Target object classification using three-dimensional geometric filtering

Patent metadata

Field	Value
Publication number	US-10242294-B2
Application number	US-201715582864-A
Country	US
Kind code	B2
Filing date	May 1, 2017
Priority date	May 1, 2017
Publication date	Mar 26, 2019
Grant date	Mar 26, 2019

Abstract

An example apparatus for classifying target objects using three-dimensional geometric filtering includes a patch receiver to receive patches with objects to be classified. The apparatus also includes a geometric filter to filter out patches including objects with sizes outside a target range using three dimensional geometry to generate filtered patches. The apparatus further includes a background remover to remove background pixels from the filtered patches to generate preprocessed patches. The apparatus includes a classification score calculator to calculate a classification score for each of the preprocessed patches.

First claim

What is claimed is:

1. An apparatus for classifying target objects using three-dimensional geometric filtering, comprising:
a patch receiver to receive patches with objects to be classified;
a geometric filter to filter out patches corresponding to objects with sizes outside a target range of object sizes using three dimensional geometry to generate filtered patches;
a background remover to remove background pixels from the filtered patches to generate preprocessed patches without the removed background pixels;
a classification score calculator to calculate a classification score for each of the preprocessed patches;
a target object detector to detect a target object in response to detecting a classification score exceeding a threshold score; and
a virtual target object displayer to display the detected target object as a virtual target object in a virtual reality display in real time.

2. The apparatus of claim 1 , comprising a score averager to average classification scores received over a predetermined time for each of the preprocessed patches.

3. The apparatus of claim 1 , wherein the target object comprises a human hand.

4. The apparatus of claim 1 , wherein the background remover is to remove background pixels based on an image segmentation.

5. The apparatus of claim 1 , wherein the target object range comprises lower target object bounds and upper target object bounds for a volume and a surface area of a bounding to be calculated for an object in each patch.

6. The apparatus of claim 1 , wherein the background remover comprises a convolutional neural network trained to remove background noise based on a depth image segmentation.

7. The apparatus of claim 1 , wherein the classification score calculator comprises a convolutional neural network trained to classify patches using a set of training images of target objects.

8. The apparatus of claim 1 , wherein the apparatus comprises a pipelined image processor.

9. A method for classifying target objects in images, comprising:
receiving, via a processor, patches with objects to be classified;
filtering out, via the processor, patches corresponding to objects with sizes outside a target range using three dimensional geometry to generate filtered patches;
removing, via the processor, background pixels from the filtered patches to generate preprocessed patches without the removed background pixels;
calculating, via the processor, a classification score for each of the preprocessed patches;
detecting, via the processor, a target object in response to detecting a classification score exceeding a threshold score; and
displaying, via the processor, the detected target object as a virtual target object in a virtual reality display in real time.

10. The method of claim 9 , comprising averaging, via the processor, classification scores received over a predetermined time for each of the preprocessed patches.

11. The method of claim 9 , wherein using three dimensional geometry to generate filtered patches comprises calculating, via the processor, a bounding box for each patch using a principal component analysis and comparing, via the processor, a volume and a surface area of the bounding box for each patch to a lower bounding box threshold and a higher bounding box threshold.

12. The method of claim 9 , wherein removing the background pixels comprises segmenting the image.

13. At least one computer readable medium for classifying target objects using three-dimensional geometric filtering having instructions stored therein that, in response to being executed on a computing device, cause the computing device to:
receive patches with objects to be classified;
filter out patches corresponding to objects with sizes outside a target range of object sizes using three dimensional geometry to generate filtered patches;
remove background pixels from the filtered patches using image segmentation to generate preprocessed patches without the removed background pixels;
calculate a classification score indicating a probability that a patch is a target object for each of the preprocessed patches;
detect a target object in response to detecting a classification score exceeding a threshold score; and
display the detected target object as a virtual target object in a virtual reality display in real time.

14. The at least one computer readable medium of claim 13 , comprising instructions to average classification scores received over a predetermined time for each of the preprocessed patches.

15. The at least one computer readable medium of claim 13 , comprising instructions to detect a target object in response to detecting an average classification score of a preprocessed patch exceeds a threshold classification score.

16. The at least one computer readable medium of claim 13 , wherein the target object comprises a human hand.

Assignees

• Intel Corp

Inventors

• Eldar Avigdor
• Menadeva Ovadya
• Viente Kfir

Classifications

• G06T2207/10028

  Range image; Depth image; 3D point clouds · CPC title

• G06T7/11

  Region-based segmentation · CPC title

• G06T7/194

  involving foreground-background segmentation · CPC title

• G06T7/187

  involving region growing; involving region merging; involving connected component labelling · CPC title

• G06V10/82 Primary

  using neural networks · CPC title