Patent Application
20220138575
The present invention relates to a computer-implemented method for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle.
The present invention further relates to a computer-implemented method for providing a trained, artificial neural network for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle.
The invention also further relates to a test unit for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle. The present invention further relates to a computer program and a computer-readable data carrier.
Driver assistance systems such as adaptive cruise control and/or highly automated driving functions can be verified and validated using a variety of verification methods. In particular, hardware-in-the-loop methods, software-in-the-loop methods, simulations and/or test drives can be used.
The overhead, in particular the time and/or cost, for testing such vehicle functions using the above-mentioned verification methods is typically very high, since a large number of potentially possible driving situations must be tested.
This can lead in particular to a high expenditure for test drives as well as for simulations. DE 10 2017 200 180 A1 specifies a method for verifying and/or validating a vehicle function, which is intended to autonomously guide a vehicle in the longitudinal and/or transverse direction.
Based on environmental data relating to an environment of the vehicle, the method includes determining a test control instruction of the vehicle function to an actuator of the vehicle, wherein the test control instruction is not implemented by the actuator.
Based on environmental data and using a road user model with respect to at least one road user in the vicinity of the vehicle, the method further includes simulating a fictitious traffic situation that would be present if the test control instruction had been implemented.
The method further comprises the provision of test data relating to the fictitious traffic situation. The vehicle function is operated passively in the vehicle to determine the test control instruction.
A disadvantage of this method is that the actual operation of the vehicle is required for the verification and/or validation of the vehicle function in order to be able to determine the required data.
Accordingly, there is a need to improve existing methods and test equipment such that so-called test cases can be efficiently determined in scenario-based testing of systems and system components in highly automated driving.
It is therefore an object of the invention to provide a method, a test unit, a computer program and a computer-readable data carrier, which can determine test cases in the context of a scenario-based testing for systems and system components in highly automated driving in an efficient manner.
The object is achieved by a computer-implemented method for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle, a computer-implemented method for providing a trained, artificial neural network for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle, a test unit for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle, a computer program and a computer-readable data carrier.
The invention relates to a computer-implemented method for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle.
The method comprises receiving a first data set comprising a plurality of driving situation parameters formed of environment parameters describing the environment of the motor vehicle and EGO parameters describing the state of the motor vehicle, wherein each driving situation parameter has a predetermined domain of definition.
The method further comprises approximating a numerical value range of a target function for the domain of definition of at least one driving situation parameter using a trained, artificial neural network which is applied to the plurality of driving situation parameters, wherein the target function represents a driving maneuver of a motor vehicle in response to a control of the device generated by the plurality of driving situation parameters of the domain of definition. The method further comprises providing a second data set including the numerical value range of the target function and the associated domain of definition of the at least one driving situation parameter.
The present invention is thus advantageously able to replace the simulation of driving situations and/or driving situation parameters by means of an artificial neural network.
In the context of the present method, therefore, an artificial neural network is advantageously used, which has the object of approximating test results. These test results are test results of interest, which are the subject of virtual tests of a device, such as a control device, for the autonomous guidance of a motor vehicle.
In scenario-based testing of systems and system components for the autonomous guidance of a motor vehicle, scenarios are defined which can be described as the abstraction of a traffic situation. A logical scenario here is the abstraction of a traffic situation with the road, the driving behavior and the surrounding traffic without concrete parameter values.
By choosing concrete parameter values, the logical scenario becomes a concrete scenario. Such a concrete scenario corresponds to a particular traffic situation.
An autonomous driving function is realized by a system, for example a control unit. The control unit is conventionally tested in the real vehicle in real traffic situations or, alternatively, validated by virtual tests.
In this context, the present method approximates test results or traffic situations of interest within a predetermined domain of definition of the driving situation parameters used. The domain of definition of the driving situation parameters is chosen such that it covers, for example, safety-related driving situations such as, for example, collisions or near-collisions of involved vehicles.
The test results approximated in this way can then be advantageously validated in the context of virtual tests of the control unit, so that an efficient virtual validation of control units for the autonomous guidance of motor vehicles is possible using the method according to the invention.
According to one aspect of the invention, the method according to the invention further comprises that the EGO parameters include a speed of the motor vehicle and the environment parameters include a speed of a further motor vehicle and a distance between the motor vehicle and the further motor vehicle.
Using these parameters, a so-called cut-in scenario can thus be advantageously approximated. The cut-in scenario can be described as a traffic situation in which a highly automated or autonomous vehicle drives in a predetermined lane and another vehicle merges into the lane of the EGO vehicle from another lane, at a certain distance and at a reduced speed as compared to the speed of the EGO vehicle.
The speed of the EGO vehicle and the further vehicle, which is also referred to as a FELLOW vehicle, is constant. Since the speed of the EGO vehicle is higher than that of the FELLOW, the EGO vehicle must be braked to avoid a collision of the two vehicles.
On the basis of the above-mentioned EGO parameters and environment parameters, it is thus possible to approximate traffic situations by the method according to the invention that are relevant in the given domain of definition of the above-mentioned parameters.
The method according to the invention further comprises that the target function is a safety target function having a numerical value, which has a minimum value at a safety distance between the motor vehicle and the further motor vehicle of ≥VFELLOW×0.55, has a maximum value in a collision between the motor vehicle and the further motor vehicle, and at a safety distance between the motor vehicle and the further motor vehicle of ≤VFELLOW×0.55, has a value which is greater than the minimum value.
The safety target function indicates how safe the traffic situation is for the EGO vehicle. It is specified as follows: If the distance between the EGO vehicle and the FELLOW vehicle is always greater than or equal to the safety distance, the functional value of the safety target function is 0.
The safety distance can be defined as a distance at which, as a function of a speed difference between the EGO vehicle and the FELLOW vehicle and the distance between the EGO vehicle and the FELLOW vehicle, safe braking of the EGO vehicle is always possible without the occurrence of a collision with the FELLOW vehicle.
Such a distance is defined in the present example by a value in meters which corresponds to the velocity VFELLOW×0.55.
The smaller the distance between the EGO vehicle and the FELLOW vehicle or starting from the point when the safety distance is undershot, the target function value more and more approaches the value of 1. Accordingly, if there is a collision of the EGO vehicle and the FELLOW vehicle, the distance between the EGO vehicle and the FELLOW vehicle is less than or equal to 0 and the target function value is 1.
The method according to the invention also can comprise that the target function is a comfort target function or an energy consumption target function having a numerical value which has a minimum value if there is no change in the acceleration of the motor vehicle, has a maximum value in the event of a collision between the motor vehicle and the further motor vehicle, and in a change in the acceleration of the motor vehicle as a function of the amount of change in the acceleration, has a value between the minimum value and the maximum value.
With the aid of the comfort target function, statements can be made as to how pleasant a driving maneuver is for the driver of the EGO vehicle. Great acceleration or deceleration and frequent repetition of these operations is considered to be not comfortable.
The changes in acceleration are called jerks. The driving situation is the more comfortable, the smaller the calculated value of the comfort target function. In the event of a collision between the EGO vehicle and the FELLOW vehicle, energy consumption is a fixed maximum value. The reason for this is that in a motor vehicle with an internal combustion engine and/or a traction battery in an electrically operated vehicle or hybrid vehicle, the tank of gas can generally no longer be used in the event of an accident.
In respect of the cut-in scenario, thus, possible safety-relevant test cases may lie between collision and non-collision cases, which can be defined on the basis of the respective target functions, i.e., the safety target function, the comfort target function and the energy consumption target function.
The method according to the invention furthermore can comprise that the plurality of driving situation parameters, in particular the speed of the motor vehicle and the speed of the further motor vehicle, are generated within the predetermined domain of definition by a random algorithm.
Thus, the plurality of driving situation parameters constituting the data set that is used to approximate the test results can be easily and time efficiently generated.
The method according to the invention can further comprise that a separate artificial neural network is used for approximating the numerical value range of each target function, wherein individual hyperparameters of each artificial neural network are stored in a database.
By using a separate artificial neural network for the approximation of each target function, the artificial neural network used in each case can be trained specifically for the best possible approximation of the relevant target function and is thus advantageously able to produce a more accurate approximation result as compared to an artificial neural network which approximates a plurality of target functions by means of a common algorithm.
The method according to the invention can further comprise that as a function of the target function, the artificial neural network uses between 2 to 15 layers, between 6 and 2048 neurons per layer, and ReLU, LeakyReLU or ELU as the activation function. This way, the best possible approximation results can be achieved in the present configuration.
The second data set comprising the numerical value range of the target function and the assigned domain of definition of the at least one driving situation parameter can be graphically represented two-dimensionally or three-dimensionally as a function of a number of driving situation parameters.
Thus, advantageously, a functional course of the approximated numerical value range of the target function for the given domain of definition of the driving situation parameters used can be represented graphically. This allows for the graphical representation of non-collision cases, collision cases and a boundary between non-collision cases and collision cases.
The invention further relates to a computer-implemented method for providing a trained, artificial neural network for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle.
The method comprises receiving a first set of input training data comprising a plurality of driving situation parameters including environment parameters describing the surroundings of the motor vehicle and EGO parameters describing the state of the motor vehicle, wherein each driving situation parameter has a predetermined domain of definition.
The method further comprises receiving a second set of output training data formed of the numerical value range of the target function and the associated domain of definition of the at least one driving situation parameter, wherein the output training data is associated with the input training data.
Moreover, the method comprises training the artificial neural network to approximate a numerical value range of the target function for the domain of definition of at least one driving situation parameter, wherein the target function represents a driving maneuver of the motor vehicle in response to a control of the device generated by the plurality of driving situation parameters of the domain of definition, based on the input training data and the output training data with a training calculation unit.
Furthermore, the method comprises the provision of the trained, artificial neural network for approximating test results of the virtual test of the device for the at least partially autonomous guidance of the motor vehicle.
As a function of a number of used training data sets and/or training episodes and the optimization performed, the thus trained, artificial neural network is therefore advantageously able to attain an exact approximation result of the target function to be approximated.
The plurality of driving situation parameters, in particular the speed of the motor vehicle and the speed of the further motor vehicle, can be generated within the predetermined domain of definition by a random algorithm and/or by a simulation. The parameter sets of driving situation parameters can thus be generated in a simple manner, for example when using an artificial neural network by applying a random function within a predetermined domain of definition.
The method can also further comprise training the artificial neural network by weight settings of a plurality of data sets of driving situation parameters using a gradient descent backpropagation, in particular using the Adam optimization method.
When values are input to an artificial neural network, the outputs of the neurons are calculated layer by layer and forwarded until they reach the end of the output layer. This is called a forward pass.
A comparison is made between the actual output and the desired target output in order to determine an error. This is called error detection. The target output is specified by the data set and is called the target value. In the simplest case, the error can be determined by the absolute value of the discrepancy between the actual value and the target value. This is called an absolute error. If several data sets are considered, the mean absolute error can be used. For this, the individual absolute errors are summed up and divided by the number of data records being considered.
In classifications, the accuracy is usually the percentage value of how often the output matches the target value.
The error can be traced through the entire network to the input layer. Since the goal is to minimize the error, the gradient descent method can be used to update the weights towards a local minimum. This is called a backward pass.
By adjusting the weights, it is possible for the neural network to solve the specific problem better and better and to improve its accuracy of the predictions. The procedure of forward pass, error detection and backward pass is called backpropagation. A single backpropagation sequence is called a training step.
The Adam optimization method uses different learning rates for different parameters. These can also increase. In addition, these learning rates are influenced by a momentum. This means that successive weight updates in the same gradient direction result in a higher learning rate. In comparison, the learning rate is decreased if the gradient direction changes with successive updates.
The invention further relates to a test unit for approximating test results of a virtual test of a device for the at least partially autonomous guidance of a motor vehicle.
The test unit comprises an input for receiving a first data set comprising a plurality of driving situation parameters including environment parameters describing the surroundings of the motor vehicle and EGO parameters describing the state of the motor vehicle, wherein each driving situation parameter has a predetermined domain of definition.
Furthermore, the test unit can approximate a numerical value range of a target function for the domain of definition of at least one driving situation parameter using a trained, artificial neural network which is applied to the plurality of driving situation parameters, wherein the target function represents a driving maneuver of the motor vehicle in response to a control of the device generated by the plurality of driving situation parameters of the domain of definition.
The test unit further comprises an output for providing a second data set including the numerical value range of the target function and the associated domain of definition of the at least one driving situation parameter.
In the context of the present test unit, therefore, an artificial neural network is advantageously used which has the object of approximating test results as a function of the predetermined domain of definition of the driving situation parameters of interest.
The test results approximated in this way can then be advantageously validated in the context of virtual tests of the device, so that an efficient virtual validation of the device for the autonomous guidance of motor vehicles is possible by means of the test unit according to the invention.
In accordance with a further aspect of the invention, it is provided that a driving situation which underlies the approximation of the test results of the virtual test of the device, formed in particular as a control unit, is a lane change of another motor vehicle into the traffic lane of the motor vehicle employing the plurality of driving situation parameters.
The test unit is thus advantageously able to approximate corresponding test results of the virtual test with regard to, for example, a cut-in scenario.
The invention further relates to a computer program with program code for carrying out the approximation method according to the invention when the computer program is executed on a computer. Moreover, the invention relates to a computer-readable data carrier with program code of a computer program in order to execute the approximation method according to the invention when the computer program is executed on a computer.
The features of the method described herein can be used for approximating test results of a variety of different scenarios or driving situations. Likewise, the test unit according to the invention is designed to test a plurality of different devices or control units of, for example, automobiles, utility vehicles and/or commercial vehicles, ships or aircraft in terms of an approximation of test results.
Further scope of applicability of the present invention will become apparent from the detailed description given hereinafter. However, it should be understood that the detailed description and specific examples, while indicating preferred embodiments of the invention, are given by way of illustration only, since various changes and modifications within the spirit and scope of the invention will become apparent to those skilled in the art from this detailed description.
The present invention will become more fully understood from the detailed description given hereinbelow and the accompanying drawings which are given by way of illustration only, and thus, are not limitive of the present invention, and wherein:
The method comprises receiving S1 a first data set DS1a, DS1b, DS1c comprising a plurality of driving situation parameters including environment parameters FP1, FP2 describing the environment of the motor vehicle and EGO parameters FP3 describing the state of the motor vehicle.
Each driving situation parameter has a predetermined domain of definition. The method further comprises approximating S2 a numerical value range WB1, WB2, WB3 of a target function F1, F2, F3 for the domain of definition of at least one driving situation parameter using a trained, artificial neural network K1.
The artificial neural network K1 is applied to the plurality of driving situation parameters. The target function F1, F2, F3 represents a driving maneuver of the motor vehicle in response to a control of the device generated by the plurality of driving situation parameters of the domain of definition.
In addition, the method comprises providing a second data set DS2a, DS2b, DS2c including the numerical value range WB1, WB2, WB3 of the target function F1, F2, F3 and the associated domain of definition of the at least one driving situation parameter.
The one or more EGO parameters FP3 include a speed VEGO of the motor vehicle, and the environment parameters FP1, FP2 include a speed VFELLOW of another motor vehicle and a distance dSPUR between the motor vehicle and the further motor vehicle.
The plurality of driving situation parameters, in particular the speed VEGO of the motor vehicle and the speed VFELLOW of the further motor vehicle are generated within the predetermined domain of definition by a random algorithm.
As a function of the target function F1, the artificial neural network K1 uses between 2 to 15 layers, between 6 and 2048 neurons per layer, and ReLU as the activation function. Alternatively, for example, LeakyReLU or ELU can be used as the activation function.
The second data set DS2a including the numerical value range of the target function F1 and the associated domain of definition of the at least one driving situation parameter. As a function of the number of driving situation parameters, the above-mentioned range of values can be graphically represented, for example, two-dimensionally or three-dimensionally.
The illustrated target function F1 is a safety target function. The safety target function has a numerical value or function value approximated by the artificial neural network K1 (shown in
The function value has a minimum value at a safety distance between the motor vehicle and the further motor vehicle of ≥VFELLOW×0.55. Furthermore, the function value has a maximum value in the event of a collision between the motor vehicle and the further motor vehicle. In addition, at a safety distance between the motor vehicle and the further motor vehicle of ≤VFELLOW×0.55, the function value has a value which is greater than the minimum value.
The value range WB1 of the second data set DS2a illustrated in
The target function F2 shown in
The numerical value range of the comfort target function from 0 to 140 shown in
On the basis of the illustrated range of values, it can thus be determined whether the values shown lie in a range in which a collision of the vehicle with the other vehicle occurs, in that no collision occurs between the vehicle and the further vehicle or the numerical function value lies in an intermediate range.
The target function F3 shown in
The numerical value range of the comfort target function of 0.1 to 0.2 shown in
On the basis of the illustrated range of values, it can thus be determined whether the values shown lie in a range in which a collision of the vehicle with the other vehicle occurs, in that no collision occurs between the vehicle and the further vehicle or the numerical function value lies in an intermediate range.
The artificial neural network K1 shown in
The first function value F1 relates to a safety target function, the second function value F2 relates to a comfort target function and the third function value F3 relates to an energy consumption target function.
The artificial neural network K1 is thus able to approximate the function value F1 based on a first data set DS1a of the safety target function. On the basis of the function value of the target function F1, the second data set DS2a can then be generated, i.e., a graphical representation of the numerical value range WB1 of the safety target function can be realized.
The artificial neural network K1 receives a first data set DS1a relating to the safety target function, a second data set DS2a relating to the comfort target function and a third data set relating to the energy consumption target function.
The data sets DS1a, DS1b, DS1c each have a plurality of driving situation parameters including environment parameters describing the surroundings of the motor vehicle FP1, FP2 and EGO parameters FP3 describing the state of the motor vehicle.
On the basis of the functional values K1 approximated by the artificial neural network, a second data set DS2a, DS2b, DS2c including the numerical value range WB1, WB2, WB3 of the target function F1, F2, F3 and the assigned domain of definition of the at least one driving situation parameter can subsequently be generated.
In contrast to the embodiment shown in
The first neural network K1 has an input layer L1a, a plurality of hidden layers L2a and an output layer L3a, which outputs the function value of the safety target function F1.
The first artificial neural network K1 also receives the first data set DS1a. On the basis of the approximated function value of the first artificial neural network K1, the second data set DS2a can then be created, which includes the numerical value range of the target function F1 and the associated domain of definition of the at least one driving situation parameter.
The second neural network K2 has an input layer L1b, a plurality of hidden layers L2b, and an output layer L3b, which outputs the function value of the safety target function F2.
The second artificial neural network K2 also receives the second data set DS1b. On the basis of the approximated function value of the second artificial neural network K2, the second data set DS2b can then be created, which includes the numerical value range of the target function F2 and the associated domain of definition of the at least one driving situation parameter.
The third neural network K3 has an input layer L1c, a plurality of hidden layers L2c and an output layer L3c, which outputs the function value of the safety target function F3.
The third artificial neural network K3 also receives the third data set DS1c. On the basis of the approximated function value of the third artificial neural network K3, the second data set DS2c can then be created, which includes the numerical value range of the target function F3 and the associated domain of definition of the at least one driving situation parameter.
The method comprises receiving S1′ a first data set DS1a′, DS1b′, DS1c′ of input training data comprising a plurality of driving situation parameters including environment parameters FP1, FP2 describing the surroundings of the motor vehicle and EGO parameters FP3 describing the state of the motor vehicle. Each driving situation parameter has a predetermined domain of definition.
In addition, the method comprises receiving S2′ of a second data set DS2a′, DS2b′, DS2c′ of output training data including the numerical value range WB1, WB2, WB3 of the target function F1, F2, F3 and the associated domain of definition of the at least one driving situation parameter. The output training data is related to the input training data.
Furthermore, the method has the step of training S3′ of the artificial neural network for approximating a numerical value range WB1, WB2, WB3 of the target function F1, F2, F3 for the domain of definition of at least one driving situation parameter.
The target function F1, F2, F3 represents a driving maneuver of the motor vehicle in response to a control of the device generated by the plurality of driving situation parameters of the domain of definition based on the input training data and the output training data. The approximation is done with a training calculation unit 20.
In addition, the method comprises providing S4′ the trained, artificial neural network K1, K2, K3 for approximating test results of the virtual test of the device for the at least partially autonomous guidance of the motor vehicle.
The plurality of driving situation parameters, in particular the speed VEGO of the motor vehicle and the speed VFELLOW of the further motor vehicle are generated within the predetermined domain of definition by a random algorithm. Alternatively, these can be generated, for example, by a simulation.
The artificial neural network K1, K2, K3 is trained by weight settings of a plurality of data sets of driving situation parameters using a gradient descent backpropagation S3a′, in particular using the Adam optimization method.
The test unit comprises an input 10 for receiving a first data set DS1a, DS1b, DS1c comprising a plurality of driving situation parameters including environment parameters describing the surroundings of the motor vehicle and EGO parameters describing the state of the motor vehicle. Each driving situation parameter in this case has a predefined domain of definition.
Furthermore, the test unit 1 has a processor 12 for approximating a numerical value range WB1, WB2, WB3 of a target function F1, F2, F3 for the domain of definition of at least one driving situation parameter using a trained, artificial neural network K1, K2, K3. The artificial neural network is applied to the plurality of driving situation parameters.
In this case, the target function F1, F2, F3 represents a driving maneuver of the motor vehicle in response to a control of the device generated by the plurality of driving situation parameters of the domain of definition. Furthermore, the test unit 1 has an output 14 for providing a second data set DS2a, DS2b, DS2c including the numerical value range WB1, WB2, WB3 of the target function F1, F2, F3 and the associated domain of definition of the at least one driving situation parameter.
In the present embodiment, a driving situation underlying the approximation of the test results of the virtual test of the device, formed in particular as a control unit, is a lane change of another motor vehicle into the lane of the motor vehicle employing the plurality of driving situation parameters.
Alternatively, an underlying driving situation may be another driving situation such as approaching a traffic light with and/or without a lane change or with one or more vehicles involved.
Although specific embodiments have been illustrated and described herein, it will be understood by those skilled in the art that a variety of alternative and/or equivalent implementations exist. It should be noted that the exemplary embodiment or embodiments are only examples and are not intended to limit the scope, applicability, or configuration in any way.
Rather, the above-mentioned summary and detailed description will provide those skilled in the art with a convenient guide to implementing at least one exemplary embodiment, it being understood that various changes in the functionality and arrangement of the elements may be made without departing from the scope of the appended claims and their legal equivalents.
In general, this application intends to cover changes or adaptations, or variations of the embodiments presented herein.
The invention being thus described, it will be obvious that the same may be varied in many ways. Such variations are not to be regarded as a departure from the spirit and scope of the invention, and all such modifications as would be obvious to one skilled in the art are to be included within the scope of the following claims.
| Number | Date | Country | Kind |
|---|---|---|---|
| 10 2019 122 414.4 | Aug 2019 | DE | national |
| 19192741.7 | Aug 2019 | EP | regional |
| 19192743.3 | Aug 2019 | EP | regional |
This nonprovisional application is a continuation of International Application No. PCT/EP2020/073057, filed on Aug. 18, 2020, which claims priority under 35 U.S.C. § 119(a) to European Patent Application No. 19192743.3, which was filed in Europe on Aug. 21, 2019, and which are both herein incorporated by reference.
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/EP2020/073057 | Aug 2020 | US |
| Child | 17559716 | US |
