DE102019114049A1 - Method for validating a driver assistance system using further generated test input data sets - Google Patents

Abstract

The invention relates to a method for validating a driver assistance system (1) of a vehicle (2), wherein the driver assistance system has at least one module (3) for modeling a process, comprising the following steps: calculating at least one first output value (21) containing a first information (22), by means of a first mapping (4) on the basis of a first input data set (23), wherein the first mapping (4) approximates the process when calculating the first output value (21), - generating a plurality of different auxiliary data sets (26) using a second map (5) and at least the first input data set (23), generating respective different test input data sets (29) using a third map (6) and the corresponding different auxiliary data sets (26) and the first information (22) as input data the third figure (6), - validation of the module (3) using the different test inputs sets (29) and the first output value (21).

Description

The invention relates to a method for validating a driver assistance system of a vehicle, wherein the driver assistance system has at least one module for modeling a process.

It is known that such a driver assistance system can be designed as an accident assistance system, which determines an accident probability on the basis of images taken with a camera. The module can be used for image recognition. In the US 2017/0316285 A1 a module is described, which is designed in the form of a neural network and can perform image recognition. Furthermore, it is known to test such a module and thus the driver assistance system using measurement data sets, ie to validate. The more measurement data records are available for validating the driver assistance system, the more reliably it can be checked whether the driver assistance system is functioning reliably, in particular does not exceed a predefined error threshold. However, collecting a large number of measurement data records requires a great deal of effort and costs.

The object of the present invention is therefore to provide a method for validating a driver assistance system of a vehicle, with which a cost of validation can be reduced.

This object is achieved by a method having the features of claim 1 and a validation system having the features of claim 11. Advantageous embodiments of the method are the subject of the dependent claims.

To solve the problem, a method for validating a driver assistance system of a vehicle is proposed, wherein the driver assistance system has at least one module for modeling a process. The procedure has the following steps. In a first step, at least one first output value having a first information is calculated by means of a first mapping and with reference to a first input data set, wherein the first mapping approximates the process when calculating the first output value. In a second step, several different auxiliary data sets are generated using a second map and at least the first input data set. A third step of the method provides for generating respective different test input data sets with the aid of a third image and based on the corresponding different auxiliary data sets and the first information as input data of the third image. In a fourth step, the module is validated using the different test input data sets and the first output value.

The numbering of steps one to four used here does not specify an order of the steps. Furthermore, a single sub-step of a single step of the four steps may be performed, followed by a single sub-step of another step of the four steps followed by another sub-step of the single step. As used herein, the terms "module" and "figure" each describe any known or later developed hardware, software, hardware or combination of hardware and software capable of operating with the particular "module" or "module" perform the associated "figure" associated functionality.

The process may be a physical process, for example an approach of the vehicle to another vehicle driving in front of the vehicle or a deceleration of the vehicle. The first input data set preferably contains sensor values detected by sensors of the vehicle. The first information may be, for example, information about whether the other vehicle is at a predetermined distance from the vehicle or not.

Based on the different test input data sets and the first output value can now be checked whether the module calculates the first output value with the first information when one of the different test input data sets is sent to an input of the module. If this is the case for a frequency that is higher than a frequency to be achieved, validation can be considered successful.

By means of the proposed method, the generated different test input data sets can be used to determine the frequency in addition to measurement data records obtained by a measurement. This can reduce the effort required to validate the module.

Before performing steps one through four, values of parameters of the first map are preferably adjusted during training of the first map such that the first map represents the first map Process with a given accuracy can approximate. For this purpose, the measurement data sets are advantageously used, which can be obtained by measurements on the vehicle using the sensors. The measurement records are divided into training input records and training output records. The respective training input data records may include, for example, the measured sensor values, such as brightness values of individual pixels of a camera of the driver assistance system. The respective training output records may include information as to whether or not the other vehicle is within a predetermined range in front of the vehicle.

If a respective training input data record is passed to an input of the first mapping, this calculates a corresponding output value. Subsequently, comparisons are made between the respective output values and the training output data sets corresponding to the corresponding training input data sets. Based on the comparisons, the values of the parameters can be adjusted.

A preferred embodiment provides that the first mapping from the first input data set additionally calculates a transformed data record and a second mapping at least from the transformed data record calculates the different auxiliary data records. Advantageously, before and / or after generating the different auxiliary data records, the first mapping is modified using a first optimization criterion. The first optimization criterion specifies that an information content of the transformed data set together with an information content of the first output value should be equal to an information content of the first input data set.

wobei der Operator L eine Länge in Bit des Argumentes ausgibt und D ein Operator ist, der bevorzugt eine DEFLATE Operation durchführt, und G eine Länge in Bit einer Implementierung, beispielsweise in Form eines Programmcodes, des Operators D ist. Der jeweilige Informationsgehalt des transformierten Datensatzes I(x), des ersten Ausgangswertes I(u) und des ersten Eingangsdatensatzes I(y) kann für eine Formulierung des ersten Optimierungskriteriums gleich der Obergrenze des Informationsgehaltes des transformierten Datensatzes I(x), des ersten Ausgangswertes I(u) beziehungsweise des ersten Eingangsdatensatzes I(y) gesetzt werden und entspricht bevorzugt näherungsweise jeweils einer Kolmogorow-Komplexität des transformierten Datensatzes K(x), des ersten Ausgangswertes K(u) bzw. des ersten Eingangsdatensatzes K(y). Das erste Optimierungskriterium kann beispielsweise folgendermaßen formuliert werden: $$L\left[D\left(y\right)\right]+G\cong L\left[D\left(x\right)\right]+L\left[D\left(u\right)\right]+G;\text{ bzw}\text{. }L\left[D\left(y\right)\right]\cong L\left[D\left(x\right)\right]+L\left[D\left(u\right)\right];$$

The information content of the transformed data set, the first output value and the first input data set can be determined according to a first variant by first applying at least one operation of a data compression algorithm to the transformed data set x, the first output value u or the first input data set y. The data compression algorithm may be, for example, the algorithm DEFLATE with which the transformed data set, the first output value and the first input data set can each be compressed losslessly. Using the data compression algorithm, an upper limit of the information content of the transformed data set I (x), the first output value I (u) and the first input data set I (y) can be estimated as follows: $$\begin{array}{ccc}I\left(x\right)\le L\left[D\left(x\right)\right]+G;& I\left(y\right)\le L\left[D\left(y\right)\right]+G;& I\left(u\right)\le L\left[D\left(u\right)\right]+G;\end{array}$$

wherein the operator L outputs a length in bits of the argument and D is an operator that preferably performs a DEFLATE operation, and G is a length in bits of an implementation, for example in the form of a program code, of the operator D. The respective information content of the transformed data set I (x), the first output value I (u) and the first input data set I (y) can be equal to the upper limit of the information content of the transformed data set I (x), the first output value I, for a formulation of the first optimization criterion (u) or the first input data set I (y) are set and preferably corresponds approximately to a Kolmogorow complexity of the transformed data set K (x), the first output value K (u) or the first input data set K (y). The first optimization criterion can be formulated, for example, as follows: $$L\left[D\left(y\right)\right]+G\cong L\left[D\left(x\right)\right]+L\left[D\left(u\right)\right]+G;\text{or}\text{,}L\left[D\left(y\right)\right]\cong L\left[D\left(x\right)\right]+L\left[D\left(u\right)\right];$$

wobei y[j]_i,u[j]_i,x[j]_i jeweils die i-te Bitposition innerhalb von y[j], u[j] beziehungsweise x[j] ist und n jeweils die entsprechende Anzahl an Bitpositionen in y[j], u[j] beziehungsweise x[j] ist und m die entsprechende Anzahl an Daten innerhalb von y, u beziehungsweise x ist. Das erste Optimierungskriterium kann entsprechend der zweiten Variante wie folgt formuliert werden: $$S_y=S_x+S_u$$

According to a second variant, it is possible, the respective upper limit of the information content of the transformed data set I (x), the first output value I (u) and the first input data I (y) using a respective total number S _x , S _u , S _y those bits with which the transformed data set, the first output value or the first input data set is reproduced. The respective total number S _x , S _u , S _{y of} the bits can be calculated, for example, in each case as follows: $$S_y={\displaystyle {\Sigma }_{j=1}^m\left({\displaystyle {\Sigma }_{i=1}^{n}y{\left[j\right]}_i}\right);}{S}_x={\displaystyle {\Sigma }_{j=1}^m\left({\displaystyle {\Sigma }_{i=1}^{n}x{\left[j\right]}_i}\right);}{S}_u={\displaystyle {\Sigma }_{j=1}^m\left({\displaystyle {\Sigma }_{i=1}^{n}u{\left[j\right]}_i}\right)}$$

where y [j] _i , u [j] _i , x [j] _{i are} each the i-th bit position within y [j], u [j], and x [j], respectively, and n are the corresponding number of bit positions in y [j], u [j] or x [j] respectively and m is the corresponding number of data within y, u or x. The first optimization criterion can be formulated according to the second variant as follows: $$S_y=S_x+S_u$$

The more the first optimization criterion, preferably according to the first and / or second variant, is fulfilled, the less information the transformed data record contains for determining the first output value.

The transformed data record is preferably distinguished by the fact that it has a lower information content than the first input data record. Preferably, the first output value can not be determined with the first information using the transformed data set.

For example, in the case where a classification is performed with the first map and the first input data set, such a classification can only be performed with a lower accuracy by means of a further map that has the transformed data record as input. This is due to the fact that the first input data set additionally has at least information, in particular the first information, which can be used for the classification compared to the transformed data record. Preferably, the first input data set additionally has at least as much information that is necessary for a successful classification compared to the transformed data record. Particularly preferably, the first input data record has all the information that is necessary for a successful classification.

A particular application of the method provides that the first input data set reproduces information of an image taken with the camera of the driver assistance system. Furthermore, it is possible that the captured image a road sign, a road and / or an environment of the road, such. B. adjacent to the road trees and houses contains. The module and the first image can be trained to recognize on the basis of the first input data set whether a traffic sign is located on the respective corresponding image. The transformed data set may, according to this example, each contain information representing a transformed image comprising all elements of the image, except those elements with which the first image and the module can recognize the traffic sign.

In the context of an advantageous development, the third mapping is modified using at least one second optimization criterion. Such a second optimization can take place before or after the second step. The second optimization criterion specifies that the third mapping should be inverse to the first mapping. If the second optimization takes place before the second step, then the third mapping calculates an exemplary test input data record based on the first output value with the first information and the transformed data record. Subsequently, the first map from the example test input data set calculates a second output value with the second information and a second transformed data record. Inverse means that the first transformed data set is equal to the second transformed data set and the first output value is equal to the second output value.

If the second optimization is carried out after the second step, the second optimization preferably takes place in the same way, except that an exemplary auxiliary data record is used instead of the transformed data record. Inverse means in this case that the exemplary auxiliary data record is equal to the second transformed data record and the first output value is equal to the second output value.

If the third mapping is changed in such a way that the second optimization criterion is satisfied, this has the advantage that the first mapping with each of the different test input data sets calculates a respective output value that is equal to the first output value. Thus, the validation of the module can be performed with a higher level of security.

According to an advantageous development, an information of a state of the first mapping is detected, which the first mapping adopts when the first mapping calculates a second output value as a function of a second input data set. This development further provides that the third image is changed depending on the state of the first image. The second input data set is preferably one of the training input data sets for adjusting the values of the parameters of the first map.

The state of the first map is preferably acquired by storing values of parameters and / or functions of the first map in a calculation of the second baseline value. The parameters and / or functions may be filtering and / or discrediting elements. For example, the entirety of the first image may be constructed as a filter, a group of multiple filters, a neural network, and / or a combination of at least one filter and a neural network. For example, a state of the first map may be detected by forming a state vector of the first map. In this case, the state vector may contain intermediate values which are needed for a calculation of the second output value. If the first mapping is a filter, the intermediate values may be values of coefficients of a Fourier transform.

The first image is preferably in the form of a neural network with neurons and connection weights between the neurons. The neural network can be mapped with software on a processor according to a first variant. In this case, the processor can be integrated in a validation system or in the driver assistance system. According to a second variant, the neural network can be imaged by means of a mechanical and / or biological structure. For example, this can be realized in that the structure comprises at least elements such as an AND gate and an inverter, these elements preferably emulate the respective function of the AND gate or inverter by a deformation of a, in particular biological, material. The deformations can be tapped off and converted into electrical signals.

The neural network has at least one input layer with neurons, it being possible to apply to each neuron of the input layer in each case a datum of the second input dataset, at least a first hidden layer with neurons and an output layer with at least one neuron. Created means that the date is electronically sent to the corresponding neuron or mechanically impressed. The output layer outputs the second output value.

If the first mapping is in the form of a neural network, the information of the state of the first mapping may be information about which of the neurons are activated when the second input data set is applied to the input of the first mapping and the network then applies the first output value calculated. In this case, a state vector of the network may be formed, which preferably has binary entries, the entries each being associated with a single neuron of the network. If the respective entry of the state vector of the network assumes the value one, then the corresponding neuron is activated in a calculation of the second output value. The respective entry accordingly has the value zero in the opposite case.

By means of the state vector of the network, it is possible inter alia to detect that information which is contained in the second input data record but does not contribute to a calculation of the second output value. For example, a first activated neuron may cause output values from other neurons of the network not to be used to calculate the second output value. The first neuron works in this case as part of a logical NAND operator that can be mapped into part of the network. The first neuron blocked in the activated state, the other neurons. A first part of data of the second input data set, with which the output values of the other blocked neurons are calculated, with a high probability makes no contribution to the calculation of the second output value. Using the state vector of the network, however, it is possible to store information about the first part of the data of the second input data set.

Advantageously, the state vector of the network includes the state of the first activated neuron and states of the other neurons blocked by the first activated neuron. In this way, information about the first part of the data of the second input data set can be stored in the state vector of the network which is no longer contained in the second output value.

In the following it will be described how the state vector of the first map, in particular of the mesh, can be used for the modification of the third map, such as an adaptation of values of parameters of the third map. For this purpose, preferably not only for the second input data set, the second output value but also for each of the above-mentioned training input data sets a corresponding second output value and preferably a corresponding transformed data set are calculated in succession using the first mapping. The transformed data sets thus calculated may have the same characteristics as the transformed data set described above.

In addition, in the calculation of the respective second output value, a respective corresponding state vector of the network is determined in the same way as the state vector of the network described above. In a special variant of the invention, it can be added to the respective second initial value corresponding transformed data set equal to the state vector corresponding to the respective second output value of the first map, in particular of the network. In general, the transformed data set can be calculated using the state vector of the first mapping.

Furthermore, for each training input data set, a respective new training input data set is formed for adapting the values of the parameters of the third mapping, referred to below as training of the third mapping. The new training input data record comprises the respective second output value and the state vector of the network corresponding to the respective second output value and preferably the transformed data record corresponding to the respective second output value. In the latter case, the state vector of the network is contained in the transformed data set.

Since the third map should be inverse to the first map, for training the third map, the respective training input record is used as a training map for the third map. In the training of the third mapping, the new training input data sets are successively applied to an input of the third mapping, and respective third output data sets are calculated using the third mapping. The respective third output data sets are compared with the training input data records corresponding to the new training input data sets. Preferably, a least squares sum may be calculated from a norm of respective differences between the respective third output data sets and the corresponding training input data sets. The least square error is preferably derived in each case according to the individual parameters of the third mapping, and a change in the value of the parameter is determined from a value of a derivative thus formed. The third map may be in the form of a neural network and the parameters may be in the form of connection weights. The alteration of the third mapping can occur both before and after the second step of the proposed method. It is possible that the third mapping is initialized by a random initialization of the parameters of the third mapping and changed by the training of the third mapping.

The different auxiliary data sets can be determined in a simple variant as follows. First, each of the above measurement data sets is converted into two individual vector pairs. The sensor values are located in a first vector of the respective vector pair and at least one corresponding target value in a second vector of the respective vector pair, for example information about whether the further vehicle is within a predetermined range in front of the vehicle or not.

Subsequently, individual values of individual elements of the respective first vector of each vector pair are successively changed and it is checked how much an output value calculated with the first image changes as a function of a change of a single value of the element of the first vector. The individual values of the elements of the first vector of a respective vector pair are preferably changed by multiplication by a known factor.

In the same way, two individual elements or several individual elements of the first vector of the respective vector pairs can be changed simultaneously and it is possible to check how strongly the respective calculated output value of the first image subsequently changes. Here, the individual elements of the first vector can be changed such that a change of a first element of the first vector is in a functional relationship with a change of a second element of the first vector.

In this way, certain elements of the first vector can be found which change the calculated output value of the first map less than other elements of the first vector. Those elements of the first vector which cause a comparatively large change in the calculated output value are preferably kept constant for the construction of the auxiliary data sets. Those elements which cause a comparatively weak change of the calculated output value are varied for a construction of several auxiliary data sets.

In order to easily generate the different test input data sets, an iteration loop is preferably run through several times. The iteration loop is formed by first passing a single auxiliary data set generated by the second map to a first input of the third map. Following this, a single test input data set calculated using the third figure, the individual auxiliary data set and the first output value is passed to an input of the first figure. The first map uses the single test input record to compute a single transformed record and another output with further information.

Then the second map uses the single transformed record to determine another auxiliary record. Subsequently, the third figure calculates another test input data record using the further auxiliary data record and the further information. The iteration loop is run through several times to optimize the first image, the second image and / or the third image using at least the first and / or second optimization criterion.

According to an advantageous variant, the first map generates a corresponding further transformed data record each time the iteration loop passes on the basis of the respective further test input data record. In addition, the respective further transformed data record is stored. In this variant, the second mapping determines the respective further auxiliary data record on the basis of at least two of the stored transformed data records. The further transformed data sets can each be determined using further state vectors of the first map, which describe a state of the first map in a respective calculation of the further initial value. The further state vectors are preferably calculated in the same way as the above-described state vector of the first map in the calculation of the second output value.

If the first mapping is modified in such a way that the first optimization criterion is met, the largest possible subspace can be formed with the aid of the stored transformed data records, the subspace representing an amount of data representing those variations of the first input data record which have as little, preferably no, influence to have a calculation of the first output value.

A development of this variant provides that the respective further transformed data record is then stored if an information content of the stored transformed data records is increased by storing the respective further transformed data record. The information content of the stored transformed data records can be increased, for example, if a sum of respective paired scalar products of two transformed data records becomes minimal. To form the sum, preferably all possible scalar products are formed from two possible transformed data sets. The possible transformed data records include the stored transformed data records and the further transformed data record.

The respective further auxiliary data record can be generated, for example, with the aid of a linear combination of at least two of the stored further transformed data records. Here, the further auxiliary data set is calculated as a sum of a first product, which has a first of the two stored further transformed data sets as a first factor and a first multiplier as a second factor, and a second product, which transformed a second of the two stored further ones Records as a first factor and a second multiplier as a second factor. Here, the two stored further transformed data sets, the first and second product and the sum are preferably vectors and the two multipliers each scalars.

By changing at least one of the two multipliers, a new additional auxiliary data record can be generated in addition to the further auxiliary data record. The change of one of the two multipliers causes the new additional auxiliary data set to be different from the further auxiliary data record. The first and the second multiplier respectively represent a first and a second parameter of the second image.

The first map preferably has a first sensitivity to a mean random change of elements of the first input data set. Furthermore, the first map preferably has a second sensitivity to a random mean change of the further test input data set which is calculated from the further auxiliary data record. The second sensitivity is preferably lower than the first sensitivity. The greater the first sensitivity to the second sensitivity, the lower is an information content of the further auxiliary data set, from which the first information can be obtained. If the second sensitivity is equal to zero, it is not possible to determine the first information from the further auxiliary data record.

In order to obtain the random mean change of the first input data set, preferably first individual individual random changes of the elements of the first input data set are determined. This can be implemented, for example, by randomly changing a single respective value of a single element of the first input data set, wherein a respective change in the individual values is preferably less than 5 percent in absolute value. Becomes a classification using the first image carried out, the respective change of the individual values is preferably in a range of -30 and 30 percent.

Preferably, all individual values of the first input data set are changed by means of a random number generator. A vector with all the individual values of the first input data set changed as entries of the vector is a modified first input data set. Preferably, a first difference vector is calculated from a difference between the first input data set and the changed first input data set.

For the modified first input data set, the first mapping can be used to calculate a changed output value. Furthermore, a second difference vector can be determined from a difference between the first output value and the changed output value. From the first and the second difference vector, a first or a second norm is formed. The first sensitivity can be calculated as a quotient with the second norm as a dividend and the first norm as a divisor.

The second sensitivity can be determined by changing a value of the first parameter of the second image. Preferably, the change of the value of the first parameter is twice as high as the maximum change of a single value of the first input data set. By changing the value of the first parameter, the further auxiliary data set is calculated differently and thus an altered further auxiliary data set is generated. From the modified further auxiliary data set, the third figure generates accordingly a modified further test input data record. Such a modified calculation of the modified further test input data set with respect to the further test input data record only causes changes in a subspace of a space formed by the different test input data records or a space formed by the randomly selected input data records.

Preferably, a third difference vector is calculated from a difference between the further test input data record and the modified further test input data record. For the modified further test input data record, a changed second output value can be calculated using the first mapping. Furthermore, a fourth difference vector can be determined from a difference between the first output value and the changed second output value. From the third and the fourth difference vector, a third or a fourth norm is formed. The second sensitivity can be calculated as a quotient with the third norm as a dividend and the fourth norm as a divisor.

The different auxiliary data sets are generated with the aid of the second mapping such that the different test input data records have a content such that the second sensitivity is lower than the first sensitivity.

By making the second sensitivity lower than the first sensitivity, it can be ensured that a change in the different auxiliary data sets has less influence on a calculation of the output values of the first image compared to a change of the first input data set.

Preferably, the second sensitivity is zero. This means that the first mapping always calculates the first output value with the first information with modified test input data sets. This represents an optimal case of the proposed method.

In a special variant of the method, it is provided that the stored further transformed data records are used for training of a generative adversarial network. Using the Generative Adversarial Network, the different auxiliary data records are generated according to this variant. The stored transformed data records are each assigned the function of a real data record during the training of the generative adversarial network.

The goal of training the Generative Adversarial Network is to create fake datasets that are similar to the real datasets. The fake records correspond to the different auxiliary records when the Generative Adversary Network training is completed. In the case of optimal training of the generative adversarial network, a discriminator can no longer distinguish a real data record from a fake data record. How to build the Generative Adversarial Network can be found in the publication "Generative Adversarial Nets" by lan J. Goodfellow; Department of Computer Science and Research, University of Montreal, Montreal, QC H3C 3J7. In this embodiment, for example, the third image may be formed such that it an auxiliary data set generated by the Generative Adversarial Network, which may have information for displaying an image, adding information, in particular the first information, for displaying an image of the traffic sign.

A preferred development provides that a generation of the different auxiliary data sets with the second image takes place at least under a boundary condition that restricts a space that can be formed with the different auxiliary data sets. The constraint may dictate that only auxiliary data sets representing a vehicle's travel at a speed of over 80 km / h or only traffic situations near road intersections be generated.

Another object of the invention is a validation system for validation of the driver assistance system. The validation system has at least one memory on which is stored a computer program which is set up to perform each step of the method according to one of the variants described above. Furthermore, the validation system has at least one interface adapted to exchange information between the module and respective inputs and outputs of the second mapping and the third mapping, wherein the second mapping and the third mapping are simulated upon execution of commands of the computer program.

• 1 eine schematische Ansicht eines Fahrzeugs mit einem Modul und eines Validierungssystems mit einer ersten, zweiten und dritten Abbildung zum Validieren des Moduls;
• 2 eine schematische Ansicht eines Datenflusses zwischen der ersten, zweiten und dritten Abbildung bei einer Erzeugung eines T estei ngangsdatensatzes;
• 3 einzelne Schritte eines Verfahrens zum Validieren des Moduls aus 1;
• 4 eine schematische Ansicht eines Datenflusses zwischen der ersten, zweiten und dritten Abbildung bei einer Erzeugung eines Testeingangsdatensatzes nach einer speziellen Variante des Verfahrens;
• 5 eine schematische Ansicht eines Datenflusses zwischen der ersten, zweiten und dritten Abbildung bei einer Erzeugung mehrerer T estei ngangsdatensätze;
• 6 eine schematische Darstellung einer Aufteilung von Messdatensätzen für ein Training der ersten Abbildung;
• 7 eine schematische Darstellung einer Erzeugung von Zustandsvektoren der ersten Abbildung;
• 8 eine schematische Darstellung von Datensätzen für ein Training der dritten Abbildung.

Further advantages, features and details of the invention will become apparent from the following description and from the figures. In this case, a reference symbol used repeatedly denotes the same component. The figures show in:

• 1 a schematic view of a vehicle with a module and a validation system with a first, second and third image for validating the module;
• 2 a schematic view of a data flow between the first, second and third mapping in generating a T estei ngangsdatensatzes;
• 3 single steps of a method to validate the module 1 ;
• 4 a schematic view of a data flow between the first, second and third mapping in a generation of a test input data set according to a specific variant of the method;
• 5 a schematic view of a data flow between the first, second and third image in a generation of multiple Tingtei ngangsdatensätze;
• 6 a schematic representation of a division of measurement data sets for a training of the first figure;
• 7 a schematic representation of a generation of state vectors of the first map;
• 8th a schematic representation of data sets for training the third figure.

1 shows a driver assistance system 1 of a vehicle 2 , The driver assistance system 1 has at least one module 3 for modeling a process. To the proposed method for validation of the driver assistance system 1 be a first , a second and a third needed. According to a first variant, the first , the second and the third in a test device 7 be arranged, the test device 7 via a communication connection 8th with the module 3 and at least one input and one output of the first connected is.

According to a second variant, not shown, the first , the second and the third inside the vehicle 2 , preferably within the module 3 be arranged. In this case, it is possible to use the proposed method for validating the driver assistance system 1 during normal operation of the vehicle 2 perform. This can be useful if the driver assistance system 1 is a changing system. For example, a camera can 11 of the driver assistance system 1 during a lifetime of the vehicle 2 change. This can also be the first change, for example, changed lens properties of the camera 11 can compensate. Due to such changes, it makes sense if the driver assistance system 1 even during normal operation of the vehicle 2 can be validated again. The driver assistance system 1 can continue an airbag 10 have, depending on one with the module 3 calculated signal can be triggered.

2 shows how individual records with the help of the first, second and third . . be calculated in an implementation of the proposed method. In a first step 31 becomes at least a first initial value 21 with the help of the first and based on a first input data set 23 calculated, the first approximates the process. The first output value 21 has at least a first piece of information 22 on.

In a second step 32 will be several different auxiliary records 26 with the help of the second and at least the first input data set 23 and preferably by a multiple variation of a value 27 a parameter 28 The second generated.

In a third step 33 become respective different test input data sets 29 with the help of the third and based on the corresponding different auxiliary data sets 26 and the first information 22 generated. In this case, preferably one of the auxiliary data sets 26 and the first information 22 at the same time for a calculation of one of the test input data sets 29 to an entrance of the third Posted. In a fourth step 34 becomes the module 3 using the different test input data records 29 and the first output value 21 validated. Validation can be done as described above.

The numbering of the steps 31 . 32 . 33 . 34 does not specify a mandatory order. Rather, steps, in particular the steps 32 . 33 and 34 be carried out in parallel. In particular, the steps 32 . 33 and 34 each divided into sub-steps. It is possible that a first step of the step 33 after a first step of the step 32 is executed and then a second sub-step of the step 32 is implemented. This is particularly advantageous when an iteration loop is formed to the different test input data sets 29 to create.

3 shows a possible sequence of the individual steps 31 . 32 . 33 and 34 of the proposed method.

4 schematically shows a data flow in an advantageous embodiment of the method, in which the first from the first input data set 23 additionally a transformed dataset 43 calculated, and the second mapping at least from the transformed data set 43 the different auxiliary data sets 26 calculated.

5 schematically shows a data flow in a further advantageous embodiment of the method in which an iteration loop is traversed. The iteration loop is formed by first using the second one generated individual auxiliary data set 51 to a first entrance 52 the third is directed. Following this, one will use the third , the individual auxiliary data set 51 and the first output value 21 calculated single test input data set 53 to an entrance 64 the first directed. The first calculated from the single test input data set 53 a single transformed record 60 and another output value 58 with another information 59 ,

Subsequently, the second determined using the single transformed record 60 another auxiliary data set 55 , Subsequently, the third calculates using the additional auxiliary data set 55 and further information 59 another test input record 63 , The iteration loop will be used to optimize the first one , The second and / or the third traverses several times using at least the above-mentioned first and / or second optimization criterion.

Particularly advantageous generates the first every iteration of the iteration loop based on the respective further test input data set 63 a corresponding further transformed data record 56 , in particular each further transformed records 56 ₁ . 56 ₂ . 56 _i and 56 _n , In this case, the respective further test input data record 63 to the entrance 64 the first picture sent.

The respective further transformed data sets become advantageous 56 ₁ . 56 ₂ . 56 _i and 56 _n in a store 57 of the driver assistance system 1 saved. The second determines the respective additional auxiliary data record 55 based on at least two of the stored further transformed records 56 ₁ . 56 ₂ . 56 _i . 56 _n , Here, the further auxiliary data set 55 from a linear combination of the stored further transformed data sets 56 ₁ . 56 ₂ . 56 _i . 56 _n be calculated. For this purpose, the second figure preferably has at least the parameter 28 with the changeable value 27 on. The value 27 of the parameter 28 may be a factor with which one of the stored further transformed records 56 ₁ . 56 ₂ . 56 _i . 56 _n is multiplied. The second preferably has several parameters for the calculation of several further auxiliary data sets when passing through the iteration loop several times.

The following describes an application of the method in which the module 3 approximates if the vehicle 2 bounces or bounces within a predetermined time on an advancing vehicle. For this purpose, the driver assistance system 1 one with the camera 11 connected evaluation unit 12 on. The evaluation unit 12 determined using with the camera 11 taken a distance 24 of the preceding vehicle and a relative speed 25 of the preceding vehicle respectively with respect to the vehicle 2 , The evaluation unit 12 sends during operation of the vehicle 2 an input that indicates the distance 24 and the relative speed 25 contains, to the module 3 , The module 3 determines an issue with an information as to whether the vehicle 2 bounces or bounces within the predetermined time on the vehicle ahead. The specified time should be equal to 3 seconds in the following.

To the module 3 Using the proposed method, at least the first input data set is obtained by means of a measurement 23 generates the distance 24 and the relative speed 25 contains. Based on the first input data set 23 determines the first the first output value 21 with the first information 22 , where the first information 22 Information about this is whether the vehicle 2 bounces or bounces within the predetermined time on the vehicle ahead.

Preferably, the first determined based on the relative speed 25 and the distance 24 a time, hereinafter referred to as collision time, within which the vehicle 2 on the preceding vehicle at a constant relative speed 25 would bounce. The collision time is preferably calculated using a quotient that measures the distance 24 as dividend and the relative speed 25 as a divisor, and converted to a first binary number, preferably of 8-bit length, unsigned. This conversion is performed such that the first binary number assumes a value of less than 10,000,000 for a collision time of less than 3 seconds and assumes a corresponding larger value for a collision time of greater than or equal to 3 seconds. The first figure sets the first output value 21 equal to the most significant bit of the first binary number.

In this example, the first output value 21 same as the first information 22 , Take the first information 22 the value 1 on, so the vehicle bounces 2 not within 3 seconds on the vehicle ahead. Accordingly, the first information 22 the value 0 in the event that an impact occurs within 3 seconds. The driver assistance system 1 processes the first information 22 preferably such that in the event that the first information 22 is zero, a pressure in a pressure vessel of the airbag 10 is set to a target pressure.

The first can be the transformed record 60 form in the form of a second binary number. The second binary number contains a value of a product resulting from the relative velocity 24 and the distance 25 is formed as factors, in the form of a third binary number and a fourth binary number, which has all bits except the most significant bit of the first binary number. This contains the transformed record 60 no information about whether the vehicle 2 within 3 seconds of the vehicle ahead.

The second generated from the transformed record 60 the individual auxiliary data set 51 by adding the third binary number of the transformed data set 60 changed and the fourth binary number is kept constant. The third generated using the first information 22 and the individual auxiliary data set 51 a fifth binary number that has the most significant bit of the first binary number as the most significant bit. The remaining bits of the fifth binary number correspond to the fourth binary number.

From the third binary number and the fifth binary number determines the third the single test input record 53 , which is another relative speed 61 and another distance 62 contains. The further distance 62 determines the third preferably in that the value of the third binary number is multiplied by the predetermined time, ie 3 seconds, and the root is formed from the product calculated therefrom. The further relative speed 61 is determined by the further distance 62 divided by the given time. Similar to the single test input record 53 is generated, the other test input data sets 63 using the additional auxiliary data sets 55 generated. The other auxiliary data sets 55 are similar to the single auxiliary data set 51 generated. By means of the further test input data sets 53 can the first be validated as to whether it determines correctly for any number of combinations of relative distances and relative speeds of the preceding vehicle, whether within 3 seconds an impact occurs or not.

The calculations described for this use case may be the first , the second and the third preferably perform after the first , the second or the third were optimized. To optimize the first , The second and / or the third to perform, it is preferable to iterate through the iteration loop several times. This will be the further test input data set 63 to the entrance 64 the first directed.

The first , the second and / or the third preferably each have parameters whose values are changed after each iteration of the iteration loop using the first and / or second optimization criterion described above. In this case, a third optimization criterion can be used, which specifies that a change in the output value 21 is minimized.

Furthermore, it can be provided that the transformed data record 60 is checked. A review of the transformed record 60 can through a visualization of the transformed dataset 60 and then entering a user. The input may include information as to whether the transformed record 60 most likely contains no information with which the initial value 21 can be determined. According to another variant, the check can be automated by means of a database and a device for comparing the visualized transformed data record 60 be done with images of the database.

Before the first step 31 of the proposed method can start, values of the parameters of the first be adapted by means of a training. The training can be done as described above. In the process, measurement data sets become 71 using at least one sensor, such as the camera 11 , of the vehicle 2 are obtained in respective corresponding training input data sets 72 and training output records 73 , as in 6 shown, split. A corresponding to a single training output record 75 the training output records 73 corresponding single training input record 74 the training input records is in 6 shown in the same line.

The respective training input records 72 may be within the scope of a possible further application of the proposed method detected intensity values of pixels of a photosensor of the camera 11 contain and the respective training output records 73 a guy in front of the vehicle 2 have vorrausfahrenden vehicle. The training output records 73 are preferably generated by a user.

Using the training input records 72 and the training output records 73 becomes the first trained to function of a presented single training input record of the training input records 72 the corresponding training output data set, for example, the type of vehicle, with a sufficient accuracy approximated. To do this, go to the individual training input records 72 respectively corresponding first output values 76 calculated and comparisons between each of the first initial values 76 and the individual corresponding training output records 73 and, based on the results of the comparisons, changed the values of the parameters of the first map.

To the third to change, can be the first output values 76 corresponding state vectors 77 be generated as in 7 shown. A single state vector of the state vectors 77 can be values of individual functions of the first Those who assume the functions when a single to a single training input record of training input records 72 corresponding output value of the first output values 76 is calculated. For a workout the third are preferred at least from the state vectors 77 and the first output values 76 new training input records 78 formed as in 8th shown. The training input records 72 be the third for training used as training output records, as the third opposite the first to be inverse. During training, the state vectors become 77 to the first entrance 52 and the first output values 76 preferably to a second input 79 the third Posted.

QUOTES INCLUDE IN THE DESCRIPTION

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

Cited patent literature

• US 2017/0316285 A1 [0002]

Claims (11)

Method for validating a driver assistance system (1) of a vehicle (2), wherein the driver assistance system has at least one module (3) for modeling a process, with the following steps: - Calculating at least a first output value (21) having a first information (22) using a first map (4) based on a first input data set (23), wherein the first map (4) in calculating the first output value (21) Process approximated, Generating a plurality of different auxiliary data sets (26) using a second map (5) and at least the first input data set (23), Generating respective different test input data sets (29) by means of a third image (6) and from the corresponding different auxiliary data sets (26) and the first information (22) as input data of the third image (6), - validation of the module (3) using the different test input data sets (29) and the first output value (21). Method according to Claim 1 , characterized in that the first mapping (4) from the first input data set (23) additionally calculates a transformed data set (43) and the second mapping (5) calculates the different auxiliary data sets (26) at least from the transformed data set (43). Method according to Claim 2 , characterized in that the first mapping (4) is modified using a first optimization criterion and the first optimization criterion specifies that an information content of the transformed data set (43; 60) together with an information content of the first output value (21) equals an information content of the first Input data set (23) should be. Method according to one of the preceding claims, characterized in that the third mapping (6) is changed using at least one second optimization criterion and the second optimization criterion specifies that the third mapping (6) should be inverse with respect to the first mapping (4). Method according to one of the preceding claims, characterized in that an information of a state of the first map (4) is detected, which the first map (4) assumes, if the first map (4) calculates a second initial value as a function of a second input data set, and the third map (6) is changed depending on the state of the first map (4). Method according to one of the preceding Claims 2 to 5 characterized in that iterating through an iteration loop by passing a single auxiliary data set (51) generated by means of said second map (5) to a first entrance (52) of said third map (6) and one using said third map (6), the individual auxiliary data set (51) and the first output value (21) is passed to an input (64) of the first map (4) and the first map (4) transforms a single one based on the single test input data set (53) Data set (60) and another output value (58) with a further information (59) calculated and the second figure (5) using the single transformed data set (60) calculated another auxiliary data set (55) and the third figure (6) using the further auxiliary data set (55) and the further information (59) calculates a further test input data record (63), and the iteration loop for optimizing the first en figure (4), the second mapping (5) and / or the third mapping (6) using at least the first and / or second optimization criterion is repeated several times. Method according to Claim 6 , characterized in that with each pass of the iteration loop, the first map (4) based on the respective further test input data set (63) generates a corresponding further transformed data set (56) and the respective further transformed data set (56) is stored and the second map (5 ) determines the respective further auxiliary data set (55) on the basis of at least two of the stored transformed data sets (56 ₁ , 56 ₂ , 56 _i , 56 _n ). Method according to Claim 7 , characterized in that the respective further transformed data set (56 _i ) is stored when an information content of the stored transformed data sets (56 ₁ , 56 ₂ , 56 _i-1 ) is increased by storing the respective further transformed data set. Method according to Claim 7 or 8th , characterized in that the stored further transformed data sets (56 ₁ , 56 ₂ , 56 _i , 56 _n ) are used for training of a generative adversarial network and using the generative adversarial network, the different auxiliary data sets (26, 55) to be generated. Method according to one of the preceding claims, characterized in that a generation of the different auxiliary data sets (26; 55) with the second image takes place at least under a boundary condition which limits a space that can be formed with the different auxiliary data sets (26, 55) , Validation system for validating a driver assistance system (1), the validation system (7) having at least one memory on which a computer program is stored, which is set up, each step of the method according to one of Claims 1 to 9 and at least one interface adapted to exchange information between the module (3) and at least respective inputs and outputs of the second map (5) and the third map (6), wherein at least the second map (5) and the third figure (6) are simulated when executing commands of the computer program.

Images