公开(公告)号：US11948400B2

公开(公告)日：2024-04-02

申请号：US18344877

申请日：2023-06-30

Applicant: HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGY

Inventor： Li Yu ,  Han Yu

IPC: G06V40/20 ,  G06T7/246 ,  G06T7/277 ,  G06V10/46 ,  G06V10/82 ,  G06V20/40 ,  G06V20/52 ,  G08B21/04

CPC classification number: G06V40/23 ,  G06T7/248 ,  G06T7/277 ,  G06V10/462 ,  G06V10/82 ,  G06V20/46 ,  G06V20/52 ,  G08B21/043 ,  G06T2207/10016 ,  G06T2207/20081 ,  G06T2207/20084 ,  G06T2207/30196

Abstract: An action detection method based on a human skeleton feature and a storage medium belong to the field of computer vision, and the method includes: for each person, extracting a series of body keypoints in every frame of the video as the human skeleton feature; calculating a body structure center point and approximating rigid motion area by using the human skeleton feature as a calculated value from the skeleton feature state, and predicting an estimated value in the next frame; performing target matching according to the estimated and calculated value, correlating the human skeleton feature belonging to the same target to obtain a skeleton feature sequence, and then correlating features of each keypoint in the temporal domain to obtain a spatial-temporal domain skeleton feature; inputting the skeleton feature into an action detection model to obtain an action category. In the disclosure, the accuracy of action detection is improved.

公开(公告)号：US20240021019A1

公开(公告)日：2024-01-18

申请号：US18344877

申请日：2023-06-30

Applicant: HUAZHONG UNIVERSITY OF SCIENCE AND TECHNOLOGY

Inventor： Li Yu ,  Han Yu

IPC: G06V40/20 ,  G06T7/246 ,  G06T7/277 ,  G06V10/82 ,  G06V20/40 ,  G06V20/52 ,  G06V10/46 ,  G08B21/04

CPC classification number: G06V40/23 ,  G06T7/248 ,  G06T7/277 ,  G06V10/82 ,  G06V20/46 ,  G06V20/52 ,  G06V10/462 ,  G08B21/043 ,  G06T2207/30196 ,  G06T2207/10016 ,  G06T2207/20081 ,  G06T2207/20084

Abstract: An action detection method based on a human skeleton feature and a storage medium belong to the field of computer vision, and the method includes: for each person, extracting a series of body keypoints in every frame of the video as the human skeleton feature; calculating a body structure center point and approximating rigid motion area by using the human skeleton feature as a calculated value from the skeleton feature state, and predicting an estimated value in the next frame; performing target matching according to the estimated and calculated value, correlating the human skeleton feature belonging to the same target to obtain a skeleton feature sequence, and then correlating features of each keypoint in the temporal domain to obtain a spatial-temporal domain skeleton feature; inputting the skeleton feature into an action detection model to obtain an action category. In the disclosure, the accuracy of action detection is improved.