Subclass partitioning in a pattern recognition classifier system

• US 20070168941A1
• Filed: 12/19/2005
• Published: 07/19/2007
• Est. Priority Date: 12/19/2005
• Status: Active Grant

• Current Assignee: TRW Automotive USLLC (Zeppelin - Stiftung der Stadt Friedrichshafen)
• Original Assignee: TRW Automotive USLLC (Zeppelin - Stiftung der Stadt Friedrichshafen)
• Inventors: Luo, Yun
• Granted Patent: US 7,483,866 B2
• Time in Patent Office: Days
• US Class Current: 717/108
• CPC Class Codes: G06V 40/103 Static body considered as a whole, e.g. static pedestrian or occupant recognition

Abstract

Systems and methods are provided for constructing a classification architecture for a pattern recognition classifier that determines an associated occupant class of an occupant within a vehicle. A plurality of occupant classes are defined, with at least one of the occupant classes having an associated set of subclasses (202). A hierarchy of subclass partitions is determined based on a degree of similarity between subclasses associated with a given occupant class (210). Each successive partition in the hierarchy is an immediate partition of the preceding partition. The following is then performed iteratively until a termination event occurs. A subclass partition is selected according to the determined hierarchy (216/230). The fitness of a classifier model having an architecture derived from the selected subclass partition is evaluated (220). A class separation associated with the selected subclass partition is retained when the evaluated fitness is better than that of the preceding subclass partition in the hierarchy (224).

Claims

1. A method for constructing a classification architecture for a pattern recognition classifier that determines an associated occupant class of an occupant within a vehicle comprising:

defining a plurality of occupant classes, at least one of the occupant classes having an associated set of subclasses;

determining a hierarchy of subclass partitions, based on a degree of similarity between subclasses associated with a given occupant class, where each successive partition in the hierarchy is an immediate partition of the preceding partition; and

performing the following steps iteratively until a termination event occurs;

selecting a subclass partition according to the determined hierarchy;

evaluating the fitness of a classifier model having an architecture derived from the selected subclass partition; and

retaining a class separation associated with the selected subclass partition when the evaluated fitness is better than that associated with the preceding subclass partition in the hierarchy.

2. The method of claim 1, wherein the iteratively performed steps further comprise the step of eliminating the class separation associated with the selected subclass partition as well as any class separations subordinate to the class separation associated with the selected subclass partition when the evaluated fitness is not better than that of the preceding subclass partition in the hierarchy.

3. The method of claim 2, wherein the termination event comprises the elimination of all class separations that have not been evaluated.

4. The method of claim 1, wherein the step of determining a hierarchy of subclass partitions comprises the steps of:

determining a minimum partition, comprising the occupant classes, and a maximum partition, comprising a plurality of subclasses associated with the occupant classes;

training a classifier having an architecture derived from the maximum partition of the classes;

classifying a verification set at the classifier; and

generating a confusion matrix for each occupant class according to the classification of the verification set.

5. The method of claim 4, wherein the step of determining a hierarchy of subclass partitions further comprises:

generating a similarity matrix for each occupant class from the confusion matrix generated for the occupant class;

determining a hierarchy of class separations for each occupant class; and

constructing a search tree, reflecting the hierarchy of class partitions, from the determined hierarchies of class separations for the occupant classes.

6. The method of claim 1, further comprising the step of training a classifier according to the determined classification architecture.

7. The method of claim 1, wherein the plurality of classes comprise an adult occupant class, a small occupant class, and an empty seat class.

8. The method of claim 7, wherein a set of subclasses associated with the small occupant class include at least one child subclass and sat least one child seat subclass.

9. The method of claim 7, wherein a set of subclasses associated with the adult occupant class includes a first subclass representing an adult in the ninety-fifth percentile of one of height and weight, a second subclass representing an adult in the fiftieth percentile of one of height and weight, and a third subclass representing an adult in the fifth percentile in one of height and weight.

10. The method of claim 1, wherein the step of evaluating the fitness of a classifier model comprises the steps of:

training a classifier having an architecture derived from the maximum partition of the classes;

classifying a verification set at the classifier; and

evaluating the performance of the classifier on the verification set.

11. A computer program product, operative in a data processing system and implemented on a computer readable medium, for establishing a classification architecture, comprising a plurality of operative classes, to distinguish between a plurality of occupant classes comprising:

a partition search component that determines a hierarchy of subclass partitions, based on a degree of similarity between subclasses associated with the plurality of occupant classes; and

a system control that iteratively selects a subclass partition according to the determined hierarchy, evaluates the fitness of a pattern recognition classifier having an architecture derived from the selected subclass partition, and retains a class separation associated with the selected subclass partition when the evaluated fitness is better than that associated with a preceding subclass partition in the hierarchy.

12. The computer program product of claim 11, wherein the pattern recognition classifier comprises an artificial neural network.

13. The computer program product of claim 11, wherein the pattern recognition classifier comprises a support vector machine.

14. The computer program product of claim 11, wherein the partition search component is operative to generate a confusion matrix for each occupant class using a maximum subclass partition, the confusion matrices representing the plurality of occupant classes being utilized to determine a hierarchy of subclass partitions.

15. The computer program product of claim 14, wherein the partition search component is operative to generate a similarity matrix for each occupant class from the confusion matrix generated for the occupant class, determine a hierarchy of class separations for each occupant class from distance values associated with the similarity matrix associated with the class, and construct a search tree, reflecting the hierarchy of class partitions, from the determined hierarchies of class separations for the occupant classes.

16. A system for establishing a classification architecture, comprising a plurality of operative classes, for a pattern recognition classifier operative to distinguish among a plurality of occupant classes as part of a vehicle occupant protection system, comprising:

a partition search component that determines a hierarchy of subclass partitions, based on a degree of similarity between subclasses associated with the plurality of occupant classes; and

a system control that iteratively selects a subclass partition according to the determined hierarchy, evaluates the fitness of a pattern recognition classifier having an architecture derived from the selected subclass partition, and retains a class separation associated with the selected subclass partition when the evaluated fitness is better than that associated with a preceding subclass partition in the hierarchy.

17. The system of claim 16, where the system control is operative to eliminate the class separation associated with the selected subclass partition as well as any class separations subordinate to the class separation associated with the selected subclass partition when the evaluated fitness is not better than that of the preceding subclass partition in the hierarchy.

18. The system of claim 16, further comprising a system memory that contains a set of training samples and a set of verification samples for the pattern recognition classifier, the system control being operative to train the pattern recognition classifier on the set of training samples, instruct the classifier to classify the set of verification samples, and evaluate the performance of the classifier on the verification set.

19. The system of claim 18, wherein the set of training samples comprises at least one of a plurality of images from a camera associated with the vehicle occupant protection system and a plurality of readings from an array of weight sensors associated with a vehicle seat.

20. The system of claim 16, wherein the hierarchy of subclass partitions proceeds from a minimum partition of the subclasses, in which each of the plurality of operative classes corresponds to an occupant class, through a series of immediate partitions to a maximum partition, in which each of the plurality of operative classes corresponds to one of a plurality of subclasses associated with the plurality of occupant classes.