서지주요정보
Context-driven spatio-temporal classifier for articulated hand gesture recognition = 연결구조의 손 제스쳐 인식을 위한 맥락 중심 시공간 분류기에 관한 연구
서명 / 저자 Context-driven spatio-temporal classifier for articulated hand gesture recognition = 연결구조의 손 제스쳐 인식을 위한 맥락 중심 시공간 분류기에 관한 연구 / Youngkyoon Jang.
발행사항 [대전 : 한국과학기술원, 2015].
Online Access 원문보기 원문인쇄

소장정보

등록번호

8028428

소장위치/청구기호

학술문화관(문화관)B1층 보존서고

DGCT 15007

휴대폰 전송

도서상태

이용가능(대출불가)

사유안내

반납예정일

리뷰정보

초록정보

Hand gesture recognition has a long and diversified history in many research communities, including computer vision (CV) and human computer interaction (HCI). However, in a wearable augmented reality (AR) or virtual reality (VR) environment, despite significant research effort for recognizing articulated hand gesture, an adequate solution has not yet been discovered to overcome a major challenge utilizing a self-occluded (missing) visual data. This is largely due to the fact that articulated hand gesture recognition in egocentric viewpoint is complicated and difficult; potential solutions must be tolerant to the incomplete data, and invariant to the translation and rotation of the hand in 3D space. Moreover, it has to be invariant to the variance of motions and postures. Despite these significant challenge, this work presents a novel framework, which is based on the proposed Spatio-temporal classifier, for articulated hand gesture recognition that not only simultaneously performs two different tasks with a significant improvement over existing traditional approaches, but also has computational efficiency and sufficient interaction performance for use in wearable AR/VR applications. As its core, this framework makes use of a basic structure of Random Forest (RF) to probabilistically infer the gestures that we defined based on a given dataset (i.e. egocentric viewed single depth image sequences). Our solutions to two applications include: (i) 3D Finger CAPE: clicking action and position estimation under self-occlusions in egocentric viewpoint and (ii) SD Gesture: static and dynamic gesture estimation for manipulating a function-equipped AR object. For both these applications we show that the proposed context-driven Spatio-temporal classifier utilizing RF are able to significantly out-perform existing state-of-the-arts, and works as a novel learning technique for supporting a novel Natural User Interface (NUI) in wearable AR/VR applications. Based on the results of the proposed Spatio-temporal classifiers to these two applications, we were able to provide a novel selection and manipulation process in an AR/VR space, which does not require any additional device (e.g. gloves or optical markers).

맨손 기반 제스쳐 인식은 컴퓨터 비젼 (CV), 인간 컴퓨터 상호작용 (HCI) 분야를 포함하는 여러 연구분야에서 오랜 기간 다양한 응용과 함께 연구되어 왔다. 하지만, 착용형 증강현실 (AR) 또는 가상현실 (VR) 환경에서는 관절구조를 가지는 손 제스처 인식에 대한 상당한 연구 노력에도 불구하고 자가가림 현상으로 인해 보이지 않는 비주얼 (Visual) 데이타를 활용해야 하는 주요한 연구 이슈에 대해서는 아직도 적절한 해법을 제시하지 못하고 있다. 앞서 언급한 자가가림 현상은 대부분의 경우 1인칭 시점에서 획득된 관절 구조의 손 영상을 활용할때 나타나는 현상으로써 이를 통한 제스처를 인식하려는 시도 자체가 어려운 문제이기 때문이다. 따라서 이를 해결할수 있는 잠재적인 해결책은 이 완전하지 않은 데이타를 활용에도 제스처 인식이 실패하지 않아야 하고, 더욱이 3차원 공간 상에서 손의 자유로운 회전/이동에도 강건한 방법이어야 한다. 더불어, 이 잠재적인 해결책은 또한 손의 자유로운 자세와 움직임 등으로 인해 생길수 있는 여러 변형들에 대해서도 강건하게 동작해야 한다. 이와 같이 앞선 언급된 주요한 문제들을 다루면서, 본 학위논문은 관절 구조의 손 제스처 인식을 위해 시공간 분류기를 기반으로 하는 새로운 프레임워크를 제안한다. 이 새로운 프레임워크는 두가지 다른 종류의 동작(tasks)을 동시에 수행할 뿐 아니라 (기존의 방법보다 월등한 성능으로), 착용형 AR/VR 응용에 대한 사용성 측면에서도 충분한 성능 및 계산 효율성을 보인다. 본 논문의 핵심 내용 중 하나로써 제안하는 프레임워크는 랜덤 포레스트 (Random Forest)의 기본 구조를 활용하는데, 이는 주어진 데이타셋을 기반으로 본 저자가 정의하는 제스처들을 확률적으로 추론하기 위해 사용한다. 제안하는 프레임워크를 활용하여 본 논문에서는 두가지 응용 및 활용 가능성에 대해 설명하는데 그 중 하나는 (1) 3D Finger CAPE: 1인칭 시점에서 상호작용시 자가가림 현상이 발생하는 상황에서 손가락을 활용한 클릭하는 동작과 클릭된 위치 추정, 그리고 다른 하나는 (2) SD Gesture: 기능이 탑재된 AR 객체를 조작하기 위한 정적 그리고 동적 제스쳐 인식이다. 두 응용 모두의 경우에서 우리는 제안하는 맥락 중심 시공간 분류기가 현존하는 최신의 연구 결과 성능을 뛰어 넘는 것을 확인할수 있었으며, 착용형 AR/VR 응용에서 이 방법들이 새로운 형태의 NUI (Natural User Interface)를 지원하는 적절한 학습 기법으로 동작하는 것을 확인하였다. 이 제안하는 시공간 분류기 결과를 기반으로 하는 두가지 응용으로써 본 논문에서는 AR/VR 공간에서의 새로운 선택/조작 프로세스를 제시하고 있다. 본 연구는 글로브 (glove) 또는 시각 마커(optical markers)와 같은 추가적인 장비 없이 상호작용을 지원할수 있다.

서지기타정보

서지기타정보
청구기호 {DGCT 15007
형태사항 vii, 63 p. : 삽화 ; 30 cm
언어 영어
일반주기 저자명의 한글표기 : 장영균
지도교수의 영문표기 : Woontack Woo
지도교수의 한글표기 : 우운택
Including Appendix
학위논문 학위논문(박사) - 한국과학기술원 : 문화기술대학원,
서지주기 References : p.
QR CODE

책소개

전체보기

목차

전체보기

이 주제의 인기대출도서

Theproposed spatio-temporal classifier base: examplesof (a) configuring the spatio-tempora feature and (b) the process of the base classifier.

Chapter organization.

Ph.D. research positioning in hand 기 2 estimation.

Ph.D. research positioning in hand-based natural user interface (NUI).

The Ph.D. research direction.

Our system, called 3D Finger CAPE, supports both 3D finger clicking action detection and clicked position estimation at the same time. In egocentric viewpoint, self-occlusion is caused when a user interacts with VR objects. The proposed spatio-temporal forest estimates 3D clicking positions (purple cube) when the pre-learnt action has occurred (yellow cube) on the screen. 3D Finger CAPE could be a

Graphical representation of the proposed framework (gray nodes: input features)

Fingertips detection concept based on a depth image as a 2D gray image: (a) fingertip detection concept (b) differentiation graph along the contour points.

The proposed clicking action detection & fingertipposition estimation model: (a) an example of configuring the ST feature based on the detected fingertips and the estimated joints ofthe hand (b) examples of the process of the forest.

Examples of the modified geodesic maxima-based detection (red dots) [6, 33] and the 3D hand posture estimation (skeletal model) [39] for the various hand posture cases.

VR environment setting for object selection experiments

Example ofexperimental scenarios of sparse and dense object selections in static scene: (a) objects are spaced farther apart in the sparse environment (b) objects are placed closer together in the dense environment. In dynamic scene, the objects move in the VR environment.

Experimental results showingthe Equal Error Rate (EER) ofaction detection fortwodifferent cases (see text).

ROC curves of clicking action classification for five fingers by the proposed ST forest

The results of the distance errors between the target position (0,0,0) coordinates and the and the estimated fingertip position, based on GM-based (red dots) [6, 33], HPE-based (green dots) [39], 3D Finger CAPE (blue dots), plotted in the camera coordinates gathered in the four different scenarios. composed of: (a) sparse and static objects (scenario#-1) (b) sparse and dynamic objects (scenario#-2

Post-Questionnaire

Experimental results showing the number of trials from a subject and failures triggered by GM-based method [6, 33]: (gray) number of total trials (black) failure counts.

Statistical results of the first session, based on ANOVA and T-test as a post-hoc test.

Mean and standard deviation results of user's preference on each method based on the post-questionnaire. (GM: 6, 33], HPE: [39])

Statistical results based on the questionmaire, based on Friedman and T-test as a post-hoc test. Question# Friedman GM [6, 33] VS. CAPE HPE [39 VS. CAPE

Our system, which we refer to as SD Gesture, estimates both static and dynamic gestures simultaneously. More specifically, the proposed static-dynamic forest estimates static 3D hand postures for triggering a functional object on hand (yellow circle), and estimates its action (i.e. function) status for manipulating the AR object (red circle). Thus, SD Gesture can be applied to the manipulation of

Graphical representation of the proposed framework

Examples ofa local coordinates of the hand: (a) an example ofcoordinate transformation from camera to hand coordinates (b) examples of two different 3D static gestures, highlighting different positions of a voxel containing hand point clouds in the local coordinates of the hand.

Example of the proposed Static-Dynamic (SD) voxel feature configuration

Example of selecting a voxel plane along a designated axis.

Example ofthe proposed 3D Layered Shape Pattern (LSP): (a) a layered shape on a selected voxel plane is changed through the sequence ofvideo (assuming index finger is bending) (b) LSP pattern values calculated from two different frames based on octal number system.

The proposed Static-Dynamic (SD) forest: (a) an example of 3D Layered Shape Pattern (LSP) value calculation based on quaternary number system (b) an illustration of the proposed static- dynamic gesture forest process.

Examples of the technical implementations: (first row) utilized 3D hand posture estimation results shows stable palm pose, while some partial finger joints are mislocated: (second row) a layered shape is changed while finger is bending; (third row) voxels represent the inclusion of the hand point clouds.

Experimental results showing the static gesture estimation rates when we utilize different voxel counts configuringa cubesurrounding a hand: (a) static gesture estimation accuracies are saturated when voxel counts exceed 196 (7x7x4) (b) when voxel counts is at 32 (4x4x2), estimation rates is at 75.80% (c) when voxel counts is at 196 (7x7x 4), estimation rates is at 91.36%.

A confusion matrix of ten-digit static gestures using the dataset captured from various viewpoints: (a) (left) Liang's results (84.70%, 94.70% excluding bad cases (gesture 3)) [30, 32], and (right) our results (97.80%, 97.50% excluding bad cases (gesture 3)) using a frontal viewpoint dataset (b) (left) Liang's results (60.00%) [30, 32] and (right) our results (91.36%) using a various- viewpoint da

ROC curves ofclicking action classification comparing the proposed SD Gesture (ours) with 3D Finger CAPE [21]: (a) thumb (Ours: 95.59%, 3D Finger CAPE: 89.80%) (b) index finger (Ours: 96.55%, 3D Finger CAPE: 96.90%) (c) ROC curves representing average accuracies for all five fingers (Ours: 93.87%, 3D Finger CAPE: 93.91%).

Examples of an AR object manufacturing (assembling) utilizing the SD Gesture results: (a) because a pinch gesture for selecting a thin and small component in a cluttered scene is difficult, a better tool is summoned (b) and (c) a user can point a very specific component easily and manipulate it with different tools, thus SD Gesture supports an elaborative AR object manufacturing scenario.

Examples of in-air drawing and annotation utilizing the SD Gesture results: by summoning different types of functional AR objects, a user is able to paint color on the wall or annotate a writing based on different objects, such as a an AR pen and (b an AR spraycan.

Examples of an AR camera application taking a photograph of a VR scene.

Examples ofimmersive gaming utilizing the SD Gesture results: a user can summon different types of weapon, such as an AR gun or an AR sword, intuitively.