Object detection method, object detection device, and program

Patent metadata

Field	Value
Publication number	US-12475676-B2
Application number	US-202118000338-A
Country	US
Kind code	B2
Filing date	May 24, 2021
Priority date	Jun 5, 2020
Publication date	Nov 18, 2025
Grant date	Nov 18, 2025

Abstract

An object detection method includes a key point estimation step of estimating key point candidates for each object in an image; and a detection step of detecting key points for each object based on the estimated key point candidates. Considering an object model that models shape of an object, the key points are points that satisfy a defined condition among points indicating a boundary of the object model that are projected onto defined coordinate axes. The defined coordinate axes have an origin at a geometric center of the object model and each forms a defined angle relative to a polar axis in a polar coordinate system set for the object model.

First claim

The invention claimed is:

1 . An object detection method for detecting each object in an image containing one or more objects of one or more defined categories, comprising:
a key point estimation step of estimating key point candidates for each object in the image; and a detection step of detecting key points for each object based on the estimated key point candidates, wherein considering an object model that models shape of an object, the key points are points that satisfy a defined condition among points indicating a boundary of the object model that are projected onto defined coordinate axes, the defined coordinate axes have an origin at a geometric center of the object model and each forms a defined angle relative to a polar axis in a polar coordinate system set for the object model, and two key points of the key points are defined for each of the coordinate axes, the two key points for each of the coordinate axes are a point on the boundary of the object having a maximum value and a point on the boundary of the object having a minimum value in a positive range of the each of the coordinate axes.

2 . The object detection method of claim 1 , wherein
the key points are selected as points on the object having a local maximum or local minimum value, and the two key points for each of the coordinate axes are determined from the points on the object having a local maximum or local minimum value.

3 . The object detection method of claim 1 , further comprising
a center position estimation step of estimating a center candidate for each object in the image and a confidence indicating likelihood of accurate estimation, wherein the detection step uses the confidence to detect a center position of each object from the center candidates, and uses each detected center position in detection of the key points of each object from the key point candidates.

4 . The object detection method of claim 3 , wherein
the key point estimation step and the center position estimation step are executed by a machine-learning model trained to detect each object.

5 . The object detection method of claim 1 , wherein
the key point estimation step estimates the key point candidates as areas each having a size based on a corresponding object of the one or more objects.

6 . The object detection method of claim 1 , wherein
the key point estimation step is executed by a machine-learning model trained to detect each object.

7 . The object detection method of claim 6 , wherein
the machine-learning model is a convolutional neural network, and parameters of the convolutional neural network are defined by machine-learning based on a training image including a detection target object, a true value of a center position of the detection target object in the training image, and a true value of a key point of the detection target object in the training image.

8 . The object detection method of claim 1 , wherein each of the key points that satisfies the defined condition is a point on the surface of the object model that is a point of a local maximum coordinate value or a local minimum coordinate value among points on the surface of the object model such that the each key point protrudes from other portions of the object or is recessed from other portions of the object.

9 . The object detection method of claim 8 , further comprising
a center position estimation step of estimating a center candidate for each object in the image and a confidence indicating likelihood of accurate estimation, wherein the detection step uses the confidence to detect a center position of each object from the center candidates, and uses each detected center position in detection of the key points of each object from the key point candidates.

10 . The object detection method of claim 9 , wherein
the key point estimation step and the center position estimation step are executed by a machine-learning model trained to detect each object.

11 . The object detection method of claim 8 , wherein
the key point estimation step estimates the key point candidates as areas each having a size based on a corresponding object of the one or more objects.

12 . The object detection method of claim 8 , wherein
the key point estimation step is executed by a machine-learning model trained to detect each object.

13 . The object detection method of claim 12 , wherein
the machine-learning model is a convolutional neural network, and parameters of the convolutional neural network are defined by machine-learning based on a training image including a detection target object, a true value of a center position of the detection target object in the training image, and a true value of a key point of the detection target object in the training image.

14 . An object detection device for detecting each object in an image containing one or more objects of one or more defined categories, comprising:
a machine-learning model trained to detect each object, which executes a key point estimation process of estimating key point candidates for each object in the image; and a detection unit that detects key points for each object based on the estimated key point candidates, wherein considering an object model that models shape of an object, the key points are points that satisfy a defined condition among points indicative of a boundary of the object model that are projected onto a defined coordinate axis, and the defined coordinate axes have an origin at a geometric center of the object model and each forms a defined angle relative to a polar axis in a polar coordinate system set for the object model, and two key points of the key points are defined for each of the coordinate axes, the two key points for each of the coordinate axes are a point on the boundary of the object having a maximum value and a point on the boundary of the object having a minimum value in a positive range of the each of the coordinate axes.

15 . A non-transitory computer readable medium storing a program causing a computer to execute object detection processing for detecting each object in an image containing one or more objects of one or more defined categories, wherein
the object detection processing comprises: a key point estimation step of estimating key point candidates for each object in the image; and a detection step of detecting key points for each object based on the estimated key point candidates, wherein considering an object model that models shape of an object, the key points are points that satisfy a defined condition among points indicative of a boundary of the object model that are projected onto a defined coordinate axis, and the defined coordinate axes have an origin at a geometric center of the object model and each forms a defined angle relative to a polar axis in a polar coordinate system set for the object model, and two key points of the key points are defined for each of the coordinate axes, the two key points for each of the coordinate axes are a point on the boundary of the object having a maximum value and a point on the boundary of the object having a minimum value in a positive range of the each of the coordinate axes.

Assignees

• Konica Minolta Inc

Inventors

• Sekii Taiki

Classifications

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06V10/82

  using neural networks · CPC title

• G06T7/75

  involving models · CPC title