EP4377802A1 - Verfahren und computerprogramm zum überwachen von ausgaben eines generators eines generativen kontradiktorischen netzwerks - Google Patents
Verfahren und computerprogramm zum überwachen von ausgaben eines generators eines generativen kontradiktorischen netzwerksInfo
- Publication number
- EP4377802A1 EP4377802A1 EP22757923.2A EP22757923A EP4377802A1 EP 4377802 A1 EP4377802 A1 EP 4377802A1 EP 22757923 A EP22757923 A EP 22757923A EP 4377802 A1 EP4377802 A1 EP 4377802A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- generator
- generative adversarial
- adversarial network
- driving system
- autonomous driving
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F11/00—Error detection; Error correction; Monitoring
- G06F11/004—Error avoidance
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F11/00—Error detection; Error correction; Monitoring
- G06F11/36—Prevention of errors by analysis, debugging or testing of software
- G06F11/3668—Testing of software
- G06F11/3672—Test management
- G06F11/3688—Test management for test execution, e.g. scheduling of test suites
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0475—Generative networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/094—Adversarial learning
Definitions
- the invention relates to a method and a computer program for monitoring outputs of a generative adversarial network generator
- autonomous vehicles In order to drive complex scenes safely, efficiently and smoothly, autonomous vehicles must estimate future trajectories of other traffic objects, also referred to as agents in this context, and protect themselves against the uncertainty in the trajectory estimation in order to choose effective and robust measures.
- Another approach are deep learning based methods.
- K. Messaoud, N. Deo, M.M. Trived and F. Nashashibi discloses. “Trajectory Prediction for Autonomous Driving based on Multi-Head Attention with Joint Agent-Map Representation”.
- ArXiv:2005.02545v3 [cs.CV] 2 Sep 2020 a deep learning encoder-decoder architecture model for multimodal trajectory prediction of vehicles using the methods described in A. Vaswani, N. Shazeer, N. Parmar, J Uszkoreit, L. Jones, AN Gomez, L. Kaiser, and I. Polosukhin. “Attention is all you need”.
- Multiple observation proposed in Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017 is used.
- This model takes a joint representation of static information and surrounding agents as input and uses a deep neural network to learn the environmental context and the complex interactions between the different agents.
- the quality of this proposal when used in autonomous vehicles is difficult to assess.
- the object of the invention was how an intelligent algorithm for predicting the trajectory of an autonomous vehicle can be validated.
- the invention provides a method for monitoring outputs of a generative adversarial network generator in the validation of a safety-critical software and/or hardware module executing the generative adversarial network of an autonomous driving system, wherein after a training phase of the generative adversarial network
- a discriminator of the generative adversarial network the outputs of the generator and context information of the autonomous driving system, comprising o static context information, comprising map and/or navigation information to which the autonomous driving system has access, and o information on interactions of the autonomous driving system with traffic objects , comprising reading in data from at least one environment detection sensor of the autonomous driving system;
- the discriminator depending on the context information of the autonomous driving system, detects safety-critical outputs among the outputs of the generator
- the discriminator • in the case of a detected safety-critical output, the discriminator provides the autonomous driving system with a warning signal. This uses the discriminator as a kind of output monitor. The method is therefore self-confident or self-aware and can be used in real time in safety-critical systems such as autonomous vehicles.
- the discriminators are used during a training phase in generative adversarial networks. After training, i.e. in an operational phase, only the trained generator is accessed. The discriminator is discarded after the generator is trained. According to the invention, the discriminator is also used after the training phase of the gene nerative adversarial network, namely for monitoring the outputs of the generator during validation. This means that deep learning algorithms can also be used in safety-critical systems, since the system knows whether the deep learning algorithm cannot handle the input in some scenarios.
- the invention provides a computer program for monitoring outputs of a generative adversarial network generator.
- the computer program includes instructions which cause a computer to carry out the steps of the method according to the invention when the computer executes the computer program.
- the commands of the computer program according to the invention include machine commands, source text or object code written in assembly language, an object-oriented programming language, for example C++, or in a procedural programming language, for example C.
- the computer program is a hardware-independent application program, for example is provided via a data carrier or a data carrier signal, according to one aspect using software over the air technology, for any hardware, for example via middleware.
- the computer program is a hardware-dependent program, for example firmware for a control module of the autonomous driving system.
- the data carrier includes volatile memory, for example RAM, DRAM, SRAM, and non-volatile memory, for example ROM, flash EEPROM.
- the data carriers are, for example, flash Memory cards, USB sticks.
- the data carrier is connected to an in/out system of a microcontroller of the autonomous driving system and transmits the computer program to the microcontroller.
- the method is based on the functionalities of a generator and a discriminator in a generative adversarial network.
- the generator predicts future trajectories of the agents comprising the ego driving system.
- the discriminator checks the output of the generator during the training and deployment phase and provides feedback, e.g. valid or invalid, on the predicted trajectories based on the agent's context information.
- Generator and/or discriminator can be designed as deep neural networks.
- GANs Generative adversarial networks, abbreviated to GANs, are described, for example, by Ian J. Goodfellow et al. discussed in arXiv:1406.2661.
- Generator and discriminator are trained at the same time.
- the generative adversarial network performs a minimax two-person zero-sum game.
- the task of the generator during training is to generate a predicted (false/invalid) trajectory of the agent as close as a ground truth (real/valid) trajectory.
- the generator tries to cheat the discriminator.
- the discriminator has to classify whether the trajectory is real/valid or not. If a fake/invalid generator trajectory is not classified by a discriminator, the generator loss will decrease and the discriminator loss will increase. Training converges with well-trained generators and discriminators.
- the generator and/or the discriminator are designed as deep neural networks, comprising layers of convolutional networks and/or recurrent networks.
- Autonomous driving system designates the components and functionalities of an autonomous vehicle at the system level as well as the autonomous vehicle as such.
- the discriminator classifies the part of the safety-critical output that is invalid based on the context information. This will only classify the corresponding part as invalid instead of the entire output.
- the predicted trajectory is valid from 0 to 2 seconds. After 2 seconds the predicted trajectory is invalid.
- the generative adversarial network is a conditional generative adversarial network.
- driving maneuvers and/or road traffic regulations are mandatory conditions for the generator. This allows the predicted trajectories to be checked using the constraints.
- Another advantage is that the results of the deep learning method are limited by the constraints imposed by the modeling results.
- the safety-critical software and/or flardware module is a module for predicting the trajectory of the autonomous driving system. Based on the context information of the autonomous driving system, driving maneuver restrictions for the autonomous driving system are calculated and applied as conditions to the generator. This can be implemented particularly advantageously by means of a conditional generative adversarial network.
- the method is particularly advantageous when used in urban areas, for example during roundabout maneuvers. It generates trajectories for surrounding agents by considering the static context information and interactions between agents in complex scenes. Due to the self-monitoring of the discriminator, the method can also be used in safety-critical systems such as autonomous vehicles.
- the calculation of the driving maneuver restrictions is based, for example, on map information, for example high-resolution maps, and/or sensor data from the environment perception of the autonomous driving system.
- the output of the generator is evaluated by checking the predicted trajectories of the generator based on the given boundary conditions.
- a human could generate thousands of trajectories.
- noise is used as another input to the generator to improve the variety of the generator's outputs.
- a backup solution of the trajectory prediction is triggered in the trajectory prediction in the event of a detection generated safety-critical trajectory.
- FIG. 1 shows an exemplary embodiment of a conditional generative adversarial network for trajectory prediction.
- FIG. 3 shows an embodiment of an output of a generator that does not meet a condition
- Fig. 4 shows an embodiment of an output of a generator that satisfies a condition
- Fig. 6 shows an embodiment of an output of the generator, which was recognized as invalid by the discriminator
- Fig. 7 is a schematic representation of an embodiment of the erfindungsge Permitted method.
- Figure 1 shows a conditional generative adversarial network cGAN.
- a generator G and a discriminator D are trained simultaneously and are coupled to one another via a feedback loop, for example.
- the generator G predicts future trajectories T for an autonomous driving system E and outputs these as outputs Out.
- the generator G reads static context information, for example information from a high-resolution map Map, and interactions Int with other traffic objects/agents A, for example other vehicles.
- the generator G reads in noise signals, for example random noise. So that the diversity of spending out increases.
- Driving maneuver restrictions are determined based on the information from the map and data from at least one environment detection sensor of the autonomous driving system E as a further input.
- the maneuver restrictions are conditions Con imposed on the generator.
- Expenses Out should fulfill the conditions Con.
- the goal of the generator G is to predict trajectories T under the given conditions Con, for example in complex urban environments.
- the discriminator D reads information from the map Map, the interactions Int and reference data GT, for example a ground truth trajectory T, and based on this determines whether the trajectory prediction Out of the generator G is valid.
- the conditions Con that the trajectory predictions Out of the generator G have to fulfill are turning left and driving straight ahead.
- an invalid and possibly safety-critical output Out of the generator G is, for example, the trajectory prediction of turning right.
- the discriminator D determines this output Out as valid, see Fig. 3. In this case, if the output Out of the generator G is the trajectory prediction right turn, the discriminator D determines this output as safety-critical, ie not valid, see FIG.
- FIG 5 shows the discriminator D in the deployment phase, in which the discriminator D indicates whether a trajectory prediction Out is valid or not.
- Fig. 6 shows the case of a trajectory prediction of the generator G, which is detected by the discriminator D as safety-critical.
- the object E to be predicted for example a car, drives behind another vehicle A.
- the condition Con is go straight.
- Another vehicle A is driving in the opposite direction to the object E in the opposite lane.
- the generator G predicts the trajectory T for an overtaking maneuver. This is safety-critical under the given conditions Con.
- the discriminator D reads the outputs Out of the generator G and context information In of the autonomous driving system on.
- the discriminator D detects, depending on the context information In, safety-critical outputs Out) from the outputs Out of the generator G.
- the discriminator D provides the autonomous driving system with a warning signal.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Quality & Reliability (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Computer Hardware Design (AREA)
- Traffic Control Systems (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| DE102021208047.2A DE102021208047A1 (de) | 2021-07-27 | 2021-07-27 | Verfahren und Computerprogramm zum Überwachen von Ausgaben eines Generators eines generativen kontradiktorischen Netzwerks |
| PCT/EP2022/071040 WO2023006803A1 (de) | 2021-07-27 | 2022-07-27 | Verfahren und computerprogramm zum überwachen von ausgaben eines generators eines generativen kontradiktorischen netzwerks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| EP4377802A1 true EP4377802A1 (de) | 2024-06-05 |
Family
ID=83006157
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP22757923.2A Withdrawn EP4377802A1 (de) | 2021-07-27 | 2022-07-27 | Verfahren und computerprogramm zum überwachen von ausgaben eines generators eines generativen kontradiktorischen netzwerks |
Country Status (3)
| Country | Link |
|---|---|
| EP (1) | EP4377802A1 (de) |
| DE (1) | DE102021208047A1 (de) |
| WO (1) | WO2023006803A1 (de) |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102019206720B4 (de) | 2019-05-09 | 2021-08-26 | Volkswagen Aktiengesellschaft | Überwachung eines KI-Moduls einer Fahrfunktion eines Fahrzeugs |
-
2021
- 2021-07-27 DE DE102021208047.2A patent/DE102021208047A1/de not_active Withdrawn
-
2022
- 2022-07-27 WO PCT/EP2022/071040 patent/WO2023006803A1/de not_active Ceased
- 2022-07-27 EP EP22757923.2A patent/EP4377802A1/de not_active Withdrawn
Also Published As
| Publication number | Publication date |
|---|---|
| DE102021208047A1 (de) | 2023-02-02 |
| WO2023006803A1 (de) | 2023-02-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| DE112017006530B4 (de) | Rückmeldung für ein autonomes fahrzeug | |
| EP3748453B1 (de) | Verfahren und vorrichtung zum automatischen ausführen einer steuerfunktion eines fahrzeugs | |
| DE102018203583B4 (de) | Verfahren, Fahrerassistenzsystem sowie Kraftfahrzeug zur Prädiktion einer Position oder einer Trajektorie mittels eines graphbasierten Umgebungsmodells | |
| EP3789926B1 (de) | Verfahren zum erkennen einer adversarialen störung in eingangsdaten eines neuronalen netzes | |
| DE102021109395A1 (de) | Verfahren, systeme und vorrichtungen für benutzerverständliche erklärbare lernmodelle | |
| DE102016007563A1 (de) | Verfahren zur Trajektorienplanung | |
| WO2022106414A2 (de) | Verfahren und system zur annotation von sensordaten | |
| EP3877231A1 (de) | Prädiktion eines voraussichtlichen fahrverhaltens | |
| DE102020200876B4 (de) | Verfahren zum Verarbeiten von Sensordaten einer Sensorik eines Fahrzeugs | |
| CN111753371B (zh) | 一种车身控制网络模型的训练方法、系统、终端和存储介质 | |
| EP3968213A1 (de) | Verfahren und vorrichtung zum ermitteln eines gleisgebundenen schienenpfades in einer gleisanlage | |
| DE102021004426A1 (de) | Verfahren zum Trainieren einer autonomen Fahrfunktion | |
| DE102019217637A1 (de) | Verfahren und Vorrichtung zum Betreiben eines Fahrzeugs | |
| DE102018124161A1 (de) | Verfahren zum zumindest teilautomatisierten Führen eines Kraftfahrzeugs auf einer Fahrbahn | |
| DE102021203588A1 (de) | Verfahren und Steuergerät zum Schätzen eines Verhaltens eines Systems | |
| Agamennoni et al. | Automated extraction of driver behaviour primitives using Bayesian agglomerative sequence segmentation | |
| EP4377802A1 (de) | Verfahren und computerprogramm zum überwachen von ausgaben eines generators eines generativen kontradiktorischen netzwerks | |
| DE102019104966A1 (de) | Selbstlernende Steuervorrichtung und Verfahren für selbstlernende Steuervorrichtung | |
| EP4000011B1 (de) | Komponentenbasierte verarbeitung von eingangsgrössen | |
| DE102022131178B3 (de) | Verfahren zum automatisierten Führen eines Fahrzeugs sowie Verfahren zum Erzeugen eines hierzu fähigen Modells des Maschinellen Lernens sowie Prozessorschaltung und Fahrzeug | |
| DE102020120934A1 (de) | Verfahren zum Bereitstellen eines komprimierten neuronalen Netzes zur Multi-Label Multi-Klassen Kategorisierung, Fahrzeugassistenzeinrichtung zur Umgebungskategorisierung und Kraftfahrzeug | |
| DE102024202568A1 (de) | Verfahren zur Objektdetektion in Bilddaten | |
| DE102021208472B3 (de) | Computerimplementiertes Verfahren zum Trainieren eines Machine-Learning-Modells für ein Fahrzeug oder einen Roboter | |
| DE102019204849A1 (de) | Erkennung eines durch Personen ausgehenden Gefahrenpotentials | |
| EP4113392B1 (de) | Verfahren zum prüfen der zuverlässigkeit einer ki-basierten objekt-detektion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: UNKNOWN |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE |
|
| PUAI | Public reference made under article 153(3) epc to a published international application that has entered the european phase |
Free format text: ORIGINAL CODE: 0009012 |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE |
|
| 17P | Request for examination filed |
Effective date: 20240125 |
|
| AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR |
|
| DAV | Request for validation of the european patent (deleted) | ||
| DAX | Request for extension of the european patent (deleted) | ||
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN |
|
| 18D | Application deemed to be withdrawn |
Effective date: 20240906 |