Optimized object detection

Patent metadata

Field	Value
Publication number	US-9727800-B2
Application number	US-201514866781-A
Country	US
Kind code	B2
Filing date	Sep 25, 2015
Priority date	Sep 25, 2015
Publication date	Aug 8, 2017
Grant date	Aug 8, 2017

Abstract

Disclosed are a system, apparatus, and method for detecting objects. Input image frames may be received and distinct regions created within each frame. Descriptors may be extracted from the regions according to their associated probability. Extracted descriptors may be matched to reference descriptors. Votes or confidence is cast for particular regions according to region properties. The region properties may be determined from center voting methods based on vector intersection to other vectors or intersections with a region. The probability of selecting particular regions can increase with each vote or increase in confidence for a region. In response to updating probabilities, additional regions may be selected and additional descriptors may be extracted. Additional voting iterations can update the probability of selecting a next region. An object pose may be estimated in response to meeting one or more thresholds.

First claim

What is claimed is:

1. A method for detecting a target object, the method comprising:
receiving an input image frame;
creating a plurality of distinct regions within the input image frame, wherein the plurality of distinct regions has a first probability distribution comprising of a particular probability of being selected for each respective region in the plurality of distinct regions;
selecting, according to the first probability distribution, a first region from the plurality of distinct regions;
extracting, from the first region, a first one or more descriptors from a first one or more feature points detected within the first region;
analyzing one or more properties of the first region to update a particular probability of the first region being selected;
creating a second probability distribution for the plurality of distinct regions according to the particular probability of the first region being selected;
selecting, according to the second probability distribution, a second region from the plurality of distinct regions;
extracting, from the second region, a second one or more descriptors from a second one or more feature points detected within the second region; and
referencing at least the first one or more descriptors and the second one or more descriptors for detecting the target object.

2. The method of claim 1 , further comprising:
continuing, according to updated probability distributions for the plurality of regions, the selecting of regions, extracting of descriptors, and analyzing of one or more properties until detecting one or both of:
a threshold number of descriptors are extracted,
a threshold confidence is met for confirmation of detecting the target object, or
a threshold confidence is met for determining the target object is not present.

3. The method of claim 1 , wherein the analyzing the one or more properties comprises:
calculating, for one or more feature points in a respective region, a direction vector representing a scale and a direction to a center of the target object determined from a reference database; and
increasing probability of selection for the respective region when two or more direction vectors within the respective region intersect.

4. The method of claim 1 , wherein the analyzing the one or more properties comprises:
calculating, for one or more feature points in a respective region, a direction vector representing a scale and a direction to a center of the target object; and
increasing probability of selection for the respective region when the direction vector intersects the respective region.

5. The method of claim 1 , wherein the analyzing the one or more properties comprises:
matching the extracted one or more descriptors to a predetermined reference descriptor associated with the target object; and
increasing probability of selection for the respective region in response to the matching.

6. The method of claim 1 , wherein, in response to meeting a threshold confidence for detecting the target object:
determining the target object pose; and
ignoring an area comprising the detected target object when performing additional target object detection in a next input image.

7. The method of claim 1 , wherein, in response to meeting a threshold confidence for determining the target object is not present:
applying a latest probability distribution associated with the plurality of regions to a next input image; and
resetting a probability distribution associated with the plurality of regions after one or both of: a threshold number of input images, or set amount of time.

8. The method of claim 1 , wherein a threshold number of descriptors to extract is determined according to one or both of: a per input image descriptor threshold, or per region descriptor threshold.

9. The method of claim 1 , wherein a size of each of the plurality of regions is determined according to one of: equal area split, superpixel segmentation, or a depth map.

10. The method of claim 1 , further comprising:
increasing the probability of selection for one or more adjacent regions to a respective region in response to having confidence for the respective region.

11. A machine readable non-transitory storage medium having stored therein program instructions that are executable by a processor to:
receive an input image frame;
create a plurality of distinct regions within the input image frame, wherein the plurality of distinct regions has a first probability distribution comprising of a particular probability of being selected for each respective region in the plurality of distinct regions;
select, according to the first probability distribution, a first region from the plurality of distinct regions;
extract, from the first region, a first one or more descriptors from a first one or more feature points detected within the first region;
analyze one or more properties of the first region to update a particular probability of the first region being selected;
create a second probability distribution for the plurality of distinct regions according to the particular probability of the first region being selected;
select, according to the second probability distribution, a second region from the plurality of distinct regions;
extract, from the second region, a second one or more descriptors from a second one or more feature points detected within the second region; and
reference at least the first one or more descriptors and the second one or more descriptors for detecting a target object.

12. The medium of claim 11 , further comprising instructions to:
continue, according to updated probability distributions for the plurality of regions, the selecting of regions, extracting of descriptors, and analyzing of one or more properties until detecting one or both of:
a threshold number of descriptors are extracted,
a threshold confidence is met for confirmation of detecting the target object, or
a threshold confidence is met for determining the target object is not present.

13. The medium of claim 11 , wherein the instructions to analyze the one or more properties comprises instructions to:
calculate, for one or more feature points in a respective region, a direction vector representing a scale and a direction to a center of the target object determined from a reference database; and
increase probability of selection for the respective region when two or more direction vectors within the respective region intersect.

14. The medium of claim 11 , wherein the instructions to analyze the one or more properties comprises instructions to:
calculate, for one or more feature points in a respective region, a direction vector representing a scale and a direction to a center of the target object; and
increase probability of selection for the respective region when the direction vector intersects the respective region.

15. The medium of claim 11 , wherein the instructions to analyze the one or more properties comprises instructions to:
match the extracted one or more descriptors to a predetermined reference descriptor associated with the target object; and
increase probability of selection for the respective region in response to the matching.

16. The medium of claim 11 , wherein, in response to meeting a threshold confidence for detecting the target object:
determining the target object pose; and
ignoring an area comprising the detected target object when performing additional target object detection in a next input image.

17. The medium of claim 11 , wherein, in response to meeting a threshold confidence for determin

Assignees

• Qualcomm Inc

Inventors

• Pan Qi
• Tallonneau Romain
• Maggio Emilio
• Walk Stefan

Classifications

• G06V10/48 Primary

  by mapping characteristic values of the pattern into a parameter space, e.g. Hough transformation · CPC title

• G06F18/2415

  based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate · CPC title

• G06V10/25

  Determination of region of interest [ROI] or a volume of interest [VOI] · CPC title

• G06V10/26

  Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion · CPC title

• G06V10/462

  Salient features, e.g. scale invariant feature transforms [SIFT] · CPC title