object (5a), to avoid disadvantages in the situation (5b) and / or to adapt the speed and / or trajectory of the vehicle (4), a physical warning device (44a), a drive system (44b), a steering system (44c) and / or a braking system (44d) of the vehicle (4) that is perceptible to the driver of the vehicle (4) is controlled.

17. A computer program product, comprising a computer program having machine-readable instructions, which, when executed on a computer and / or on a control device, cause the computer and / or the control device to implement the method (100, 200, 300) according to any one of claims 1 to 8 or 12 to 16.

Images