EP2308010A1 - Verfahren zur rechnergestützten erkennung eines spezifischen objekts aus einer datenmenge basierend auf einer interaktion mit einem nutzer - Google Patents
Verfahren zur rechnergestützten erkennung eines spezifischen objekts aus einer datenmenge basierend auf einer interaktion mit einem nutzerInfo
- Publication number
- EP2308010A1 EP2308010A1 EP09780519A EP09780519A EP2308010A1 EP 2308010 A1 EP2308010 A1 EP 2308010A1 EP 09780519 A EP09780519 A EP 09780519A EP 09780519 A EP09780519 A EP 09780519A EP 2308010 A1 EP2308010 A1 EP 2308010A1
- Authority
- EP
- European Patent Office
- Prior art keywords
- user
- specific object
- real
- training examples
- learning method
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/40—Software arrangements specially adapted for pattern recognition, e.g. user interfaces or toolboxes therefor
Definitions
- the invention relates to a method for computer-aided recognition or identification of a specific object from a data volume based on an interaction with a user as well as a corresponding device and a corresponding computer program product.
- Document [1] describes an interactive training method with which a detector for a classification task is trained based on real-time learning.
- the support systems known from the prior art have the disadvantage that they do not perform a completely interactive training or are not used to recognize a specific object specified by a user.
- the object of the invention is therefore to provide a completely interactive method for the computer-aided recognition of a specific object from a data set.
- a data quantity processed using a real-time-capable learning method is displayed via a user interface, wherein a user can identify training examples for the specific object in the displayed data volume via the user interface.
- the specific object is an arbitrary object that can be specified or recognized in accordance with the user preferences in the dataset.
- object is to be construed broadly and may include any type of objects, such as objects, living beings, persons, and the like, as well as constellations of several such objects (combined into an object) has specific object and is able to recognize objects in the dataset as specific objects
- the specific object is thus no longer an object class, but rather a fixed object corresponding to the user preferences, in particular in the form of the instance of an object class.
- the mouse Being configured manner, preferably it includes a monitor for displaying the amount of data and a corresponding input unit in the form of a keyboard or mouse. Particularly preferred is the use of a mouse, the mouse can be moved to a corresponding object and the identification of the object can be made as a training example by clicking.
- a knowledge base is updated real-time-capable learning method, with the real-time learning method learning a model for identifying the specific object based on the updated knowledge base, whereby the processing tete amount of data is received. This is followed by an updated display of the data quantity according to the above step a).
- the method according to the invention is thus characterized by an iterative process of displaying and subsequently updating the data volume, the method being completely interactive, since after each user interaction in the form of identification of one or more training examples, the knowledge base is updated immediately the real-time learning method is performed.
- any of the real-time-capable learning methods known from the prior art can be used in the method according to the invention.
- the on-line learning described in the document [1] or [2] is used.
- the entire disclosure of these references is incorporated herein by reference.
- the online learning described in the document is based on a so-called boosting method in which a classifier is learned through on-line training.
- an on-line version of the adaboost learning method known from the prior art is used.
- the classifier no longer identifies classes of objects but rather a specifically defined object.
- the method according to the invention By stepwise transferring knowledge from a user to a system in the method according to the invention, the effort on the human side is drastically reduced and is increasingly limited to corrective interventions.
- the method according to the invention adapts fluently to the respective preferences of the user and not only takes work away from it, but even acts proactively in his or her sense.
- objects are identified in the amount of data processed using the real-time learning method be classified as the specific object by the learned model.
- the user selects training examples by the user such that the user of the real-time learning method correctly marks objects classified as specific objects as positive training examples and / or objects incorrectly classified as specific objects as negative training examples by the real-time-capable learning method. This creates a simple and intuitive way for a user to generate new training examples.
- a particularly fast and effective identification of training examples can be made, in particular, if only negative training examples can be marked by the user via the user interface. It is taken into account that after a plurality of iterations the number of negative training examples becomes ever smaller, since with advanced learning an ever better identification of the specific object takes place in the dataset.
- the user interface can also mark, as negative training examples, objects that correspond to the specific object and have not been classified as a specific object by the real-time learning method.
- the display in step a) contains a list of the objects identified by the real-time-capable learning method. This list can be easily and intuitively edited by a user by highlighting positive or negative training examples in the list.
- the inventive method can be used for any types of data.
- the method identifies specific objects in a dataset from one or more video streams.
- the specific object in the video stream may be a person characterized by its outward appearance, particularly its clothing and / or body measurements and / or body movements.
- This can eg be evaluated in a surveillance video, which acts a specific person performs.
- a surveillance video which acts a specific person performs.
- a user may first note the appearance of a person doing an unauthorized act (for example, a theft).
- the method according to the invention the occurrence of the same person in the video stream or in other video streams can be determined interactively. In this way, a search for wanted persons in video streams is made possible by an interactive computer-aided method.
- the method according to the invention is used to identify or identify specific objects (and not object classes) in a dataset.
- the method can also be used to annotate large amounts of data.
- the real-time learning method does not provide objects which are classified as specific objects, but objects which are assumed to have a certain annotation.
- the user can in turn identify training examples in the annotated objects, for example he can mark incorrectly annotated objects accordingly.
- the annotation effort is greatly simplified by a user, since a large part of the annotations made computer-aided and the user only has to take corrective action.
- the invention further includes an apparatus for computer-aided recognition of a specific object from a data set based on an interaction with a user.
- the device comprises a
- User interface for displaying a data set processed using a real-time learning method, wherein a user uses the user interface to provide training examples for the spe- can identify a cifical object in the displayed dataset.
- an arithmetic unit is provided which, in operation, after each identification made by the user of one or more training examples, in particular of 20 training examples or less, preferably of 15 or 10 training examples or less, and more preferably of a single training example, a knowledge base and real-time-capable learning method, wherein the real-time learning method learns a model for identifying the specific object based on the updated knowledge base, thereby obtaining the processed amount of data displayed by the user interface.
- the device is preferably designed such that any desired variant of the method according to the invention can be carried out with the device.
- the invention further relates to a computer program product with a program code stored on a machine-readable carrier for carrying out each of the above-described variants of the inventive method when the program is run on a computer.
- Fig. 1 is a schematic representation of the steps which are performed in an embodiment of the invention.
- FIG. 2 shows a schematic representation of the objects identified in the steps of an embodiment of the method according to the invention.
- a plurality of video streams is used as data quantity processed.
- the goal here is to identify an object in the form of a specific person in a computer-aided manner in the video streams.
- the particular person is, for example, a person included in another video stream and viewed by a user.
- the user is thus aware of the appearance of the person, and the user's knowledge is used interactively in the inventive method for computer-aided identification of the same person in the plurality of video streams.
- One use case may be that a person in a video stream has committed a tampering action that has been filmed by a video camera.
- the embodiment of the invention described here uses a real-time or online learning method known per se.
- a variant of such a learning method is described in the document [1] or [2].
- This learning method is used in the cited document for the classification of objects, but according to the invention is used to search for specific objects.
- the real-time learning method is first initialized with a user-identified image of the person sought as a training example, this image is generated for example by marking a point in a video that shows the person searched. Subsequently, a generalization of a descriptor for the individual description of the person is derived from several consecutive frames of the video. The derivation of general sated descriptors is well known from classification methods and is therefore not described in detail.
- Classifier learned and applied to the video streams As a result, on a user interface 1, a table of objects identified by the classifier as the searched person is displayed (step S1).
- the learning method was computer-aided carried out on the designated in Fig. 1 by reference numeral 2 arithmetic unit.
- the user interface 1 is shown schematically in FIG. 1 as a monitor with a keyboard and a connected mouse. Via the user interface, the user can make corresponding inputs, in particular he can mark displayed objects by clicking on the mouse. The user now selects from the displayed table of detected objects one or more of such objects, which were recognized by the classifier as the searched person, but are not that person. These objects are marked as a negative training example by the user with the mouse.
- an update of the known training examples immediately follows. That is, the knowledge base on which the online learning method operates is updated immediately after being marked with the mouse by including the negative training examples in the knowledge base. Based on the updated knowledge base, the online learning process for learning the classifier is performed immediately in step S2 by means of the arithmetic unit 2. Finally, the objects resulting from the updated classifier, which are classified as the searched person, are displayed in a new step S1 on the user interface. This is indicated by a corresponding arrow P from step S2 to step S1. Thus, an updated list of recognized objects is displayed in real time, it being assumed that the updated list is based on the further training example contains fewer objects that are incorrectly identified as the person you are looking for.
- step S2 is carried out, i. the knowledge base is updated with the further training examples and the online learning process is carried out.
- the updated result is displayed again in step S1.
- the method is repeated iteratively a plurality of times, the method being characterized in that any user intervention in the form of flags of one or more training examples results in an immediate update of the knowledge base and an immediate visual effect by display on the user interface 1.
- This is made possible by an on-line learning method, i. through a learning process that is fast enough to enable real-time training based on newly added training examples.
- the learning method described in the document [1] or [2] is preferably used, which is an online version of the adaboost learning method known to the person skilled in the art.
- FIG. 2 again clearly shows the filtering of a searched person from a dataset.
- a user P marks the user interface of a person P as an object to be recognized in a video image at a location in a video stream.
- step S108 persons from another video stream or elsewhere in the video stream are reproduced step by step in the form of a list.
- This list contains those persons identified as person P according to a real-time learning method.
- a user makes the identification of positive training examples, ie of correct meetings, from the corresponding lists with a mouse click.
- the persons marked as correct hits are provided in FIG. 2 with corresponding reference signs T. It can be seen that in step S100, where the filter has not yet been learned, no real hit is found. Gradually, more and more hits are found, which are always identified by the user as positive training examples. Finally, four hits are identified in step S108.
- the process finally ends in a step S109 in which a person in the video stream is recognized as the original person P from the video image of the step SlO1.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| DE102008037022A DE102008037022A1 (de) | 2008-08-08 | 2008-08-08 | Verfahren zur rechnergestützten Erkennung eines spezifischen Objekts aus einer Datenmenge basierend auf einer Interaktion mit einem Nutzer |
| PCT/EP2009/058930 WO2010015488A1 (de) | 2008-08-08 | 2009-07-13 | Verfahren zur rechnergestützten erkennung eines spezifischen objekts aus einer datenmenge basierend auf einer interaktion mit einem nutzer |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| EP2308010A1 true EP2308010A1 (de) | 2011-04-13 |
Family
ID=41078006
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| EP09780519A Ceased EP2308010A1 (de) | 2008-08-08 | 2009-07-13 | Verfahren zur rechnergestützten erkennung eines spezifischen objekts aus einer datenmenge basierend auf einer interaktion mit einem nutzer |
Country Status (3)
| Country | Link |
|---|---|
| EP (1) | EP2308010A1 (de) |
| DE (1) | DE102008037022A1 (de) |
| WO (1) | WO2010015488A1 (de) |
-
2008
- 2008-08-08 DE DE102008037022A patent/DE102008037022A1/de not_active Ceased
-
2009
- 2009-07-13 EP EP09780519A patent/EP2308010A1/de not_active Ceased
- 2009-07-13 WO PCT/EP2009/058930 patent/WO2010015488A1/de not_active Ceased
Non-Patent Citations (1)
| Title |
|---|
| See references of WO2010015488A1 * |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2010015488A1 (de) | 2010-02-11 |
| DE102008037022A1 (de) | 2010-04-08 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| DE102014210820B4 (de) | Verfahren zum Nachweis von großen und Passagierfahrzeugen von festen Kameras | |
| DE112017002821T5 (de) | Verfahren, System und Computerprogrammprodukt zur interaktiven Identifizierung von gleichen Personen oder Objekten in Videoaufnahmen | |
| WO2010040668A1 (de) | Verfahren und einrichtung zur analyse von umgebungsobjekten und/oder umgebungsszenen, wie zur objekt- und szenenklassensegmentierung | |
| DE102014213408A1 (de) | Verfahren zur Ermittlung eines dreidimensionalen Modelldatensatzes eines wenigstens ein Gefäßsegment umfassenden Blutgefäßsystems | |
| DE112022001468B4 (de) | Selektives unkenntlichmachen von bildern | |
| DE102011015730A1 (de) | Phantombilddatenbank (3D) | |
| EP3029628A1 (de) | Verfahren zum Erzeugen eines Trainingsbildes | |
| DE102012218870A1 (de) | Verbessertes Erkennen herrenloser Objekte unter Verwendung des Erfassens von Fussgängern | |
| DE112020001625T5 (de) | Bildsuchsystem und Bildsuchverfahren | |
| DE112016002275T5 (de) | Koordinierte benutzerwortauswahl zum übersetzen und erhalten von kontextinformtionen für das ausgewählte wort | |
| EP2756458A1 (de) | Maschinelles lernverfahren zum maschinellen erlernen von erscheinungsformen von objekten in bildern | |
| DE102016124277A1 (de) | Verfahren zum Ermitteln und Anzeigen von Produkten auf einer elektronischen Anzeigeeinrichtung | |
| DE102023202598A1 (de) | Computerimplementiertes Verfahren und System zum Trainieren eines Lehrernetzwerks eines teilüberwachten Lernalgorithmus | |
| Riedl | Unsupervised methods for learning and using semantics of natural language | |
| EP2359308B1 (de) | Vorrichtung zur erzeugung und/oder verarbeitung einer objektsignatur, überwachungsvorrichtung, verfahren und computerprogramm | |
| DE102009060687A1 (de) | Verfahren und Vorrichtung zum rechnergestützten Annotieren von Multimediadaten | |
| EP2308010A1 (de) | Verfahren zur rechnergestützten erkennung eines spezifischen objekts aus einer datenmenge basierend auf einer interaktion mit einem nutzer | |
| EP4435638A1 (de) | Computerimplementiertes verfahren zum erkennen eines zustandes eines innenraumes eines öffentlichen verkehrsmittels | |
| DE102022201161A1 (de) | Objektklassifizierung mit einem einstufigen metabasierten Objektdetektor unter Verwendung von Klassenprototypen | |
| Al-Behadili et al. | Teilüberwachtes Lernen von emblematischen Gesten Semi-supervised learning of emblematic gestures. | |
| EP4250246A1 (de) | Computerimplementiertes verfahren zum erkennen eines objektes in einem innenraum eines öffentlichen verkehrsmittels | |
| DE112021005555T5 (de) | Multitasking-lernen über gradienteilung zur umfangreichen menschlichen analyse | |
| DE102023202597A1 (de) | Computerimplementiertes Verfahren und System zum Trainieren eines Lehrernetzwerks eines teilüberwachten Lernalgorithmus | |
| Villalba | Prediction, detection, and correction of misunderstandings in interactive tasks | |
| DE102023203379A1 (de) | System und verfahren zur iterativen verfeinerung und kuratierung von bildern auf basis von visuellen vorlagen |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 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 |
|
| 17P | Request for examination filed |
Effective date: 20101217 |
|
| AK | Designated contracting states |
Kind code of ref document: A1 Designated state(s): 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 SE SI SK SM TR |
|
| AX | Request for extension of the european patent |
Extension state: AL BA RS |
|
| RIN1 | Information on inventor provided before grant (corrected) |
Inventor name: WINTER, MARTIN Inventor name: ROTH, PETER, MICHAEL Inventor name: NGUYEN, THUY Inventor name: GRABNER, HELMUT Inventor name: FRUEHSTUECK, BERNHARD Inventor name: BISCHOF, HORST Inventor name: BIRCHBAUER, JOSEF, ALOIS |
|
| RIN1 | Information on inventor provided before grant (corrected) |
Inventor name: WINTER, MARTIN Inventor name: ROTH, PETER, MICHAEL Inventor name: NGUYEN, THUY Inventor name: GRABNER, HELMUT Inventor name: FRUEHSTUECK, BERNHARD Inventor name: BISCHOF, HORST Inventor name: BIRCHBAUER, JOSEF, ALOIS |
|
| 17Q | First examination report despatched |
Effective date: 20110722 |
|
| DAX | Request for extension of the european patent (deleted) | ||
| RIN1 | Information on inventor provided before grant (corrected) |
Inventor name: WINTER, MARTIN Inventor name: ROTH, PETER, MICHAEL Inventor name: NGUYEN, THUY Inventor name: GRABNER, HELMUT Inventor name: FRUEHSTUECK, BERNHARD Inventor name: BISCHOF, HORST Inventor name: BIRCHBAUER, JOSEF, ALOIS |
|
| REG | Reference to a national code |
Ref country code: DE Ref legal event code: R003 |
|
| RIN1 | Information on inventor provided before grant (corrected) |
Inventor name: WINTER, MARTIN Inventor name: ROTH, PETER, MICHAEL Inventor name: NGUYEN, THUY Inventor name: GRABNER, HELMUT Inventor name: FRUEHSTUECK, BERNHARD Inventor name: BISCHOF, HORST Inventor name: BIRCHBAUER, JOSEF, ALOIS |
|
| STAA | Information on the status of an ep patent application or granted ep patent |
Free format text: STATUS: THE APPLICATION HAS BEEN REFUSED |
|
| 18R | Application refused |
Effective date: 20120129 |