WO2024200136A1 - Kalibrierung der unsicherheit eines objektdetektors - Google Patents
Kalibrierung der unsicherheit eines objektdetektors Download PDFInfo
- Publication number
- WO2024200136A1 WO2024200136A1 PCT/EP2024/057366 EP2024057366W WO2024200136A1 WO 2024200136 A1 WO2024200136 A1 WO 2024200136A1 EP 2024057366 W EP2024057366 W EP 2024057366W WO 2024200136 A1 WO2024200136 A1 WO 2024200136A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- uncertainty
- statement
- neural network
- calibration
- calibrated
- 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.)
- Ceased
Links
Classifications
-
- 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
- G06N3/09—Supervised learning
-
- 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
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural 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
- G06N3/044—Recurrent networks, e.g. Hopfield 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
- G06N3/045—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
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0499—Feedforward networks
Definitions
- the present invention relates to a method, in particular a computer-implemented method, for calibrating at least one uncertainty of a neural network of an object detector, a method, in particular a computer-implemented method, for detecting at least one object by means of an object detector comprising a neural network calibrated according to the invention, a computer program for carrying out the method according to the invention and a computer program product on which the method according to the invention is stored.
- ADAS advanced driver assistance systems
- neural networks can be used, for example, to process input data from various sensors, such as radar, lidar or ultrasonic sensors, as well as cameras.
- ADAS functions typically provided by ADAS systems serve to support the driver while the driver still retains control of driving the vehicle, or, depending on the level of automation, fully automated driving can be realized.
- Examples of ADAS functions are various methods for detecting objects or obstacles on the road, methods for detecting lane boundaries and/or keeping the vehicle in a lane, methods for detecting rain on the windshield, or methods for supporting or carrying out a parking process.
- the input data from the ADAS sensors used can be subject to a variety of problems, such as insufficient sensor resolution, problematic weather conditions or rarely occurring scenarios.
- certain objects may be obscured by other objects and thus not visible in the input data. In all of these cases, knowledge of the uncertainty of the respective forecast would be of key importance.
- DE 102020215860A1 discloses a method for correcting input image data from a plurality of cameras of an all-round vision system that is impaired by rain, light and/or dirt, in which input image data captured by the cameras is processed by a trained neural network, which also outputs an uncertainty for the determined image correction.
- a similar method for correcting input image data from a camera is known from DE102020215859A1.
- neural networks especially those that serve a regression task, i.e. networks that output continuous values as prediction results, tend to have associated uncertainties that are inaccurate and therefore not directly usable. In order to increase the accuracy, calibration or adjustment of the uncertainties is therefore required. This also applies, among other things, to object localization using an object detector, which precedes any object classification. Here, little or no information is available on the confidence of the respective prediction, i.e. on its reliability.
- DE102019217300A1 discloses a method for training a neural network to output a prediction and associated uncertainty, in which the determined uncertainty is taken into account by the loss function. In this way, it can be ensured that the network determines well-calibrated uncertainties if the distribution of the data to be processed is similar to the distribution of the training data.
- a method specifically for calibrating uncertainties in regression has become known from “Accurate Uncertainties for Deep Learning Using Calibrated Regression” by V. Kuleshov et al., published on arXiv: 1807.00263, 2018 (DOI: https://doi.org/10.48550/arXiv.1807.00263). However, the method carries the risk of over-fitting. In addition, it is based on isotonic regression and is therefore comparatively computationally intensive.
- the object underlying the invention is to provide a possibility for calibrating uncertainties, which is characterized by high accuracy and high computational efficiency.
- the object underlying the invention is achieved by the method according to claim 1, the method according to claim 10, the computer program according to claim 13 and the computer program product according to claim 14.
- the object underlying the invention is achieved by a method, in particular a computer-implemented method, for calibrating at least one uncertainty of a neural network of an object detector, which neural network is designed to make at least one statement about an object and at least one of the uncertainty associated with the statement.
- the procedure comprises the following steps:
- the calibration of the uncertainties is separated from the training of the neural network and is downstream of it.
- the calibration of the uncertainties is carried out using a, in particular separate, calibration data set.
- objects are first localized and then classified.
- the object detector thus comprises a localization unit and a classification unit.
- the present invention relates in particular to the localization of objects and the associated uncertainties with regard to object localization.
- the procedure according to the invention offers the advantage of high accuracy and resource efficiency. This makes it particularly suitable for use in connection with embedded systems, such as those frequently used in driver assistance systems.
- the normalization of the generated uncertainties and the deviation of the statement from the indication, ie the ground truth the advantageous effect that the uncertainties for each statement can be correctly calibrated regardless of the respective object size and object class.
- the neural network is a trained neural network.
- the method according to the invention can be part of a training method for training the neural network or it can be a separate method that runs separately from the training method.
- the neural network is, for example, a convolutional neural network (CNN), a mixture-density network, a transformer network, a recurrent neural network, a multilayer perceptron or a generative, adversial network.
- CNN convolutional neural network
- the neural network is, for example, a convolutional neural network (CNN), a mixture-density network, a transformer network, a recurrent neural network, a multilayer perceptron or a generative, adversial network.
- the object detector is an anchor-based, a central point-based, a multi-stage, a single-stage, a CNN-based or a transformer-based object detector.
- it can be a YOLO object detector, an SSD object detector or an EfficientDet object detector.
- the statement is a statement about a position of an object, a statement about a property of the object or a statement about a geometric size of the object.
- the position of the object can be specified, for example, using one or more coordinates for the object.
- a property of the object is, for example, an object type and a geometric size relates, for example, to its geometric dimensions, its size, cross-sectional area or a volume of the object. It should be noted that other statements about the object can also be made and fall under the present invention.
- each coordinate can be considered as a statement. However, it is also conceivable to combine all coordinates into one statement.
- the uncertainty and the deviation of the statement from the indication are standardized to a geometric size of the object, in particular a length, width, height, area or volume of the object.
- This has the advantageous effect that different object sizes, aspect ratios, geometric dimensions and sizes of bounding boxes can be taken into account for the calibration. This can, for example, prevent larger objects or bounding boxes from having a negative influence on the calibration for smaller objects or bounding boxes.
- the parameters of the model are optimized by means of an optimization method, in particular by means of an isotonic regression, a gradient method, a Platt scaling method or a histogram binning approach.
- the error function is a distance metric, in particular a mean square deviation or an absolute deviation.
- a distance metric is a metric or distance function for a metric space, which assigns a non-negative, real value to two elements or points of a space.
- a distance metric is symmetrical and satisfies the triangle inequality.
- the invention can also be used in conjunction with other error functions.
- the calibration parameter includes a scaling, which is taken into account in particular for the error function.
- the scaling can be one or more constant scaling factors, a scaling vector or a scaling matrix.
- the scaling can be used for all predicted uncertainties after the calibration has been completed. By means of scaling, a calibration for all predicted uncertainties can be achieved quickly, effectively and precisely.
- a further advantageous embodiment includes determining a calibration rule by means of which at least one uncertainty determined by means of the neural network is calibrated.
- the calibration rule is therefore taken into account in particular when using the neural network to output the statement and the uncertainty.
- the rule takes into account the at least one calibration parameter and can include a multiplication with a scalar value or also a model.
- the object underlying the invention is also achieved by a method, in particular a computer-implemented method, for detecting at least one object by means of an object detector comprising a neural network calibrated according to the invention.
- the method comprises the following method steps:
- the input data can be recorded using one or more sensors. In the case of a driver assistance system, for example, these are various ADAS sensors.
- the statement about the object is in particular a position of the object.
- the method then concerns object localization.
- the uncertainty is preferably standardized in the same way as the uncertainties generated during the execution of the calibration procedure using the calibration data. In particular, the uncertainty is normalized to a geometric size of the object.
- the calibration of the uncertainty is preferably carried out using a previously determined calibration specification.
- a calibrated, absolute uncertainty is determined and output.
- the calibrated, normalized uncertainty can be multiplied by the normalization.
- the method according to the invention for detecting at least one object by means of an object detector is preferably used for driver assistance systems. However, the method can also be used in other areas, for example in computer vision or for medical imaging methods.
- the method for calibrating at least one uncertainty of a neural network of an object detector is also preferably used in connection with a driver assistance system.
- statements about at least one object are determined using at least one ADAS sensor, for example a camera.
- the object can be another road user in the vehicle's surroundings, an animal, a road boundary, an object in the vehicle's surroundings, or a traffic sign or traffic light or light signal.
- an uncertainty for the statement made is also determined, calibrated and output.
- the object underlying the invention is further achieved by a computer program with instructions which, when the computer program is executed by a computer, cause the computer to carry out the method according to the invention for calibrating at least one uncertainty of a neural network or the method for detecting at least one object by means of an object detector, depending on one of the described embodiments, to carry out, as well as by a computer program product on which the computer program according to the invention is stored.
- Fig. 1 shows a first advantageous embodiment of the method according to the invention
- Fig. 2 shows an object detection using a neural network calibrated according to the invention.
- Fig. 1 assumes a pre-trained neural network NN in an object detector.
- the neural network NN is used to determine a statement p about an object 0 and an uncertainty a associated with the statement.
- the statement here is the position of the object 0.
- the uncertainties a output by a neural network NN are typically subject to errors and do not correspond to the actual uncertainty.
- the uncertainties a In order to enable a precise prediction of the uncertainty a, the uncertainties a must be calibrated, which can be carried out using the method according to the invention. After calibration, the uncertainty a can be interpreted probabilistically, i.e. statements can be made such as that the correct value lies in the predicted uncertainty interval with a probability of 95%.
- calibration data Ikai for the neural network NN and an indication Ind for all calibration data Ikai are initially provided.
- the calibration data Ikai are processed using the neural network and a statement p and associated uncertainty a are made based on the calibration data Ikai.
- the generated uncertainty ⁇ J is then normalized and divided, for example, by a normalization factor n, ie
- the deviation of the statement p from the indication Ind is normalized with the normalization factor n.
- the normalized, generated uncertainty a nor is then compared with the normalized deviation of the statement p from the indication Ind using an error function Error and at least one calibration parameter pkai is optimized based on the comparison.
- the error function is a distance metric for which a scaling s is taken into account and provided that the scaling s is a constant, scalar value
- a function based on a mean square deviation can be chosen as the error function Error, which is then given for N detections, for example, by
- an error function Error can be used based on an absolute deviation, which in turn is given by
- A ⁇ Ind - [i ⁇ is the absolute deviation, i.e. the difference between the indication Ind and the prediction p.
- a calibration rule KAL can be determined, which can be used for the embodiment shown here for object localization in the object detector, as illustrated in Fig. 2.
- information about a vehicle environment and in particular about at least one object 0 in the vehicle environment is first recorded with one or more ADAS sensors. This is made available to the neural network NN as input.
- the neural network NN is designed to use the input to make a statement about the position P of at least one object and to output an uncertainty o associated with the position.
- a statement can also be made about a size of the object 0, for example a length, width, height, area or volume of the object 0.
- the uncertainty o output by the network NN is then normalized o nor and calibrated using the calibration specification KAL. The result is a calibrated, normalized uncertainty Okai.nor, from which a calibrated, absolute uncertainty can also be determined.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Probability & Statistics with Applications (AREA)
- Medical Informatics (AREA)
- Multimedia (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202480022480.3A CN120917458A (zh) | 2023-03-31 | 2024-03-20 | 对对象检测器不确定性的校准 |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| DE102023203027.6 | 2023-03-31 | ||
| DE102023203027.6A DE102023203027A1 (de) | 2023-03-31 | 2023-03-31 | Kalibrierung der Unsicherheit eines Objektdetektors |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2024200136A1 true WO2024200136A1 (de) | 2024-10-03 |
Family
ID=90571649
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/EP2024/057366 Ceased WO2024200136A1 (de) | 2023-03-31 | 2024-03-20 | Kalibrierung der unsicherheit eines objektdetektors |
Country Status (3)
| Country | Link |
|---|---|
| CN (1) | CN120917458A (de) |
| DE (1) | DE102023203027A1 (de) |
| WO (1) | WO2024200136A1 (de) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102019217300A1 (de) | 2019-11-08 | 2021-05-12 | Robert Bosch Gmbh | Verfahren zum Trainieren eines künstlichen neuronalen Netzes, Computerprogramm, Speichermedium, Vorrichtung, künstliches neuronales Netz und Anwendung des künstlichen neuronalen Netzes |
| DE102020215859A1 (de) | 2020-12-15 | 2022-06-15 | Conti Temic Microelectronic Gmbh | Korrektur von Bildern einer Kamera bei Regen, Lichteinfall und Verschmutzung |
| DE102020215860A1 (de) | 2020-12-15 | 2022-06-15 | Conti Temic Microelectronic Gmbh | Korrektur von Bildern eines Rundumsichtkamerasystems bei Regen, Lichteinfall und Verschmutzung |
-
2023
- 2023-03-31 DE DE102023203027.6A patent/DE102023203027A1/de not_active Withdrawn
-
2024
- 2024-03-20 CN CN202480022480.3A patent/CN120917458A/zh active Pending
- 2024-03-20 WO PCT/EP2024/057366 patent/WO2024200136A1/de not_active Ceased
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE102019217300A1 (de) | 2019-11-08 | 2021-05-12 | Robert Bosch Gmbh | Verfahren zum Trainieren eines künstlichen neuronalen Netzes, Computerprogramm, Speichermedium, Vorrichtung, künstliches neuronales Netz und Anwendung des künstlichen neuronalen Netzes |
| DE102020215859A1 (de) | 2020-12-15 | 2022-06-15 | Conti Temic Microelectronic Gmbh | Korrektur von Bildern einer Kamera bei Regen, Lichteinfall und Verschmutzung |
| DE102020215860A1 (de) | 2020-12-15 | 2022-06-15 | Conti Temic Microelectronic Gmbh | Korrektur von Bildern eines Rundumsichtkamerasystems bei Regen, Lichteinfall und Verschmutzung |
Non-Patent Citations (4)
| Title |
|---|
| BHATT DHAIVAT ET AL: "f-Cal: Aleatoric uncertainty quantification for robot perception via calibrated neural regression", 2022 INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION (ICRA), IEEE, 23 May 2022 (2022-05-23), pages 6533 - 6539, XP034146633, DOI: 10.1109/ICRA46639.2022.9811903 * |
| DI FENG ET AL: "Towards Safe Autonomous Driving: Capture Uncertainty in the Deep Neural Network For Lidar 3D Vehicle Detection", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 14 April 2018 (2018-04-14), XP081551113 * |
| M-H. LAVES ET AL.: "Well-calibrated Model Uncertainty with Temperature Scaling for Dropout Variational Inference", ARXIV: 1909.13550, 2019 |
| V. KULESHOV ET AL.: "Accurate Uncertainties for Deep Learning Using Calibrated Regression", ARXIV: 1807.00263, 2018 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN120917458A (zh) | 2025-11-07 |
| DE102023203027A1 (de) | 2024-10-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| DE102020214283A1 (de) | Vorrichtung zur erkennung von hindernissen, fahrzeugsystem mit dieser und verfahren dafür | |
| DE112014002019B4 (de) | Kurvenformmodelliereinrichtung, Fahrzeuginformationsverarbeitungssystem, Kurvenformmodellierverfahren und Kurvenformmodellierprogramm | |
| DE112009000480T5 (de) | Dynamische Objektklassifikation | |
| DE102011119767A1 (de) | Erscheinungsbild-gestützte vereinigung von kamera- undentfernungssensordaten für mehrere objekte | |
| DE102021201445A1 (de) | Computerimplementiertes Verfahren zum Testen der Konformität zwischen realen und synthetischen Bildern für maschinelles Lernen | |
| WO2020061603A1 (de) | Verfahren und vorrichtung zur analyse eines sensordatenstroms sowie verfahren zum führen eines fahrzeugs | |
| DE102018124979A1 (de) | Fahrerassistenzsystem zur Bestimmung einer Entfernung zwischen zwei Fahrzeugen mit einer Kamera | |
| DE102024100521A1 (de) | Bildgestützte generierung von beschreibenden und wahrnehmungs-nachrichten von automobilszenen | |
| DE102023129628A1 (de) | System zur durchgängigen modellierung von wahrnehmungsstörungen eines fahrzeugs | |
| EP3961479A1 (de) | Verfahren und vorrichtung zur klassifizierung von targets | |
| DE102022127740A1 (de) | System und Verfahren zum dynamischen Dimensionieren von Fahrzeug und Anhänger aus Schattenprofilen | |
| DE102022120140B4 (de) | Verfahren und Steuervorrichtung zum Vorgeben eines Objektzustands eines Objekts in einem Umfeld einer Umfelderfassungseinrichtung | |
| DE102024205413A1 (de) | Automatische identifizierung der fehlsignaturen für autonome fahranwendungen | |
| DE102023203027A1 (de) | Kalibrierung der Unsicherheit eines Objektdetektors | |
| WO2024200137A1 (de) | Kalibrierung von unsicherheiten von neuronalen netzen | |
| DE102018213132A1 (de) | Nicht-überwachtes oder teilüberwachtes Lernen eines Klassifikators für die Umfeld-Erfassung eines Fahrzeugs | |
| DE102012008780A1 (de) | Verfahren und Einrichtung zur Erfassung wenigstens eines Fahrbahnrandes und Kraftfahrzeug | |
| DE102021133977A1 (de) | Verfahren und System zur Klassifikation von Szenarien eines virtuellen Tests sowie Trainingsverfahren | |
| DE102019220615A1 (de) | Verfahren und Vorrichtung zum Erkennen und Klassifizieren von Objekten | |
| DE102024104360B4 (de) | Fahrzeugsystem für ein fahrzeug zum vorhersagen einer zukünftigen anwesenheit eines fussgängers auf einer strasse | |
| EP4113392B1 (de) | Verfahren zum prüfen der zuverlässigkeit einer ki-basierten objekt-detektion | |
| DE102023207709A1 (de) | Verbesserung der Aussagegenauigkeit eines neuronalen Netzes | |
| DE102024210706A1 (de) | Objektdetektion mit Distanzbestimmung | |
| DE102024203903A1 (de) | Verfahren zum Trainieren eines Maschinenlernmodells für eine Detektionsanwendung in einem Fahrzeug | |
| DE102024204250A1 (de) | Verfahren zum Trainieren eines Maschinenlernmodells für eine Detektion von Verkehrslinienmarkierungen |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 24715073 Country of ref document: EP Kind code of ref document: A1 |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 202480022480.3 Country of ref document: CN |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 112024001575 Country of ref document: DE |
|
| WWP | Wipo information: published in national office |
Ref document number: 202480022480.3 Country of ref document: CN |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 24715073 Country of ref document: EP Kind code of ref document: A1 |

