TRAINING-FREE GENERIC OBJECT DETECTION IN 2-D AND 3-D USING LOCALLY ADAPTIVE REGRESSION KERNELS

• US 20110311129A1
• Filed: 12/16/2009
• Published: 12/22/2011
• Est. Priority Date: 12/18/2008
• Status: Active Grant

• Current Assignee: University of California Santa Cruz (University of California)
• Original Assignee: University of California Santa Cruz (University of California)
• Inventors: Milanfar, Peyman, Seo, Hae Jong
• Granted Patent: US 8,559,671 B2
• Time in Patent Office: Days
• US Class Current: 382/154
• CPC Class Codes: G06V 10/422 for representing the structure of the pattern or shape of an object therefor
• CPC Class Codes: G06V 10/50 by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
• CPC Class Codes: G06V 40/20 Movements or behaviour, e.g. gesture recognition recognition of facial expressions G06V40/16

Abstract

The present invention provides a method of learning-free detection and localization of actions that includes providing a query video action of interest and providing a target video, obtaining at least one query space-time localized steering kernel (3-D LSK) from the query video action of interest and obtaining at least one target 3-D LSK from the target video, determining at least one query feature from the query 3-D LSK and determining at least one target patch feature from the target 3-D LSK, and outputting a resemblance map, where the resemblance map provides a likelihood of a similarity between each the query feature and each target patch feature to output learning-free detection and localization of actions, where the steps of the method are performed by using an appropriately programmed computer.

Claims

1. A method of learning-free detection and localization of actions, comprising:

a. providing a query video action of interest and providing a target video by using an appropriately programmed computer;

b. obtaining at least one query space-time localized steering kernel (3-D LSK) from said query video action of interest and obtaining at least one target 3-D LSK from said target video by using said appropriately programmed computer;

c. determining at least one query feature from said query 3-D LSK and determining at least one target patch feature from said target 3-D LSK by using said appropriately programmed computer; and

d. outputting a resemblance map, wherein said resemblance map provides a likelihood of a similarity between each said query feature and each said target patch feature by using said appropriately programmed computer to output learning-free detection and localization of actions.

2. The method of detection and localization of actions in claim 1, wherein each said 3-D LSK is computed from covariance matrices, wherein said covariance matrices are estimated by using singular value decomposition (SVD), wherein said SVD is applied to a collection of space-time gradient vectors in a local neighborhood, wherein said covariance matrices determine a shape and size of a local kernel, wherein local geometric structures are encoded.

3. The method of detection and localization of actions in claim 1, wherein each resulting said 3-D LSK is invariant to illumination changes and robust to a presence of noise.

4. The method of detection and localization of actions in claim 1, wherein each said 3-D LSK is densely computed and normalized.

5. The method of detection and localization of actions in claim 4, wherein each said query feature volume and each said target feature volume comprises a principle component analysis (PCA) applied to each said densely computed and normalized query 3-D LSK, wherein said PCA step determines feature spaces for dimensionality reduction of said normalized query 3-D LSK.

6. The method of detection and localization of actions in claim 5, wherein said PCA is applied to each said densely computed and normalized query 3-D LSK and said target 3-D LSK.

7. The method of detection and localization of actions in claim 4, wherein lower dimensional features of each said query 3-D LSK and each said target 3-D LSK are determined by projecting each said normalized query 3-D LSK and each said target 3-D LSK onto said feature space.

8. The method of detection and localization of actions in claim 1, wherein said resemblance map comprises nonparametric significance testing and non-maxima suppression.

9. The method of detection and localization of actions in claim 1, wherein said resemblance map comprises a volume of voxels, wherein each said voxel indicates the likelihood of similarity between each said query feature and each said target patch feature.

10. The method of detection and localization of actions in claim 1, wherein said resemblance map is computed by using a matrix cosine similarity (MCS).

11. The method of detection and localization of actions in claim 1, wherein said motion estimation is not required to detect and localize actions of interest.

12. The method of detection and localization of action in claim 1, wherein a space-time saliency detection method is utilized to automatically extract a finite cube that only contains a valid action for the task of action category classification from the query video, wherein said space-time saliency detection is based on said 3-D LSKs and a self-resemblance measure.