ETRI-Knowledge Sharing Plaform

KOREAN
논문 검색
Type SCI
Year ~ Keyword

Detail

Conference Paper Kick Recognition System Using Dual RGB-D Cameras
Cited 1 time in scopus Share share facebook twitter linkedin kakaostory
Authors
Sungjin Hong, Myung-Gyu Kim, Yejin Kim
Issue Date
2018-12
Citation
International Conference on Video and Image Processing (ICVIP) 2018, pp.12-16
Language
English
Type
Conference Paper
DOI
https://dx.doi.org/10.1145/3301506.3301517
Abstract
This paper introduces a kick recognition system based on human body detection using dual RGB-D cameras. Recently, the availability of RGB-D cameras makes it possible to get the human body joints that are informative for activity analysis. However, single camera-based approaches enforce frontal-oriented action due to the occlusion problem. Using dual RGB-D cameras and a smart sandbag, the proposed system detects major joints and recognizes various kick actions from a general user in real time. For each camera, our system detects salient body parts in a kick action such as a head and feet. A local detector trained with a supervised model is used for the head detection. The detected body parts are converted into a quadtree-structured graph model to detect feet using accumulative geodesic distance (AGD). To deal with the occlusion by the sandbag, our system compares the result of each camera and selects the most reliable one based on AGD. For the kick recognition, a finite-state machine is adopted to track and to segment continuous kick movements into different states. Considering a viewpoint change and a variable kick speed, fixed size descriptors are constructed from the interpolated action to recognize user kicks. We evaluated our system using various kick actions in taekwondo and achieved a high recognition rate of 92%.
KSP Keywords
Body joints, Body parts, Finite State Machine(FSM), Geodesic distances, Head Detection, Human body detection, RGB-D cameras, Real-time, Recognition Rate, Recognition system, Supervised model