Determine the orientation of an object by means of radar or by using electromagnetic interrogating radiation

Abstract

The invention relates to a method (100) for determining the spatial orientation (11b) of an object (1) from at least one measurement signal (3), the measurement signal comprising the response (3a) of the object (1) to electromagnetic interrogation radiation (2). The invention relates to a method (200) for predicting the trajectory (1a) of at least one object (1) from at least one measurement signal (3), the measurement signal comprising the response (3a) of the object (1) to electromagnetic interrogation radiation (2), in combination with the object's scalar velocity v. The invention relates to a method (300) for training a classifier (6a) and / or a regressor (6b).

Description

Technical Field

[0001] The present invention relates to a method for identifying and monitoring an object in a sensing area based on the object's response to electromagnetic interrogation radiation. Background Art

[0002] In order to enable a vehicle to move at least partially automatically in road traffic, it is necessary to sense the vehicle's environment and take countermeasures if there is a risk of collision with objects in the vehicle's environment. In order to drive safely and automatically, it is also necessary to create an environment representation and localization.

[0003] The sensing of objects by means of radar is independent of the light level and is possible, for example, even at night at greater distances without oncoming vehicles being blinded by the high beams. In addition, the distance and speed of the object are directly known from the radar data. This information is important for determining whether a collision with the object is likely to occur. However, the type of object cannot be directly identified from the radar signal. Currently, this identification is solved by calculating properties in digital signal processing.

[0004] US 8,682,821 B2 discloses the classification of radar signals by means of machine learning in the following manner: whether the radar signal is caused by the movement of a certain object or the movement of a non-human animal. This knowledge can be used to avoid false alarms when monitoring human intruders in an area or to select the correct action for avoiding a collision in at least partially automated driving. Summary of the invention

[0005] In the context of the present invention, a method for determining the spatial orientation of an object from at least one measurement signal is proposed, which contains a response of the object to electromagnetic interrogation radiation. In particular, the response can contain a reflection of the interrogation radiation.

[0006] In this method, an at least two-dimensional representation of the spatial distribution of the contributions of different locations on the object to the measurement signal is ascertained from the measurement signal. The spatial orientation of the object is ascertained from this representation by means of a classifier and / or by means of a regressor.

[0007] Here, the measurement signal can have been sensed by means of one or more detectors. For example, the measurement signal can be an overall measurement signal aggregated from raw data provided by a plurality of detectors. In addition, the measurement signal can have passed through one or more stages of signal processing before forming the representation.

[0008] For example, a classifier can divide the spatial orientation between 0° and 360°, expressed as an angle, into discrete intervals and determine a confidence measure for each input representation of the spatial distribution of these contributions, with the aid of which the representation can be assigned to each interval of the possible intervals of the orientation. A regressor can provide the most likely spatial orientation in terms of angle, without having to be constrained by discrete intervals.

[0009] For example, the electromagnetic interrogation radiation can be radar radiation. Commercially available radar sensors for monitoring a spatial region contain one or more transmitters for the radar radiation and usually an array of a plurality of receiving antennas, with the aid of which the direction from which the reflected radar radiation is incident (“direction of arrival”, DOA) can be determined without mechanically moving the array. Radar sensors generally do not output the raw data provided by the receiving antennas, but rather extract the reflections from these raw data. For each reflection, usually at least one angular position of the reflection, at least one intensity of the reflection and at least one distance to the location of the reflection are evaluated.

[0010] It is now recognized that the merging of raw data into reflections is a strong simplification, which in particular strongly abstracts the spatial extent of the detected object. The object becomes one or more discrete sources at point-like locations that respond with the intensity of the sensed reflections. In reality, however, an extended electromagnetic wavefront strikes an extended object. According to the Huygens principle, each infinitesimal location on the surface of the object is the starting point of an isotropic elementary wave, and the object's response to the electromagnetic interrogation radiation is formed by the superposition of all these elementary waves.

[0011] The spatial distribution of the contributions of different locations on the object to the measurement signal determined according to the method is therefore ultimately a restoration of the measured response of the object to the physical origin of this response.

[0012] This distribution changes when the spatial orientation of the object changes relative to the measuring assembly. Thus, the interrogation radiation can only reach those locations on the surface of the object that are in a line of sight with the transmitter without hindrance. In addition, the spatial orientation of the object also determines the angle of incidence of the interrogation radiation on the object at each location and thus the phase relationship between the elementary waves formed at different locations.

[0013] The spatial orientation of the object can therefore be ascertained from the representation of the distribution of the contributions by classification and / or regression.

[0014] For example, the representation can be a three-dimensional point cloud, which assigns to each location in three-dimensional space a physical contribution to the response that originates from the location. However, the representation can also be a two-dimensional image, for example, which is derived explicitly from the point cloud by projecting the point cloud onto a plane or by other assignment rules. It is not critical that the representation correctly reproduces, for example, the contour of the object in order to determine the spatial orientation of the object from the representation by classification and / or regression in the desired manner. It is only important that two different spatial orientations of the object are also represented in different representations.

[0015] It has also been recognized that the possibility of determining the spatial orientation of an object obtained in the described manner improves the prediction of the trajectory of a moving object based on observation with the aid of electromagnetic interrogation radiation. In the case of a large number of objects, such as vehicles, the possible upcoming directions of movement are associated with the spatial orientation of the object. For example, a vehicle can continue to roll along a driving direction predetermined by its longitudinal axis or turn from this longitudinal axis by a steering operation. The minimum radius of curvature of such a turn is predetermined by the structure of the vehicle, such as the wheelbase of the vehicle. However, for example, a vehicle cannot move exactly perpendicular to the longitudinal axis of the vehicle due to its own drive.

[0016] For example, if the sensing area in front of the vehicle to be controlled or monitored (host vehicle) is monitored by radar radiation, then the direction in which the vehicle is likely to move can be at least limited based on the spatial orientation of other vehicles detected in the sensing area. This information is important for determining whether other vehicles are likely to be on a collision course with the host vehicle.

[0017] According to the above, in a particularly advantageous embodiment, radar radiation is selected as the electromagnetic interrogation radiation. For example, in principle, it is also possible to use light that scans the sensing area within the scope of the LiDAR method (“light detection and ranging”). However, such a scanning reception of the measurement signal makes it more difficult to calculate the spatial distribution compared to a measurement using radar radiation.

[0018] As mentioned above, advantageously, the measurement signal comprises data sets concerning reflections of the electromagnetic interrogation radiation, and each of these data sets comprises:

[0019] - at least one angular position of the reflection (e.g. azimuth and / or elevation),

[0020] - at least one intensity of the reflection and

[0021] - At least one distance to the location of the reflection.

[0022] The ascertainment of the spatial distribution of the contribution can be made more precise if a processing stage of the measurement signal is used for this purpose, which has undergone a lesser degree of abstraction and simplification than the raw data. In particular, it is also possible, for example, to ascertain a representation of the spatial distribution directly from the raw data provided by one or more detectors.

[0023] In a particularly advantageous embodiment, a physical model for converting the spatial distribution of the contributions into a measurement signal is determined based on the data set about the reflections. As mentioned above, the model can, for example, contain a rule according to which contributions from different locations are superimposed to form a composite wavefront, which carries the response of the object. In addition, the model can also contain, for example, a rule according to which the wavefronts arriving at one or more detectors are processed into a measurement signal. The more detailed the model, the more precisely the spatial distribution of the contributions can be inferred from the measurement signal.

[0024] In another particularly advantageous configuration, a sinc function is established for each reflection, the peak height of which is determined by the intensity of the reflection, and the position of which is determined by the angular position of the reflection and the distance from the location of the reflection. The sum of all established sinc functions is used to obtain the representation. Behind this is the recognition that the output signal provided by the radar sensor specifically contains a sinc function in an internal processing stage, which is compressed into the mentioned information about the reflection in a subsequent processing stage. Just as the sinc function describes the superposition of elementary waves when the wavefront diffracts at the gap, the sinc function is also suitable for describing the superposition of contributions made by multiple locations of an object to a determined reflection.

[0025] In this case, the 3 dB width of the sinc function can be determined particularly advantageously as a function of the physical measurement accuracy of the measuring device which has provided the measurement signal. Thus, for example, different measurement accuracies used to determine different reflections can be weighted against each other.

[0026] In another particularly advantageous embodiment, an artificial neural network, KNN and / or auxiliary vector machine, i.e. SVM, is selected as the classifier and / or regressor. KNN and SVM are classifiers or regressors that can be trained with the aid of machine learning. This means that the classifier or regressor can be prepared for an infinite number of unknown situations using a limited number of learned representations of the spatial distribution of contributions and the associated spatial orientations of objects. A list of all possible representations of the spatial distribution of contributions cannot be given conclusively in advance, since the diversity of possible objects (e.g. vehicle types) cannot be overviewed.

[0027] In another particularly advantageous embodiment, the type of object is preclassified from the physical observation of the object. The classifier or regressor for determining the spatial orientation is then selected based on this type. Thus, for example, a truck has a completely different shape than a passenger car, and accordingly the spatial distribution of the contribution of locations on these vehicles to the measurement signal from the same angle when interrogated with the interrogation radiation also differs.

[0028] As mentioned above, the determined spatial orientation of a moving object can be used in particular to predict the future course of the movement. This applies in particular, for example, when the same electromagnetic interrogation radiation (e.g. radar radiation) is also used to determine the speed of the moving object. This speed shifts the frequency of the response relative to the frequency of the incident interrogation radiation by the Doppler frequency shift, i.e., the speed can be measured by evaluating the frequency of the response. However, the speed determined in this way is a scalar. This means that the speed only indicates how quickly the object approaches the measuring device, or how quickly the object moves away from the measuring device. Now, a plurality of different movements of the object relative to the measuring device can be compatible with the same scalar speed. By measuring the spatial orientation of the object with the aid of the previously described method, the ambiguity regarding the future trajectory of the moving object can be at least partially resolved.

[0029] The invention therefore also relates to a method for predicting the trajectory of at least one object from at least one measurement signal containing the object's response to an electromagnetic interrogation radiation in conjunction with its scalar velocity v.

[0030] In this method, the position of the object is determined from the measurement signal. The spatial orientation of the object is determined using the method described above. Starting from the position of the object, the trajectory of the object is derived from the spatial orientation of the object while taking into account the scalar velocity v.

[0031] In this context, "based on" can mean, for example, in the case of a vehicle as an object, that the possible future driving direction of the vehicle is determined by the longitudinal axis of the vehicle, which is known from the spatial orientation of the vehicle, and by possible steering maneuvers away from this axis. However, depending on the time frame for the derivation, it may also be sufficient to take into account, for example, the current longitudinal axis of the vehicle.

[0032] As already mentioned above, the scalar velocity v can be determined particularly advantageously from the Doppler frequency shift of the electromagnetic interrogation radiation.

[0033] In another particularly advantageous embodiment, the predicted trajectory of the object is compared with the current driving trajectory and / or the planned trajectory of the vehicle to be monitored and / or controlled. In this way, it is particularly possible to detect whether there is a risk that the trajectory of the observed object will be tangent to the current driving trajectory or the planned trajectory of the host vehicle. Such tangency indicates a risk of collision.

[0034] In another particularly advantageous embodiment, in response to the predicted trajectory of the object being tangential to the current driving trajectory or planned trajectory of the vehicle, a physical warning device of the vehicle that is perceptible to the driver of the vehicle is actuated. Alternatively or also in combination therewith, the steering system, drive system and / or braking system of the vehicle is actuated such that the new trajectory of the vehicle is no longer tangential to the predicted trajectory of the object.

[0035] In particular, when radar radiation is used as electromagnetic interrogation radiation, a collision risk can be recognized earlier, for example at night, than is possible for a human driver based on his visual observation in the range of action of the low beam headlights and can be reduced by the mentioned interventions.

[0036] As mentioned above, in the above-described method, in particular classifiers and / or regressors that can be trained by means of machine learning can be used. Therefore, the present invention also relates to a method for training a classifier and / or regressor for use in the above-described method.

[0037] In the method, a learning set of learning data sets is provided. These learning data sets each contain as learning input an at least two-dimensional representation of the spatial distribution of the contribution of different locations on the object to a measurement signal containing the object's response to electromagnetic interrogation radiation. In addition, the learning data sets contain as learning output the spatial orientation present in each case during the measurement.

[0038] The behavior of the classifier and / or regressor is determined by parameters. In KNN, for example, the parameters can be weights used to calculate the input of each neuron to activate the neuron.

[0039] For each learning data set, the corresponding learning input is fed to the classifier and / or regressor. The spatial orientation of the object outputted by the classifier and / or regressor is compared with the learning output associated with the learning input. The parameters of the classifier and / or regressor are optimized in such a way that the classifier and / or regressor maps the learning input contained in the learning set to the associated learning output at least with a predetermined accuracy according to the error function.

[0040] As mentioned above, in this way it is possible to prepare a classifier and / or regressor for an unlimited number of new situations based on a limited number of learned situations.

[0041] As mentioned above, advantageously, the sensor hardware for receiving the measurement signal can be changed in such a way that the sensor hardware provides an interface for deriving more detailed information about the reflection than the currently typical output. However, in principle, the method is also sufficient to cope with the information currently provided, i.e., without changing the hardware. In particular, these methods can be implemented in software, which brings direct customer benefits, i.e., the spatial orientation of the object or the future trajectory of the object can be determined more accurately than before. Therefore, the present invention also relates to a computer program, which contains machine-readable instructions, when the instructions are implemented on a computer and / or on a controller, the instructions arrange the computer and / or the controller to implement a method according to one of the described methods. Similarly, the present invention also relates to a machine-readable data carrier or download product with a computer program.

[0042] Further measures for improving the invention are shown in more detail below together with the description of preferred exemplary embodiments of the invention with reference to the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings show:

[0044] Figure 1 An exemplary superposition of different contributions 31a-31c for a response 3a to an electromagnetic interrogation radiation 2;

[0045] Figure 2 Embodiments of method 100;

[0046] Figure 3 An exemplary representation 4 of the spatial distribution of contributions 31a - 31c for different spatial orientations 11b of the same object 1 ;

[0047] Figure 4 Embodiments of method 200;

[0048] Figure 5 An exemplary application of the method 200 at an intersection of four roads 81 , 82 , 83 , 84 ;

[0049] Figure 6 Embodiment of method 300 . DETAILED DESCRIPTION

[0050] Figure 1 An exemplary situation is shown in which an extended object 1 forms a response 3a to an electromagnetic interrogation radiation 2. The measuring device 5 used comprises a transmitter 5a for the interrogation radiation 2 and a receiver 5b for the response 3a of the object 1. Figure 1In the example shown in FIG. 1 , the object 1 is a vehicle which is located at a position 11 a and occupies a spatial orientation 11 b in the form of a rotation angle with respect to an axis extending vertically through the object 1 . The measuring device 5 uses the Figure 1 The unmarked device extracts the reflections 32 from the raw data received by the receiver 5a and outputs these reflections 32 as the measurement signal 3. Specifically, for the reflections 32, at least one angular position 32a, at least one intensity 32b and at least one distance 32c from the location of the reflection 32 are sensed respectively.

[0051] The method 100 is used to evaluate the spatial orientation 11b of the object 1 from the measurement signal 3. The method 100 uses the fact that each location 12a-12c on the object 1 responds to the electromagnetic interrogation radiation 2 with its own contribution 31a-31c and that these contributions 31a-31c are superimposed to form an overall response 3a. An at least two-dimensional representation 4 of the spatial distribution of these contributions 31a-31c is evaluated from the measurement signal 3 and the sought spatial orientation 11b of the object 1 is determined therefrom by means of a classifier 6b and / or by means of a regressor 6b.

[0052] The method 100 Figure 2 In an optional step 110, a type 14 of the object 1 is first preclassified from a physical observation 13 of the object 1. In particular, the physical observation 13 can completely or partially coincide with the measurement data 3. In an optional step 120, a classifier 6a and / or a regressor 6b is selected depending on the type 14.

[0053] In step 130 , the classifier 6 a and / or the regressor 6 b is used in order to analyze the spatial orientation 11 b of the evaluation object 1 from the measurement signals 3 , 32 .

[0054] For this purpose, according to block 131 , a physical model for converting the spatial distribution of contributions 31a - 31c into measurement signal 3 can be ascertained, for example, using a data set contained in measurement signal 3 for reflections 32. The model provides a representation 4 of this distribution.

[0055] Alternatively or also in combination, according to block 132, a sinc function 32d can be generated for each reflection 32. In this case, in particular, for example, according to block 132a, the 3 dB width of the sinc function 32d can be determined with the aid of the physical measurement accuracy of the measuring device 5 that has provided the measurement signal 3. According to block 133, the sum of the sinc functions 32d generated in each case for all reflections 32 is taken into account for determining the representation 4.

[0056] Regardless of the manner in which the representation 4 is ultimately obtained, in step 140 it is fed to a classifier 6 a and / or a regressor 6 b and processed further there to form the sought spatial orientation 11 b of the object 1 .

[0057] Figure 3 By way of example, some representations 4 obtained from the measurement signals 3 for the same object 1 in a plurality of different spatial orientations 11b are schematically shown. It is noteworthy here that even with a fixed spatial orientation 11, the representations 4 determined from a plurality of successively received measurements differ slightly from one another. Nevertheless, the task of mapping all representations 4 into the correct spatial orientation 11b of the object 1 can still be solved very well with the aid of machine learning.

[0058] Figure 4 An exemplary embodiment of a method 200 is shown. In step 210, the position 11a of the object 11 is ascertained from the measurement signal 3, which contains the response 3a of the object 1 to the electromagnetic interrogation radiation 2. In step 220, the spatial orientation 11b of the object 1 is ascertained from the same measurement signal 3 with the aid of the method 100. Then, in an optional step 225, the measurement signal 3 is used to ascertain the scalar velocity v of the object 1. For this purpose, the Doppler frequency shift is evaluated relative to the electromagnetic interrogation radiation 2 originally emitted.

[0059] In step 230, starting from position 11a, a trajectory 1a of object 1 is derived from the previously determined spatial orientation 11b, taking into account the scalar velocity v. In step 240, the trajectory 1a of object 1 predicted in this way is compared with the current driving trajectory 50a and / or the planned trajectory 50b of the vehicle 50 to be monitored and / or controlled.

[0060] Specifically, in step 250, it is checked whether the predicted trajectory 1a of the object 1 is tangential to, in particular, intersects, for example, the current driving trajectory 50a and / or the planned trajectory 50b of the vehicle 50. If this is the case (true value 1), then in step 260, a physical warning device 51 of the vehicle 50 that is perceptible to the driver of the vehicle 50 is activated. Alternatively or also in combination with this, in step 270, the steering system 52, the drive system 53 and / or the brake system 54 of the vehicle 50 can be controlled in such a way that the vehicle 50 travels along a new trajectory 50c that is no longer tangential to the predicted trajectory 1a of the object 1.

[0061] exist Figure 5 An example of application of method 200 is shown in FIG. Figure 5In the illustrated scenario, the vehicle 50 to be controlled on the road 81 approaches an intersection with three other roads 82, 83 and 84. The vehicle 50 is currently traveling on the trajectory 50a. It is planned that the vehicle 50 continues its travel in a straight line along the trajectory 50b in the direction of the road 83.

[0062] The vehicle 50 monitors the sensing region 2a by means of the measuring device 5. For this purpose, the measuring device 5 emits electromagnetic interrogation radiation 2 into the sensing region 2a and receives a response 3a therefrom.

[0063] exist Figure 5 In the situation shown in , in the sensing area 2a, the other vehicle is recorded as an object 1 at the position 11a. It is also recorded that the other vehicle 1 approaches the host vehicle 50 at a scalar speed v. However, the expected intention of the other vehicle 1 cannot be determined from this. It is possible that the driver of the other vehicle 1 does not perceive the host vehicle 50 and intends to turn left into the road 82. In this case, the trajectory 1a of the other vehicle 1 will be tangential to the planned trajectory 50b of the host vehicle 50, that is, a collision may occur. However, it is also possible that the driver of the other vehicle 1 intends to continue driving straight along the trajectory 1b into the road 81. In this case, there will be no conflict with the host vehicle 50.

[0064] Within the scope of the above-described method 200 , a spatial orientation 11 b of other vehicle 1 is additionally determined from the measurement data 3 recorded by means of measuring device 5 and is used to derive trajectory 1 a of other vehicle 1 starting from its current position 11 a .

[0065] exist Figure 5 In the example shown, it is known from the spatial orientation 11b of the other vehicle 1 that this vehicle 1 is expected to follow a trajectory 1a which leads in the direction of the road 82. Therefore, countermeasures are taken to avoid a collision. Figure 5 In the example shown in , an avoidance trajectory 50 c is determined, on which the host vehicle 50 turns into the road 84 .

[0066] This example shows that it is desirable to predict the intention of other vehicles 1 with a certain degree of reliability. If the real intention of vehicle 1 is to continue driving on trajectory 1b in the direction of road 81, then avoiding on trajectory 50c is exactly an incorrect reaction of host vehicle 50 and a collision may occur.

[0067] Figure 6 An embodiment of a method 300 for training a classifier 6a or a regressor 6b is shown. In step 310, a learning set of learning data sets 7 is provided, each containing the previously described representation 4 as learning input 7a and the spatial orientation 11b of the object in the underlying measurement as learning output 7b.

[0068] In step 320, the learning input 7a is fed to the classifier 6a and / or the regressor 6b. In step 330, the spatial orientation 11b of the object 1 output by the classifier 6a and / or the regressor 6b is compared with the learning output 7b. In step 340, the parameters 6c, 6d of the classifier 6a and / or the regressor 6b are optimized so that the learning input 7a is mapped as accurately as possible to the learning output 7b over the entire learning set.

Claims

1. A method (100) for determining a spatial orientation (11b) of an object (1) from at least one measurement signal (3), the measurement signal containing a response (3a) of the object (1) to an electromagnetic interrogation radiation (2), the method comprising the following steps: - determining (130) from the measurement signal (3) an at least two-dimensional representation (4) of the spatial distribution of the contribution (31a-31c) made by different locations (12a-12c) on the object (1) to the measurement signal (3); - determining (140) the spatial orientation (11b) of the object (1) from the representation (4) by means of a classifier (6a) and / or by means of a regressor (6b), in, The measurement signal (3) comprises data sets concerning reflections (32) of the electromagnetic interrogation radiation (2), and each of the data sets comprises: - at least one angular position (32a) of said reflection (32), - at least one intensity (32b) of said reflection (32) and - at least one distance (32c) to the location of said reflection (32), in, - establishing (132) for each reflection (32) a sinc function (32d), the peak height of the sinc function being determined by the intensity (32b) of the reflection (32) and the position of the sinc function being determined by the angular position (32a) of the reflection (32) and the distance (32c) to the location of the reflection (32); and - The sum of all created sinc functions (32d) is used (133) to obtain the representation (4).

2. The method (100) according to claim 1, wherein: The radar radiation is selected as the electromagnetic interrogation radiation (2).

3. The method (100) according to claim 1 or 2, wherein: A physical model for converting the spatial distribution of the contributions (31a-31c) into the measurement signal (3) is determined (131) based on a data set regarding the reflections (32).

4. The method (100) according to claim 1 or 2, wherein: The 3 dB width of the sinc function (32d) is determined (132a) as a function of the physical measurement accuracy of the measuring device (5) which has provided the measurement signal (3).

5. The method (100) according to claim 1 or 2, wherein: An artificial neural network and / or an auxiliary vector machine is selected as a classifier (6a) and / or a regressor (6b).

6. The method (100) according to claim 1 or 2, wherein: A type (14) of the object (1) is pre-classified (110) from a physical observation (13) of the object (1), and wherein the classifier (6a) and / or the regressor (6b) is selected (120) based on the type (14).

7. A method (200) for predicting the trajectory (1a) of at least one object (1) from at least one measurement signal (3) containing a response (3a) of the object (1) to an electromagnetic interrogation radiation (2) in conjunction with a scalar velocity v of the object, the method comprising the following steps: - determining (210) the position (11a) of the object (1) from the measurement signal (3); - determining (220) the spatial orientation (11b) of the object (1) by means of a method (100) according to any one of claims 1 to 6; Starting from the position (11a) of the object, taking into account the scalar velocity v, a trajectory (1a) of the object (1) is derived (230) as a function of the spatial orientation (11b) of the object (1).

8. The method (200) of claim 7, wherein: The scalar velocity v is determined (225) from the Doppler frequency shift of the electromagnetic interrogation radiation (2).

9. The method (200) according to any one of claims 7 to 8, wherein: The predicted trajectory (1a) of the object (1) is compared (240) with a current driving trajectory (50a) and / or a planned trajectory (50b) of a vehicle (50) to be monitored and / or controlled.

10. The method according to claim 9, wherein: In response to the predicted trajectory (1a) of the object (1) being tangent (250) to the current driving trajectory (50a) or the planned trajectory (50b) of the vehicle (50), - operating (260) a physical warning device (51) of the vehicle that is perceptible to a driver of the vehicle (50); and / or - controlling (270) the steering system (52), the drive system (53) and / or the braking system (54) of the vehicle (50) in such a way that the vehicle's subsequent new trajectory (50c) is no longer tangential to the predicted trajectory (1a) of the object (1).

11. A method (300) for training a classifier (6a) and / or a regressor (6b) for use in a method (100) according to any one of claims 1 to 6, the method comprising the following steps: - providing (310) a learning set of learning data sets (7), each of which comprises: an at least two-dimensional representation (4) of the spatial distribution of the contribution (31a-31c) of different locations (12a-12c) on the object (1) to a measurement signal (3), the measurement signal comprising a response (3a) of the object (1) to an electromagnetic interrogation radiation (2), as a learning input (7a), and The spatial orientation (11b) of the object (1) when the measurement is performed as a learning output (7b); - for each learning data set (7), feeding (320) the classifier (6a) and / or the regressor (6b) with the corresponding learning input (7a), wherein: Determining the behavior of the classifier (6a) and / or the regressor (6b) by a set of parameters (6c, 6d); Comparing (330) the spatial orientation (11b) of the object (1) outputted by the classifier (6a) and / or by the regressor (6b) respectively with the learning output (7b) corresponding to the learning input (7a); Parameters (6c, 6d) of the classifier (6a) and / or the regressor (6b) are optimized (340) in such a way that the classifier (6a) and / or the regressor (6b) maps the learning input (7a) contained in the learning set to the corresponding learning output (7b) at least with a predetermined accuracy according to an error function.

12. A computer program product comprising machine-readable instructions which, when executed on a computer and / or on a controller, arrange the computer and / or the controller to implement the method (100, 200, 300) according to any one of claims 1 to 11.

Images