Training Neural Networks Using a Neural Network

Abstract

The disclosure relates to a method for training a first neural network, in particular for generating training data for at least one second neural network, using a controller, wherein measurement data ascertained by at least one surroundings sensor or artificially generated data of initially ten traffic scenarios is received, the received measurement data is fed to the first neural network as input data in order to train the first neural network, and the first neural network which is trained on the basis of the input data is used to generate data of traffic scenarios which differ from the initial traffic scenarios. Furthermore, the disclosure relates to a method for training at least one second neural network, to a controller, to a computer program, and to a machine-readable storage medium.

Claims

1. A method for training a first neural network using a control device, the method comprising: receiving measurement data of initial traffic scenarios that are one of (i) ascertained by at least one environment sensor and (ii) artificially generated; training the first neural network by feeding the received measurement data of the initial traffic scenarios to the first neural network as input data; and generating, using the first neural network, data of traffic scenarios which differ from the initial traffic scenarios.

2. The method as claimed in claim 1 further comprising: receiving measurement data of ideal traffic scenarios that are ascertained by at least one environment sensor of different vehicles; and training the first neural network using the received measurement data of the ideal traffic scenarios.

3. The method as claimed in claim 1, the feeding the measurement data of initial traffic scenarios to the first neural network further comprising: feeding the measurement data of the initial traffic scenarios to the first neural network in a manner depending on at least one of: at least one sensor type of the at least one environment sensor; at least one mounting position of the at least one environment sensor; and a position of at least one vehicle.

4. A method for training at least one second neural network using a control device, the method comprising: generating data of a plurality of traffic scenarios using a first neural networks; and training the at least one second neural network by feeding the generated data of the plurality of traffic scenarios to the at least one second neural network as input data.

5. The method as claimed in claim 4, wherein the data of the plurality of traffic scenarios generated by the first neural network are configured to train an object classifier of at least one environment sensor, the object classifier being based on the at least one second neural network, depending on a sensor type and a mounting position of the at least one environment sensor.

6. The method as claimed in claim 4, the training the at least one second neural network further comprising: training the at least one second neural network using the data of the plurality of traffic scenarios generated by the first neural network and using measurement data ascertained by environment sensors.

7. The method as claimed in claim 6, the training the at least one second neural network further comprising: using the measurement data ascertained by the environment sensors to check reactions of the at least one second neural network to traffic situations.

8. The method as claimed in claim 1, wherein the method is carried out by a control device.

9. The method as claimed in claim 1, wherein the method is carried out by a computer program having instructions that, when carried out by one of a computer and a control device, cause the one of the computer and the control device to carry out the method.

10. A non-transitory machine-readable storage medium that stores a computer program for training a first neural network, the computer program having instructions that, when carried out by one of a computer and a control device, cause the one of the computer and the control device to: receive measurement data of initial traffic scenarios that are one of (i) ascertained by at least one environment sensor and (ii) artificially generated; train the first neural network by feeding the received measurement data of the initial traffic scenarios to the first neural network as input data; and generate, using the first neural network, data of traffic scenarios which differ from the initial traffic scenarios.

11. The method as claimed in claim 1, wherein the first neural network generates training data for at least one second neural network.

12. The method as claimed in claim 4, wherein the method is carried out by a control device.

13. The method as claimed in claim 4, wherein the method is carried out by a computer program having instructions that, when carried out by one of a computer and a control device, cause the one of the computer and the control device to carry out the method.

Description

[0042] The following text explains preferred exemplary embodiments of the invention in more detail based on highly simplified schematic illustrations, in which

[0043] FIG. 1 shows a schematic illustration of a traffic scenario with a vehicle and

[0044] FIG. 2 shows a schematic flow chart for illustrating a method according to one embodiment.

[0045] FIG. 1 shows a schematic illustration of a traffic scenario 1 with a vehicle 2. For the sake of clarity, the traffic scenario 1 has one vehicle 2 but may preferably have a plurality of vehicles 2 which ascertain measurement data and can use same to train neural networks.

[0046] The measurement data of the traffic situation 1 can preferably be ascertained by environment sensors 4, 6. The ascertained measurement data can be preprocessed or at least intermediately stored by vehicle-based control units 8. In a subsequent step, the measurement data of the traffic situation 1 ascertained by the environment sensors 4, 6 can be transmitted to a control device 12 via a communication link 10.

[0047] The environment sensors 4, 6 may be configured, for example, as radar sensors, Lidar sensors, camera sensors, ultrasonic sensors, position sensors and the like.

[0048] In the exemplary embodiment illustrated, the vehicle 2 has a camera sensor 4 and a Lidar sensor 6. The vehicle 2 thus has two different sensor types which are secured to the vehicle 2 in different mounting positions 5, 7. By way of example, the camera sensor 4 can be secured in the region of a windshield 5 of the vehicle 2 and the Lidar sensor 6 can be secured to the roof 7 of the vehicle 2.

[0049] The measurement data ascertained by the environment sensors 4, 6 are transmitted to the central control device 12 by the vehicle-based control unit 8 and are used there as input data to train neural networks. To this end, the control device 12 can be configured, for example, as a high-performance computer.

[0050] The vehicle 2 can preferably be configured as a vehicle able to be operated in automated fashion according to the BASt standard.

[0051] FIG. 2 shows a schematic flow chart for illustrating a first method 14 and a second method 16 according to one embodiment.

[0052] A first method 14 consists in training a first neural network. A second method 16 serves for using the first neural network to train other neural networks. The methods 14, 16 are carried out substantially by the control device 12.

[0053] In one step 18 of the first method 14, measurement data ascertained by at least one environment sensor 4, 6 or artificially generated data of initial traffic scenarios 1 are received. The traffic scenarios 1 are preferably ideal traffic scenarios 1 or traffic scenarios without accidents.

[0054] The measurement data received are subsequently fed the first neural network as input data in order to train the first neural network 20.

[0055] The first neural network which is trained based on the input data is then used to generate 22 data from traffic scenarios 1′ which differ from the initial traffic scenarios 1.

[0056] In the second method 16, the data 22 of traffic scenarios 1′ which differ from the initial traffic scenarios 1, said data being generated by the first neural network, are fed to the second neural network as input data in order to train the second neural network 24.

[0057] In another step 26, the trained second neural network is implemented, for example, in a vehicle-based control unit 8 or an environment sensor 4, 6 of the vehicle 1. The second neural network is then used in at least one object identification, at least one autonomous driving function, at least one environment sensor or for at least one vehicle control device of a vehicle able to be operated in automated fashion.