Object recognition system, position information acquisition method, and program

1. An object recognition system, comprising at least one processor configured to:

input an input image to a learning model for recognizing a first object;

acquire a heat map about a position that is a basis for recognition of the first object by the learning model having the input image input thereto based on a Class Activation Map method;

acquire a motion image about a position of a moving body that is included in the input image based on motion detection processing;

acquire another image by retaining a part of the input image in which a sum of a pixel value in the heat map and a pixel value in the motion image is equal to or more than a threshold value; and

acquire position information about a position of the first object included in the input image, based on the heat map and the motion image.

2. The object recognition system according to claim 1 ,

wherein the first information is information by which a value indicating a strength as a basis for recognition by the learning model is identifiable for each pixel of the input image, and

wherein the second information is information by which a value indicating a strength as the moving body or the foreground is identifiable for each pixel of the input image.

3. The object recognition system according to claim 2 , wherein the at least one processer is configured to acquire the position information based on a result of adding together, with a use of a predetermined mathematical expression, a value identified from the first information as a value that is associated with a pixel or pixels in the input image and a value identified from the second information as a value that is associated with a pixel or pixels in the input image.

4. The object recognition system according to claim 1 ,

wherein the at least one processer is configured to acquire third information about a position of a second object, which is included in the input image and which grasps the first object, and

wherein the at least one processer is configured to acquire the position information based further on the third information.

5. The object recognition system according to claim 4 ,

wherein the first information is information by which a value indicating a strength as a basis for recognition by the learning model is identifiable for each pixel of the input image,

wherein the second information is information by which a value indicating a strength as the moving body or the foreground is identifiable for each pixel of the input image, and

wherein the third information is information by which a value indicating a strength as the position of the second object is identifiable for each pixel of the input image.

6. The object recognition system according to claim 5 , wherein the at least one processer is configured to acquire the position information based on a result of adding together, with a use of a predetermined mathematical expression, a value identified from the first information as a value that is associated with a pixel or pixels in the input image and a value identified from the second information as a value that is associated with a pixel or pixels in the input image, and subtracting, with a use of a predetermined mathematical expression, a value identified from the third information as a value that is associated with a pixel or pixels in the input image.

7. The object recognition system according to claim 1 , wherein the at least one processer is configured to acquire the position information based further on weighting that is heavier in the heat map than the motion image.

8. The object recognition system according to claim 1 ,

wherein the position information indicates an area including the first object, and

wherein the at least one processer is configured to acquire a plurality of candidates for the area, and acquire the position information based on one of the plurality of candidates that is largest in size.

9. The object recognition system according to claim 1 ,

wherein the input image includes a plurality of types of objects each of which is the first object, and

wherein the at least one processer is configured to acquire the position information for each object type out of the plurality of types of objects included in the input image.

10. The object recognition system according to claim 1 , wherein the at least one processer is configured to execute, based on teaching data that is generated based on the input image and the position information, learning processing of another learning model for recognizing a position of the first object.

11. The object recognition system according to claim 10 ,

wherein the input image is a frame in a moving image,

wherein the at least one processer is configured to acquire the position information for each frame in the moving image, and

wherein the at least one processer is configured to execute the learning processing based on a teaching data set including a plurality of pieces of the teaching data that are generated from the moving image.

12. The object recognition system according to claim 1 ,

wherein the input image is a frame in a moving image, and

wherein the at least one processer is configured to track the first object included in the moving image, based on the position information.

13. The object recognition system according to claim 1 , wherein the learning model is a classifier configured to classify a type of the first object, or a determiner configured to determine whether the first object is present.

14. The object recognition system according to claim 1 , wherein the position information is information about a bounding box including the first object.

15. A position information acquisition method, comprising:

inputting an input image to a learning model for recognizing a first object;

acquiring a heat map about a position that is a basis for recognition of the first object by the learning model having the input image input thereto based on a Class Activation Map method;

acquiring a motion image about a position of a moving body, that is included in the input image based on motion detection processing;

acquire another image by retaining a part of the input image in which a sum of a pixel value in the heat map and a pixel value in the motion image is equal to or more than a threshold value; and

acquiring position information about a position of the first object included in the input image, based on the heat map and the motion image.

16. A non-transitory computer-readable information storage medium for storing a program for causing a computer to:

input an input image to a learning model for recognizing a first object;

acquire a heat map about a position that is a basis for recognition of the first object by the learning model having the input image input thereto based on a Class Activation Map method;

acquire a motion image about a position of a moving body, that is included in the input image based on motion detection processing;

acquire another image by retaining a part of the input image in which a sum of a pixel value in the heat map and a pixel value in the motion image is equal to or more than a threshold value; and

acquire position information about a position of the first object included in the input image, based on the heat map and the motion image.

17. The object recognition system according to claim 1 , wherein the learning model performs skin detection processing to form a skin image and subtracts the skin image from the input image.