Object detection device, object detection method, and computer program for object detection

1. An object detection device, comprising

a processor configured to

calculate, by inputting images acquired in time-series to a classifier that calculates, for each of a plurality of types of an object to be detected, a confidence score indicating a degree of certainty that the type of the object is represented, for each of the time-series of images, the confidence score of each of the plurality of types for a plurality of regions on the image;

detect, for each of the time-series of images, in a region in which the confidence score of any of the plurality of types is equal to or higher than a confidence threshold of the type among the plurality of regions on the image, an object of the type;

track the detected object, based on the region in which the detected object is represented in each of the time-series of images;

count a frequency of occurrence for each type of the detected object in a period in which the detected object is tracked; and

update the confidence threshold for at least any of two or more types of the object detected in the tracked period in such a way that the confidence threshold of a type of the object having the higher frequency of occurrence among the two or more types is lower than the confidence threshold of a type of the object having the lower frequency of occurrence among the two or more types, wherein

the processor uses the updated confidence threshold for a comparison with the confidence score for a region in which the detected object is estimated to be located in an image acquired after the time-series of images.

2. The object detection device according to claim 1 , wherein

the processor updates, only when there is a predetermined similarity relationship between the two or more types of the object detected in the tracked period, the confidence threshold for at least any of the two or more types.

3. The object detection device according to claim 2 , wherein

the predetermined similarity relationship is set between types of an object having a common feature in terms of an appearance.

4. The object detection device according to claim 1 , wherein

the processor detects, for each of the time-series of images, in a region in which the confidence score is maximum of two or more regions in which the confidence score of any type belonging to the same class of the plurality of types is equal to or higher than a confidence threshold of the type and a degree of overlapping between each other is equal to or higher than a predetermined overlapping threshold among the plurality of regions on the image, an object of a type corresponding to the maximum confidence score, and

updates, only when two or more types of the object detected in the tracked period belong to the same class, the confidence threshold for at least any of the two or more types.

5. The object detection device according to claim 1 , wherein

the period when any of two or more types of the detected object has an appearance feature that varies over time is set to be shorter than the period when none of two or more types of the detected object has an appearance feature that varies over time.

6. An object detection device, comprising

a processor configured to

calculate, by inputting images acquired in time-series to a classifier that calculates, for each of a plurality of types of an object to be detected, a confidence score indicating a degree of certainty that the type of the object is represented, for each of the time-series of images, the confidence score of each of the plurality of types for a plurality of regions on the image;

detect, for each of the time-series of images, in a region in which the confidence score is maximum of two or more regions in which the confidence score of any type belonging to the same class of the plurality of types is equal to or higher than a confidence threshold of the type and a degree of overlapping between each other is equal to or higher than a predetermined overlapping threshold among the plurality of regions on the image, an object of a type corresponding to the maximum confidence score;

track the detected object, based on the region in which the detected object is represented in each of the time-series of images;

count a frequency of occurrence for each type of the detected object in a period in which the detected object is tracked; and

update the class in such a way as to include two or more types of the detected object when the frequency of occurrence for each of the two or more types is equal to or higher than a frequency threshold in the tracked period, wherein

the processor applies the updated class to a region in which the detected object is estimated to be located in an image acquired after the time-series of images.

7. The object detection device according to claim 6 , wherein

the processor updates, when a first sub-class to which one of the two or more types belongs and in which whether or not the degree of overlapping is equal to or higher than the predetermined overlapping threshold is determined and a second sub-class to which the other one of the two or more types belongs and in which whether or not the degree of overlapping is equal to or higher than the predetermined overlapping threshold is determined are different, the class in such a way that the class includes both of the first sub-class and the second sub-class.

8. The object detection device according to claim 6 , wherein

the period when any of two or more types of the detected object has an appearance feature that varies over time is set to be shorter than the period when none of two or more types of the detected object has an appearance feature that varies over time.

9. A non-transitory recording medium in which a computer program for object detection is recorded for causing a computer to execute:

calculating, by inputting images acquired in time-series to a classifier that calculates, for each of a plurality of types of an object to be detected, a confidence score indicating a degree of certainty that the type of the object is represented, for each of the time-series of images, the confidence score of each of the plurality of types for a plurality of regions on the image;

detecting, for each of the time-series of images, in a region in which the confidence score of any of the plurality of types is equal to or higher than a confidence threshold of the type among the plurality of regions on the image, an object of the type;

tracking the detected object, based on the region in which the detected object is represented in each of the time-series of images;

counting a frequency of occurrence for each type of the detected object in a period in which the detected object is tracked; and

updating the confidence threshold for at least any of two or more types of the object detected in the tracked period in such a way that the confidence threshold of a type of the object having the higher frequency of occurrence among the two or more types is lower than the confidence threshold of a type of the object having the lower frequency of occurrence among the two or more types, wherein

detection of the object uses the updated confidence threshold for a comparison with the confidence score for a region in which the detected object is estimated to be located in an image acquired after the time-series of images.

10. A non-transitory recording medium in which a computer program for object detection is recorded for causing a computer to execute:

calculating, by inputting images acquired in time-series to a classifier that calculates, for each of a plurality of types of an object to be detected, a confidence score indicating a degree of certainty that the type of the object is represented, for each of the time-series of images, the confidence score of each of the plurality of types for a plurality of regions on the image;

detecting, for each of the time-series of images, in a region in which the confidence score is maximum of two or more regions in which the confidence score of any type belonging to the same class of the plurality of types is equal to or higher than a confidence threshold of the type and a degree of overlapping between each other is equal to or higher than a predetermined overlapping threshold among the plurality of regions on the image, an object of a type corresponding to the maximum confidence score;

tracking the detected object, based on a region in which the detected object is represented in each of the time-series of images;

counting a frequency of occurrence for each type of the detected object in a period in which the detected object is tracked; and

updating the class in such a way as to include two or more types of the detected object when the frequency for each of the two or more types is equal to or higher than a frequency threshold in the tracked period, wherein

detection of the object applies the updated class to a region in which the detected object is estimated to be located in an image acquired after the time-series of images.