Method for detecting object

• US 8,533,162 B2
• Filed: 08/16/2010
• Issued: 09/10/2013
• Est. Priority Date: 08/18/2009
• Status: Expired due to Fees

• Current Assignee: Osaka Prefcture University
• Original Assignee: Osaka Prefcture University
• Inventors: Kise, Koichi, Maus, Heiko, Dengel, Andreas, Klinkigt, Martin
• Primary Examiner(s): Coby, Frantz
• Application Number:
  US13/386,924
• Publication Number:
  US 20120143856A1
• Time in Patent Office: 1,121 Days
• Field of Search: 707/669, 707/705, 707/687, 707/769, 707/758, 707/748, 707/749, 382/253, 382/190, 382/199, 706/14, 345/166, 358/583
• US Class Current: 707/669
• CPC Class Codes: G06F 18/22 Matching criteria, e.g. proximity measures
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]
• CPC Class Codes: G06V 10/75 Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
• CPC Class Codes: G06V 10/757 Matching configurations of points or features
• CPC Class Codes: G06V 10/759 Region-based matching

Abstract

An object detection method that is provided with a step for extracting a plurality of reference feature vectors related to a local area from an image representing an object, and extracting a plurality of query feature vectors related to the local area from a search query image; a step for matching each query feature vector against each reference feature vector, and calculating a similarity score having a value that is higher the closer the distance between both vectors, the larger the local area for which the query feature vector has been extracted, and the larger the local area for which a matching reference feature vector has been extracted; a step for determining a reference feature vector for which a similarity score is highest as the similar vector for each query feature vector; and a step for acquiring a final score by object associated with the similar vectors, and setting the object returning the highest score as the detection result; and wherein the score is calculated by dividing a sum of the similarity score for each similar vector by the number of feature vectors that have matched the object.

Claims

1. A method of detecting an object through a retrieve of an image representing an object matched to an object represented by a query image from an image database that includes a plurality of reference images and is created by extracting a plurality of reference feature vectors from each reference image that represents an object, each of the reference feature vectors representing a feature of one of local areas in each reference image, and by storing the reference feature vectors in such a manner that each of the reference feature vectors is associated with the corresponding reference image and with an object ID that identifies the object represented by each reference image, the method comprising the steps of:

extracting a plurality of vectors from the query image as query feature vectors, each of the query feature vectors representing a feature of one of local areas in the query image;

comparing each query feature vector with each reference feature vector, and calculating a similarity score that is determined to have a greater value in case where the query feature vector and the reference feature vector are closer, in case where a local area from which the query feature vector is extracted is greater, and in case where a local area from which the reference feature vector is extracted is greater;

determining a reference feature vector which provides the highest similarity score as a similar vector for each query feature vector; and

obtaining resulting scores through a predetermined calculation procedure according to the respective object IDs, each object ID being associated with the similar vector, and determining, as a detection result, at least one object that is specified by an object ID giving the highest resulting score,

wherein each resulting score is calculated according to the following equation ;

2. The method according to claim 1, wherein

the similarity score is calculated according to the following equation:

3. The method according to claim 1, wherein

the image database further stores distances and orientations from a predetermined reference point with respect to respective stored images, the distances and orientations being associated with the corresponding reference feature vector extracted from the image,

the method further comprising the steps of:

determining locations of imaginary reference points to the similar vectors for each similar vector by using the distance and orientation, which are stored as being associated with the corresponding reference feature vector;

determining a plurality of clusters by collecting the similar vectors, which have imaginary reference points close to one another, and with which the same object ID is associated; and

obtaining a locality score, which represents a degree of similarity of each cluster obtained as a result of the step of determining clusters, through a predetermined process,

wherein each resulting score is calculated according to the following equation instead of the equation (1);

4. The method according to claim 3, wherein the locality score r_c is defined as the total number of the imaginary reference points included in each cluster.

5. The method according to claim 3, further comprising the steps of:

obtaining a shape context according to respective feature vectors by applying a procedure of a shape context that is defined as a histogram obtained by dividing a region around one feature vector into a plurality of sub-regions as bins, and counting a number of other feature vectors belonging to the respective bins, each of the other feature vectors being distributed around the feature vector,

calculating a context score through a predetermined procedure by using the obtained shape context, and

calculating the context score according to the query feature vector by applying the shape context to each query feature vector,

wherein each resulting score is calculated according to the following equation instead of the equation (3);

6. The method according to claim 5, wherein

the context score is defined as a length of a vector by coupling elements of respective representative vectors,

7. The method according to claim 2, wherein

the image database further stores distances and orientations from a predetermined reference point with respect to respective stored images, the distances and orientations being associated with the corresponding reference feature vector extracted from the image,

the method further comprising the steps of:

determining locations of imaginary reference points to the similar vectors for each similar vector by using the distance and orientation, which are stored as being associated with the corresponding reference feature vector;

determining a plurality of clusters by collecting the similar vectors, which have imaginary reference points close to one another, and with which the same object ID is associated; and

obtaining a locality score, which represents a degree of similarity of each cluster obtained as a result of the step of determining clusters, through a predetermined process,

wherein each resulting score is calculated according to the following equation instead of the equation (1);

8. The method according to claim 4, further comprising the steps of:

obtaining a shape context according to respective feature vectors by applying a procedure of a shape context that is defined as a histogram obtained by dividing a region around one feature vector into a plurality of sub-regions as bins, and counting a number of other feature vectors belonging to the respective bins, each of the other feature vectors being distributed around the feature vector,

calculating a context score through a predetermined procedure by using the obtained shape context, and

calculating the context score according to the query feature vector by applying the shape context to each query feature vector,

wherein each resulting score is calculated according to the following equation instead of the equation (3);