DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Kim, Changick | - |
dc.contributor.advisor | 김창익 | - |
dc.contributor.author | Choi, Seokeon | - |
dc.date.accessioned | 2018-06-20T06:22:35Z | - |
dc.date.available | 2018-06-20T06:22:35Z | - |
dc.date.issued | 2017 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=675443&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/243335 | - |
dc.description | 학위논문(석사) - 한국과학기술원 : 전기및전자공학부, 2017.2,[iii, 43 p. :] | - |
dc.description.abstract | Gait is a useful biometric feature for person identification in video surveillance applications, since it requires neither constrained condition nor attention of the individuals. Recently, as a result of advances in the Kinect sensor and research on 3D pose estimation, gait recognition methods based on kinematic analysis of human joints have been widely studied. However, these model-based methods have major problems. First, modeling with complex analysis of all gait-poses causes high computational complexity. In addition, since 3D pose estimation errors always exist in gait sequences, discriminative feature vectors are inevitably contaminated by feature vectors including the error in the process of aggregating feature vectors in video units. Thus, these problems greatly reduce both recognition speed and performance in real video surveillance applications. To address these problems, we propose a candidate selection method which chooses representative gait poses through rank-level fusion taking into account the conditions of smoothness and periodicity. Through the selection process, both computational complexity and memory requirement can be reduced. Next, we suggest a Pose-aware Decision Fusion (PDF) method, which merges the recognition scores and results of each gait-pose computed by repetitive global linear matching. The method, which is a new approach of model-based methods, increases the discrimination power and makes it robust to estimation error for each pose, compared with conventional dissimilarity-space classification. To evaluate the proposed method, we basically use the UPCV Gait dataset extracted from the Kinect sensor. In addition, we intentionally add a 3D joint position error from 0% to 4.5% for the existing dataset to evaluate the joint position error robustness of the proposed method in the real world environment. By the experiments, we show that the proposed method achieves the highest performance compared with state-of-the-art model-based gait recognition methods. In addition, the proposed method contributes to the realization of a real-time application system that can verify the identity of an individual as soon as a gait-pose information is received. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | Model-based gait recognition | - |
dc.subject | Kinect sensor | - |
dc.subject | candidate selection | - |
dc.subject | Pose-aware Decision Fusion (PDF) | - |
dc.subject | repetitive global linear matching | - |
dc.subject | weighted voting | - |
dc.subject | rank-level fusion | - |
dc.subject | real-time applications | - |
dc.subject | 모델기반 보행인식 | - |
dc.subject | 키넥트 센서 | - |
dc.subject | 후보 선택 | - |
dc.subject | 자세단위 결정융합 | - |
dc.subject | 반복 전역 선형 매칭 | - |
dc.subject | 가중 투표 | - |
dc.subject | 순위기반 융합 | - |
dc.subject | 실시간 응용 | - |
dc.title | Robust model-based gait recognition via candidate selection and pose-aware decision fusion | - |
dc.title.alternative | 후보선택과 자세단위 결정융합을 통한 강인한 모델기반 보행인식 방법 | - |
dc.type | Thesis(Master) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :전기및전자공학부, | - |
dc.contributor.alternativeauthor | 최석언 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.