System and method of fast object detection using parts to whole fragment detection

Patent metadata

Field	Value
Publication number	US-9400945-B2
Application number	US-201113241908-A
Country	US
Kind code	B2
Filing date	Sep 23, 2011
Priority date	Sep 23, 2011
Publication date	Jul 26, 2016
Grant date	Jul 26, 2016

Abstract

A system and method may compare an image vector representing an image feature of a first image fragment of an image to database vectors representing the image feature of database image fragments of database images. It may be determined based on the comparison a first matching database vector of the database vectors which most closely, among the database vectors, describes the first image feature represented by the image vector. The system or method may determine, using a data structure in conjunction with the first matching database vector and previously matched database vectors, a second of the database vectors which includes the first matching database vector and the previously matched database vectors and most closely describes a second image fragment including the first image fragment. The system or method may determine an object feature based on the second database vector.

First claim

What is claimed is:

1. A method for object detection in an image comprising:
providing a hierarchal database of database-vectors, representing features of objects' fragments of database-objects, the database including a list of the database-vectors and their containing database-vectors; and
using a processor for:
identifying at least two first database-vectors from the database of vectors, which are most closely matching with two or more image-vectors describing features of at least two respective first image-fragments of the image, wherein the at least two first image-fragments are different one from another or partially overlapping, and
detecting at least one object of the database-objects, responsive to identifying of a second database-vector from the database of vectors, wherein the second database-vector:
comprises at least two of the first database-vectors, and
is most closely matching with one or more image-vectors describing features of a second image-fragment of the image, further wherein the second image fragment comprises at least one of the at least two first image-fragments.

2. The method of claim 1 , wherein the processor is further configured to render the database-vectors into codewords.

3. The method of claim 1 , wherein the database comprises hash table or k-d tree for the database-vectors.

4. The method of claim 1 , wherein the identifying of the most closely first and/or second database-vectors is implemented using an approximate nearest-neighbor structure.

5. A system for object detection in an image comprising:
a memory loaded with an hierarchal database of database-vectors, representing features of objects' fragments of database-objects, the database including a list of the database-vectors and their containing the database-vectors; and
a processor configured to:
identify at least two first database-vectors from the database of vectors, which are most closely matching with two or more image-vectors describing features of at least two respective first image-fragments of the image, wherein the at least two first image-fragments are different one from another or partially overlapping, and
detect at least one object of the database-objects, responsive to identification of a second database-vector from the database of vectors, wherein the second database-vector:
comprises at least two of the first database-vectors, and
is most closely matching with one or more image-vectors describing features of a second image-fragment of the image, further wherein the second image fragment comprises at least one of the at least two first image-fragments.

6. The system of claim 5 , wherein the processor is further configured to render the database-vectors into codewords.

7. The system of claim 5 , wherein the identification of the most closely first and/or second database-vectors is implemented using an approximate nearest-neighbor structure.

8. The system of claim 5 , wherein the database comprises hash table or k-d tree for the database-vectors.

Assignees

• Bar Hillel Aharon
• Levi Dan
• Gm Global Tech Operations Llc

Inventors

• Bar Hillel Aharon
• Levi Dan

Classifications

• G06V10/75 Primary

  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 title

• G06K9/68 Primary

  Physics · mapped topic