Granted Patent B2

US 12,561,815 · App. 18/310,216 · Granted Feb 24, 2026

Inventors: Kichang Yang (Seoul, KR); Youngki Lee (Seoul, KR); Juheon Yi (Seoul, KR); Kyungjin Lee (Seoul, KR)

Assignee: SEOUL NATIONAL UNIVERSITY R&DB FOUNDATION

G06T7/20G06V10/25G06V10/44G06V10/762G06V10/764G06V2201/07

Abstract

Claims (29)

1 . A video analysis terminal comprising:

at least one processor, and

memory storing instructions,

wherein the instructions, when executed by the at least one processor, cause the video analysis terminal to:

recommend tracking-failure patches and new-object patches in a current frame of a video image; and

generate a first patch cluster by collecting the tracking-failure patches recommended in the current frame, or generate a second patch cluster by collecting the new-object patches recommended in the current frame,

wherein the tracking-failure patches indicate regions for which tracking has failed in the current frame,

wherein the new-object patches indicate regions in which a new object is likely to be present but has not been detected and thus tracking has not been performed,

wherein the tracking-failure patches are extracted based on features, which are extracted from the current frame and imply a tracking failure, and machine learning for predicting a tracking failure degree based on the extracted features, and

wherein the extracted features comprise normalized cross correlation (NCC) between a bounding box in a frame before tracking and a bounding box after the tracking, a velocity of a bounding box, an acceleration of the bounding box, a gradient of a region around the bounding box, and confidence of detection.

2 . The video analysis terminal of claim 1 , wherein the instructions, when executed by the at least one processor, cause the vide analysis terminal to receive the first patch cluster or the second patch cluster, and detect an object, so as to improve an object detection speed.

3 . The video analysis terminal of claim 1 , wherein the first patch cluster or the second patch cluster has a rectangular shape.

4 . The video analysis terminal of claim 1 , wherein a size of the first patch cluster or the second patch cluster is adjusted according to a size and number of tracking-failure patches or new-object patches included in each of the first patch cluster or the second patch cluster.

5 . The video analysis terminal of claim 1 , wherein the instructions, when executed by the at least one processor, cause the video analysis terminal to collect the new-object patches and the tracking-failure patches in every t frame of the video image, before performing object detection.

6 . The video analysis terminal of claim 1 , wherein the instructions, when executed by the at least one processor, cause the video analysis terminal to recommend the new-object patches by using an edge intensity and a refresh interval.

7 . The video analysis terminal of claim 1 , wherein the instructions, when executed by the at least one processor, cause the video analysis terminal to generate the first patch cluster by classifying and arranging the collected tracking-failure patches according to error values.

8 . The video analysis terminal of claim 1 , wherein machine learning is performed by using a decision tree classification model, based on the extracted features, and then a degree of tracking failure is predicted by identifying an intersection over union (IoU), which is a degree of overlap between a tracked bounding box and a real object.

9 . A method, performed by a terminal, of performing video object detection, the method comprising:

recommending tracking-failure patches and new-object patches in a current frame of a video image; and

generating a first patch cluster by collecting the tracking-failure patches recommended in the current frame, or generating a second patch cluster by collecting the new-object patches recommended in the current frame,

wherein the tracking-failure patches indicate regions for which tracking has failed in the current frame, and

wherein the new-object patches indicate regions in which a new object is likely to be present but has not been detected and thus tracking has not been performed,

wherein the tracking-failure patches are extracted based on features, which are extracted from the current frame and imply a tracking failure, and machine learning for predicting a tracking failure degree extracted features, and

wherein the extracted features comprise normalized cross correlation (NCC) between a bounding box in a frame before tracking and a bounding box after the tracking, velocity of a bounding box, an acceleration of the bounding box, a gradient of a region around the bounding box, and confidence of detection.

10 . The method of claim 9 , further comprising receiving the first patch cluster or the second patch cluster, and detecting an object, so as to improve an object detection speed.

11 . The method of claim 9 , wherein a size of the first patch cluster or the second patch cluster is adjusted according to a size and number of tracking-failure patches or new-object patches included in each of the first patch cluster or the second patch cluster.

12 . The method of claim 9 , wherein the recommending comprises recommending the new-object patches by using an edge intensity and a refresh interval.

13 . The method of claim 9 , wherein the recommending comprises recommending the tracking-failure patches based on features, which are extracted from the current frame and imply a tracking failure, and machine learning for predicting a tracking failure degree based on the extracted features.

14 . A computer program stored in a non-transitory computer-readable recording medium, for executing, on the terminal, the method of performing the video object detection of claim 9 .

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (13)