ROBOT WITH VISION-BASED 3D SHAPE RECOGNITION

• US 20100286827A1
• Filed: 05/05/2010
• Published: 11/11/2010
• Est. Priority Date: 05/08/2009
• Status: Active Grant

• Current Assignee: Honda Research Institute Europe GmbH (Honda Motor Company)
• Original Assignee: Honda Research Institute Europe GmbH (Honda Motor Company)
• Inventors: FRANZIUS, Mathias, WERSING, Heiko
• Granted Patent: US 8,731,719 B2
• Time in Patent Office: Days
• US Class Current: 700/259
• CPC Class Codes: G06F 18/213 Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
• CPC Class Codes: G06V 10/7715 Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
• CPC Class Codes: G06V 20/64 Three-dimensional objects
• CPC Class Codes: G06V 2201/06 Recognition of objects for industrial automation

Abstract

The invention relates to a method for processing video signals from a video sensor, in order to extract 3d shape information about objects represented in the video signals, the method comprising the following steps:

• • providing a memory in which objects are stored in a 3d shape space, the shape space being an abstract feature space encoding the objects'"'"' 3d shape properties, and
  • mapping a 2d video signal representation of an object in the shape space, the coordinates of the object in the shape space indicating the object'"'"'s 3d shape.

Claims

1. A method for processing video signals from a video sensor, in order to extract 3d shape information and/or information on at least one physical property about objects represented in the video signals, the method comprising the following steps:

providing a memory in which objects are stored in a 3d shape space, the shape space being an abstract feature space encoding the objects'"'"' 3d shape properties, and

mapping a 2d video signal representation of an object in the shape space, the coordinates of the object in the shape space indicating the 3d shape and/or a physical property of the object.

2. The method according to claim 1, wherein 2d video signals and additionally depth information on the position of objects relative to the video sensor is supplied.

3. Use of a method according to claim 1, in order to control a mechanical actuator manipulating an object dependent on the thus gained information on the object'"'"'s shape.

4. A computing unit for an autonomous robot, being designed to perform a method according to claim 1.

5. An autonomous robot, being provided with at least one video sensor providing video signals to a computing unit according to claim 4, the computing unit controlling a mechanical object manipulator of the robot.

6. A computer software program product, performing a method according to claim 1 when run on a computing device.

7. An unsupervised learning method for setting up a computing system able to extract information on the 3d shape and/or at least one physical property of objects on the basis of video signals representing said objects, the method comprising the following steps:

presenting in the input field of a 2d video camera training objects respectively in different positions or the trajectory of an object upon a defined motion-inducing stimulus,

extracting slowly varying features of different rotational views of the training objects and clustering the extracted features in order to parameterize a shape space representation of the training objects, the shape space being an abstract feature space encoding the objects'"'"' 3d shape properties.

8. The method according to claim 7, being followed by a supervised learning method.

9. A method for processing video signals from a video sensor comprising the steps of:

providing a shape feature space in a memory structure,

presenting to the video sensor training objects in different positions or a trajectory of an object for setting up the shape feature space,

extracting at least one feature form each training object,

calculating feature clusters from the extracted features in the shape feature space, so that different views of the same object produce similar features belonging to one cluster, while similar views of distinct objects are separated,

identifying the feature cluster centers,

obtaining a view of an object from a video signal,

mapping the view of the object to the shape feature space, the coordinates of the object in the shape feature space indicating the object'"'"'s shape, and

characterizing the object'"'"'s shape as a shape similarity represented by distances to the feature cluster centers.

10. The method of claim 9, comprising further the step of:

extracting slowly varying features of different rotational views of the training objects and clustering the extracted features in order to parameterize the shape feature space representation of the training objects, the shape feature space being an abstract feature space encoding the objects'"'"' 3d shape properties.