DE10110038B4 - Method for the autonomous adaptation of a classifier - Google Patents

Method for the autonomous adaptation of a classifier Download PDF

Info

Publication number
DE10110038B4
DE10110038B4 DE10110038A DE10110038A DE10110038B4 DE 10110038 B4 DE10110038 B4 DE 10110038B4 DE 10110038 A DE10110038 A DE 10110038A DE 10110038 A DE10110038 A DE 10110038A DE 10110038 B4 DE10110038 B4 DE 10110038B4
Authority
DE
Germany
Prior art keywords
classifier
class
classification
classified
objects
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.)
Expired - Fee Related
Application number
DE10110038A
Other languages
German (de)
Other versions
DE10110038A1 (en
Inventor
Ulrich Dr.-Ing. Kreßel
Christian Dr.rer.nat. Wöhler
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Daimler AG
Original Assignee
DaimlerChrysler AG
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by DaimlerChrysler AG filed Critical DaimlerChrysler AG
Priority to DE10110038A priority Critical patent/DE10110038B4/en
Publication of DE10110038A1 publication Critical patent/DE10110038A1/en
Application granted granted Critical
Publication of DE10110038B4 publication Critical patent/DE10110038B4/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/62Methods or arrangements for recognition using electronic means
    • G06K9/64Methods or arrangements for recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references, e.g. resistor matrix
    • G06K9/66Methods or arrangements for recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references, e.g. resistor matrix references adjustable by an adaptive method, e.g. learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/20Analysis of motion
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/20Image acquisition
    • G06K9/32Aligning or centering of the image pick-up or image-field
    • G06K2009/3291Pattern tracking

Abstract

Method for the autonomous adaptation of a classifier within a system for the classification of time-varying scenarios,
whereby the classifier was trained in advance of the classification with typical patterns of objects to be classified (training set),
and the classifier assigns patterns, which can not be assigned to any of the objects to be classified, to a rejection class,
characterized,
that, by evaluating the timing of the results of the classification, certain objects previously assigned to a rejection class are classed as belonging to a particular class by associating class labels with them,
that the training set will be extended to include these objects and class labels,
and that the classifier is trained again with this new training set.

Description

  • The The invention relates to a method according to the preamble of the claim In general, in the methods known from the prior art for the classification of objects (for example: pictures, picture objects, Communication or radar signals), a detection and a classification stage connected in series. The detection level provides information about that, that at a respective position (for example: coordinates in pictures) is an object of a certain size. Through the subsequent classification stage is checked by means of any suitable classification method whether the detected object belongs to a certain object class or Not. The detection level has always over one Tracking algorithm that detects a detected object's identification number (ID), which and does not change the same physical object from image to image as long as This object is sufficiently well recognizable. The classification module is trained with an initial learning set on objects created by be delivered to the detection stage.
  • Out DE 19636028 C1 For example, a method for object detection is known which subjects stereo image data to classification as input data. But it is also conceivable how out DE 19831413 A1 known to detect by means of template matching objects in image data and subjected to a subsequent classification. This is also suitable for the analysis of time-varying scenarios DE 19534230 A1 Known detection method of color segmentation.
  • During the Use of the classifier in the image processing system according to Training phase no adaptation takes place, since it is at the training method used is always supervised learning in which before training a human 'teacher' manually correct each training example Object class that has to assign 'label', which is not possible in the operation of the system in the real world.
  • From the DE 19518993 A1 a method for the classification of objects is known in which, in addition to the assignment of the object data to an object to be classified, the assignment to a rejection class (garbage class) is possible. The assignment to the rejection class takes place in cases in which no clear classification can be performed or in which the classification result is not unique. The object data associated with the rejection class can be used to re-train the classifier to extend its class set with respect to new object classes and thereby increase its classification performance.
  • The The object of the invention is a method for adapting a classifier to find which of its existing set of object classes to novel object expressions respectively changing environmental conditions adapts.
  • The Task is by a method with the features of the patent claim 1 solved. Advantageous embodiments and further developments of the invention are in the subordinate claims described.
  • In innovative way of solving the novel method for adapting a classifier per se temporally changeable Scenarios the task by the temporal sequence of the results the classification are evaluated. There are certain objects, which has not previously been assigned to any class or rejection class were classified as belonging to a specific class. In this way becomes the original one training set used to adapt the classifier Objects are extended, whereupon the classifier with this new training set is trained again.
  • Based of figures and embodiments the invention will be described in detail below.
  • 1 shows the flowchart for explaining the autonomous generation of an extended data set (training set) for classifier training.
  • In particularly profitable way, the method is not on the Limited interaction with a certain type of classifier. There However, it in the automatic according to the invention Assignment of labels, i. the automatic classification of certain Objects to certain classes, bad labels comes when a complete sequence is classified incorrectly, it is advantageous Use classifiers that train with square mean approaches become. Polynomial classifiers are particularly useful for this purpose or neural networks, like one with the backpropagation algorithm trained multi-layer perceptron. Classifiers of this type work generally robust against mislabeled (ranked) Pattern in training set. Equally conceivable is the use of classifiers, which refers to methods such as template matching or color segmentation support.
  • The Basis of the method according to the invention is founded in the observation, in the case of the use of a classifier in an image processing system, the probability of a particular object, e.g. a pedestrian or a background pattern, correct at a given time often only 80% or less. That the error probability in the classification of a picture sequence is per time step in general at over 20%. This is due both to inadequacies of the classifier as well as the quality the object segmentation a role.
  • In particularly advantageous manner now pursues the inventive method an object by means of the detection stage over a longer period, or over several Snapshots, so that by suitable integration of the classifier provided detection probabilities per snapshot integrated Recognition probability for the period concerned. Such an integrated Recognition probability is much more reliable, as the probability of detection of a single snapshot.
  • In profitable way can be such an integrated probability of recognition through education one of the mean of the detection probabilities of the generate individual snapshots. However, it is just as conceivable alternatively generate an integrated detection probability, in which these with the percentage of times during the Tracing the object to which the object passes through the classifier assigned to a particular class (1-by-k assignment), equated becomes.
  • In a subsequent Step then advantageously this integrated recognition probability can be checked, whether she over or below a predetermined threshold. If the threshold is exceeded, then all in the previously considered sequence will pass through the detection stage captured and provided with the same ID image sections labeled, which of the largest integrated Detection probability is assigned. This is the one Label chosen, which belongs to the class, which was classified as the most probable. Be in this way So also those objects provided with this label, which previously not or classified as belonging to any other class.
  • It is self-evident, that this procedure of automatic generation and assignment of Labeling within the method according to the invention directly to the Case of the presence of multiple object and garbage classes extended can be.
  • It Alternatively, in a profitable way, also conceivable, alone the fact that an object is over a certain period of time could be safely followed shut down, that this object is a relevant object and not to a garbage pattern (a pattern that none of the interesting Heard classes) or a non-classifiable pattern. Following the Tracking, then all patterns are assigned to the same class and provided with the same label. For identification and determination of the corresponding class, it is conceivable to refer to the above Procedure for generation and evaluation (threshold comparison) of a integrated recognition probability.
  • Such a procedure is as a flowchart in 1 shown. The starting point here is that an object was tracked over an interval of N time steps t i , i = 1... N. The classifier will assign the object ROI (t i ) to one of the classes K for each of the time steps t i . This time step-specific class assignment c (t i ) with c (t i ) = 1... K is determined by the respectively used classifier according to its assignment rules. If n k corresponds to the number of time steps for which an object ROI (t i ) has been assigned to a class k, according to Equation 1:
    Figure 00060001
  • The tracking specific probability P k that an object ROI (t i ) belongs to a class k is defined according to Equation 2:
    Figure 00060002
  • From this probability measure P k , a tracking-specific class assignment C can be derived by means of the rule described in Equation 3 ('the winner-takes-all rule'):
    Figure 00060003
  • Although this tracking-specific class assignment C is significantly more reliable than a single class assignment c (t i ), it is no sooner than the end of N-time tracking available. After completion of the tracking then all tracked objects ROI (t i ) are assigned to the class C and provided with the appropriate class label. If a sufficient number of such newly-labeled objects are available, they will be added to the training set.
  • The The training set of the classifier will follow the one described above Procedure then extended to the newly provided with labels pattern. With this advanced training set, the classifier is re-trained. The higher here the share of new training examples (newly labeled patterns) is, the higher it is also their influence the coefficients and thus the recognition behavior of the classifier. contains For example, the new training set is much newer than initial Training examples, the classifier is the one in the initial Training phase learned patterns largely "forgotten".
  • The inventive method can profitably also repeatedly using ever newer scenes accomplished with the training set becoming more powerful. Here are from time to time the newly learned object characteristics or environmental conditions Always stronger different from the initial ones learned.
  • The inventive method can be profitable using a polynomial classifier accomplished by the classwise moment matrices of the polynomial classifier with the newly gained training examples and autonomously generated Class labels successively recursive according to Schürmann (Schürmann, J .; Pattern Classification, Wiley-Interscience, New York, 1996) to be updated. Are enough new examples added has been determined, the updated coefficient matrix according to Schürmann. The advantage of this approach is that the required memory with increasing number of new training examples does not increase, but remains constant.
  • The inventive method of course can also be used in other areas in which objects apply in temporally variable Classify scenarios; such as in the radar or Communications technology. When used in communication technology the objects to be classified are usually different Broadcasts.

Claims (16)

  1. Method for the autonomous adaptation of a classifier within a system for the classification of temporally variable scenarios, wherein the classifier was trained in advance of the classification with typical patterns of objects to be classified (training set), and the classifier patterns, which can not be assigned to any of the objects to be classified assigned to a rejection class, characterized that by evaluating the timing of the results of the classification, classifying certain objects previously assigned to a rejection class as belonging to a particular class by associating class labels with those classes, expanding the training set around those objects and class labels, and that the classifier is trained again with this new training set.
  2. Method according to claim 1, characterized in that that the patterns of the objects to be classified image data and / or Communication signals and / or radar signals are.
  3. Method according to one of claims 1 or 2, characterized that the classifier is a polynomial classifier.
  4. Method according to claim 3, characterized that the classwise moment matrices of the polynomial classifier with the newly gained training examples and autonomously generated Class labels are successively updated recursively, and that for the Case, that enough new training examples in the update in the updated coefficient matrix is determined.
  5. Method according to one of claims 1 or 2, characterized that the classifier is a support vector machine.
  6. Method according to one of claims 1 or 2, characterized that the classifier is a neural network.
  7. Method according to one of claims 1 or 2, characterized that the classifier uses a template matching method power.
  8. Method according to one of claims 1 or 2, characterized that the classifier is a process for color segmentation makes use of.
  9. Method according to one of claims 1 or 2, characterized that the classifier is a method of analyzing stereo image pairs makes use of.
  10. Method according to one of claims 1 to 9, characterized that in evaluating the timing of the results of the Classification, from the recognition probabilities of snapshots the temporally variable Scenarios integrated recognition probabilities are calculated, and that in a subsequent Step, this integrated detection probability is then checked, whether she over or below a predetermined threshold.
  11. Method according to claim 10, characterized in that that of the detection probabilities exceed the threshold the maximum probability is selected that's this Detection probability assigned class is determined, and that at least one of the manifestations of a Object, then classified as belonging to this class.
  12. Method according to claim 10, characterized in that that in evaluating the timing of the results of the Classification while the duration of a successful tracking of a particular object appearing manifestations are used, and that at least one of the manifestations of this object, then belonging to the class is classified as having the most probable classifier while has classified the persecution.
  13. Method according to claim 12, characterized in that that the greatest likelihood with which the classifier classifies an object during a pursuit has, is determined by an integrated detection probability determined and then checked, whether she over or below a predetermined threshold.
  14. Use of the method according to one of claims 1 to 13, in the analysis of image data.
  15. Use of the method according to one of claims 1 to 13, in the analysis of radar data.
  16. Use of the method according to one of claims 1 to 13, in the analysis of communication signals.
DE10110038A 2001-03-02 2001-03-02 Method for the autonomous adaptation of a classifier Expired - Fee Related DE10110038B4 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
DE10110038A DE10110038B4 (en) 2001-03-02 2001-03-02 Method for the autonomous adaptation of a classifier

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
DE10110038A DE10110038B4 (en) 2001-03-02 2001-03-02 Method for the autonomous adaptation of a classifier

Publications (2)

Publication Number Publication Date
DE10110038A1 DE10110038A1 (en) 2002-09-19
DE10110038B4 true DE10110038B4 (en) 2005-10-06

Family

ID=7676046

Family Applications (1)

Application Number Title Priority Date Filing Date
DE10110038A Expired - Fee Related DE10110038B4 (en) 2001-03-02 2001-03-02 Method for the autonomous adaptation of a classifier

Country Status (1)

Country Link
DE (1) DE10110038B4 (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE10344299B4 (en) * 2003-09-23 2016-08-04 Volkswagen Ag Classification of the objects located in an environment of a motor vehicle
DE102005062154B4 (en) * 2005-12-22 2007-10-31 Daimlerchrysler Ag Generation of large realistic training samples for classifiers from a few examples
DE102016215509A1 (en) * 2016-08-18 2018-02-22 Conti Temic Microelectronic Gmbh Mirror target detection in a radar system in a vehicle

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE19518993A1 (en) * 1995-05-29 1996-12-05 Sel Alcatel Ag Device and method for automatic detection or classification of objects
DE19802261A1 (en) * 1998-01-22 1999-07-29 Daimler Chrysler Ag Processing of a time sequence of digitized images, e.g. for interpretation of road traffic situations from a vehicle
DE19942223A1 (en) * 1999-09-03 2001-03-15 Daimler Chrysler Ag Classification procedure with rejection classes e.g. for classification of road sign recognition, involves defining rejection class R as an additional class to which are assigned

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE19518993A1 (en) * 1995-05-29 1996-12-05 Sel Alcatel Ag Device and method for automatic detection or classification of objects
DE19802261A1 (en) * 1998-01-22 1999-07-29 Daimler Chrysler Ag Processing of a time sequence of digitized images, e.g. for interpretation of road traffic situations from a vehicle
DE19942223A1 (en) * 1999-09-03 2001-03-15 Daimler Chrysler Ag Classification procedure with rejection classes e.g. for classification of road sign recognition, involves defining rejection class R as an additional class to which are assigned

Also Published As

Publication number Publication date
DE10110038A1 (en) 2002-09-19

Similar Documents

Publication Publication Date Title
US10127445B2 (en) Video object classification with object size calibration
Li et al. Deeptrack: Learning discriminative feature representations by convolutional neural networks for visual tracking.
JP6018674B2 (en) System and method for subject re-identification
Sun et al. A real-time precrash vehicle detection system
Javed et al. A hierarchical approach to robust background subtraction using color and gradient information
Javed et al. Online detection and classification of moving objects using progressively improving detectors
US5537488A (en) Pattern recognition system with statistical classification
CN104809443B (en) Detection method of license plate and system based on convolutional neural networks
DE102017107264A1 (en) Visible rain recognition using deep learning
US6697504B2 (en) Method of multi-level facial image recognition and system using the same
JP4159794B2 (en) Image processing apparatus and method
US20180253866A1 (en) Image analysis neural network systems
Bart et al. Cross-generalization: Learning novel classes from a single example by feature replacement
US9053391B2 (en) Supervised and semi-supervised online boosting algorithm in machine learning framework
CN105184271A (en) Automatic vehicle detection method based on deep learning
Koetsier et al. Kernel maximum likelihood hebbian learning
US8175333B2 (en) Estimator identifier component for behavioral recognition system
JP2004252940A (en) System and method for detecting object in image
EP2518661A2 (en) System and method for human detection and counting using background modeling, hog and haar features
US7280697B2 (en) Unsupervised learning of object categories from cluttered images
US9501703B2 (en) Apparatus and method for recognizing traffic sign board
Huwer et al. Adaptive change detection for real-time surveillance applications
EP2189955B1 (en) Sensor network system, transfer protocol, method for recognising an object and computer program
US20080152236A1 (en) Image processing method and apparatus
Zhang et al. Detection and tracking of multiple humans with extensive pose articulation

Legal Events

Date Code Title Description
OP8 Request for examination as to paragraph 44 patent law
8364 No opposition during term of opposition
8339 Ceased/non-payment of the annual fee