• Proceedings of the Korea Information Processing Society Conference
• /
• 2018.10a
• /
• pp.686-689
• /
• 2018

3D 손 포즈 추정(Hand Pose Estimation, HPE)은 스마트 인간 컴퓨터 인터페이스를 위해서 중요한 기술이다. 이 연구에서는 딥러닝 방법을 기반으로 하여 단일 RGB-Depth 카메라로 촬영한 양손의 3D 손 자세를 실시간으로 인식하는 손 포즈 추정 시스템을 제시한다. 손 포즈 추정 시스템은 4단계로 구성된다. 첫째, Skin Detection 및 Depth cutting 알고리즘을 사용하여 양손을 RGB와 깊이 영상에서 감지하고 추출한다. 둘째, Convolutional Neural Network(CNN) Classifier는 오른손과 왼손을 구별하는데 사용된다. CNN Classifier 는 3개의 convolution layer와 2개의 Fully-Connected Layer로 구성되어 있으며, 추출된 깊이 영상을 입력으로 사용한다. 셋째, 학습된 CNN regressor는 추출된 왼쪽 및 오른쪽 손의 깊이 영상에서 손 관절을 추정하기 위해 다수의 Convolutional Layers, Pooling Layers, Fully Connected Layers로 구성된다. CNN classifier와 regressor는 22,000개 깊이 영상 데이터셋으로 학습된다. 마지막으로, 각 손의 3D 손 자세는 추정된 손 관절 정보로부터 재구성된다. 테스트 결과, CNN classifier는 오른쪽 손과 왼쪽 손을 96.9%의 정확도로 구별할 수 있으며, CNN regressor는 형균 8.48mm의 오차 범위로 3D 손 관절 정보를 추정할 수 있다. 본 연구에서 제안하는 손 포즈 추정 시스템은 가상 현실(virtual reality, VR), 증강 현실(Augmented Reality, AR) 및 융합 현실 (Mixed Reality, MR) 응용 프로그램을 포함한 다양한 응용 분야에서 사용할 수 있다.

• Proceedings of the Korea Multimedia Society Conference
• /
• 2001.11a
• /
• pp.849-853
• /
• 2001

본 논문에서는 연속적인 인간의 제스처 영상을 저차원 제스처 특징 공간과 HMM 이용하여 인식할 수 있는 방법에 대해 소개한다. 일반적으로 제스처 공간에서 모델 패턴들과 매칭하기 위해서는 모든 모델 영상과 연속적인 입력영상들간의 거리평가로 인식을 수행하게 된다. 여기서 제안한 방법은 연속성을 가진 모델영상들을 HMM로 포즈들의 시공간적 특성을 매칭에 이용하였다. 이 방법은 동작의 구분뿐만 인식결과를 학습에 이용할 수 있는 장점이 있다.

• Proceedings of the IEEK Conference
• /
• 2003.07e
• /
• pp.2024-2027
• /
• 2003

본 논문에서는 얼굴 인식 분야에서 사용되는 PCA/LDA 알고리즘을 대신하기 위해 DCT/LDA 알고리즘을 제안하였다. PCA/LDA를 이용한 얼굴 인식의 경우 PCA 를 이용하여 얼굴 영상을 적은 수의 특징 값으로 표현한 다음 LDA를 수행한다. 그러나 PCA는 트레이닝 과정의 계산량이 많고 트레이닝 셋이 변할 때마다 기저 벡터가 변화한다. PCA/LDA의 단점을 개선하기 위해 계산량이 적고 기저 벡터가 일정한 DCT의 계수를 사용한다. DCT/LDA를 사용할 경우 특징 값을 빠르게 추출하면서 PCP/LDA와 유사한 성능을 얻을 수 있다. 실험을 통하여 포즈 변화와 조명 변화가 있는 얼굴 데이터 셋에서 최고 97.8%의 인식률을 보였다.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.43 no.1 s.307
• /
• pp.17-26
• /
• 2006

This paper proposes a novel hidden Markov model (HMM)-based gesture recognition method and applies it to an HCI to control a computer game. The novelty of the proposed method is two-fold: 1) the proposed method uses a continuous streaming of human motion as the input to the HMM instead of isolated data sequences or pre-segmented sequences of data and 2) the gesture segmentation and recognition are performed simultaneously. The proposed method consists of a single HMM composed of thirteen gesture-specific HMMs that independently recognize certain gestures. It takes a continuous stream of pose symbols as an input, where a pose is composed of coordinates that indicate the face, left hand, and right hand. Whenever a new input Pose arrives, the HMM continuously updates its state probabilities, then recognizes a gesture if the probability of a distinctive state exceeds a predefined threshold. To assess the validity of the proposed method, it was applied to a real game, Quake II, and the results demonstrated that the proposed HMM could provide very useful information to enhance the discrimination between different classes and reduce the computational cost.

• Journal of KIISE:Software and Applications
• /
• v.28 no.12
• /
• pp.947-954
• /
• 2001

In this paper, we propose a method for the recognition of various head gestures with automata technique applied to the sequence of facial pose states. Facial regions as detected by using the optimum facial color of I-component in YIQ model and the difference of images adaptively selected. And eye regions are extracted by using Sobel operator, projection, and the geometric location of eyes Hierarchical feature analysis is used to classify facial states, and automata technique is applied to the sequence of facial pose states to recognize 13 gestures: Gaze Upward, Downward, Left ward, Rightward, Forward, Backward Left Wink Right Wink Left Double Wink, Left Double Wink , Right Double Wink Yes, and No As an experimental result with total 1,488 frames acquired from 8 persons, it shows 99.3% extraction rate for facial regions, 95.3% extraction rate for eye regions 94.1% recognition rate for facial states and finally 99.3% recognition rate for head gestures. .

• Proceedings of the Korean Information Science Society Conference
• /
• 2004.10a
• /
• pp.223-225
• /
• 2004

얼굴 인식 및 검출에 있어서 어려운 문제가 조명의 변화와 포즈의 변화에 따른 성능 면에서의 신뢰성이다. 이러한 상황(Context)의 변화를 고려하여 영상을 처리하기 위하여 얼굴 영상에 주어진 조명의 상황을 SOM으로 분석하며, 영상에 따라 다른 전처리 기법의 필요성에 대해 제안한다. SOM은 비 지도학습으로써 얼굴 이미지들을 수집하여 그룹화 함으로써 상황분석을 위한 알고리즘으로 활용한다 이는 상황분석 기법을 적용하기 위한 응용에 활용할 수 있으며, 적절한 전처리 기법은 얼굴 인식의 성능을 향상시킴을 알 수 있었다.

• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.66 no.4
• /
• pp.672-681
• /
• 2017

This paper is concerned with the design methodology of face recognition system based on pose estimation. In 2-dimensional face recognition, the variations of facial pose cause the deterioration of recognition performance because object recognition is carried out by using brightness of each pixel on image. To alleviate such problem, the proposed face recognition system deals with Learning Vector Quantizatioin(LVQ) or K-Nearest Neighbor(K-NN) to estimate facial pose on image and then the images obtained from LVQ or K-NN are used as the inputs of networks such as Convolution Neural Networks(CNNs) and Radial Basis Function Neural Networks(RBFNNs). The effectiveness and efficiency of the post estimation using LVQ and K-NN as well as face recognition rate using CNNs and RBFNNs are discussed through experiments carried out by using ICPR and CMU PIE databases.

• Journal of Korea Multimedia Society
• /
• v.15 no.8
• /
• pp.1067-1074
• /
• 2012

In this paper, we proposed a contents of Chinese characters learning based on gesture recognition using HMM(hidden markov model) algorithm. Input image of the system is obtained in 3-dimensional information from the TOF camera, and the method of gesture recognition is consisted of part of forecasting user's posture in two infrared images and part of recognizing gestures from continuous poses. In the communication between human and computer, this system provided convenience that user can manipulate it easily by not using any further equipment but action. Because this system raise immersion and interest by using two large display and various multimedia factor, it can maximize information transmission. The edutainment Chinese character contents proposed in this paper provide educational effect that use can master Chinese character naturally with interest, and it can be expected a synergy effect via content experience because it is based on gesture recognition.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.45 no.5
• /
• pp.63-71
• /
• 2008

In this paper, we propose a color based hierarchically propagated particle filter that extends the color based particle filter into the articulated upper body tracking. Since color feature is robust to partial occlusion and rotation, the color based particle filter is widely used for object tracking. However, in articulated body tacking, it is not desirable to use the traditional particle filter because the dimension of the state vector usually is high and thus, many samples are required for robust hacking. To overcome this problem, we use a hierarchical tracking method for each body part based on the blown body part. By using a hierarchical tracking method, we can reduce the number of samples for robust tracking in the cluttered environment. Also for human pose recognition, we classify the human pose into eight categories using Support Vector Machine(SVM) according to the angle between upper- arm and fore-arm. Experimental results show that our proposed method is more efficient than the traditional particle filter.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.24 no.4
• /
• pp.607-616
• /
• 2014

Face recognition is a big challenge in surveillance system since different rotation angles of the face make the difficulty to recognize the face of the same person. This paper proposes a novel method to recognize face with different head poses by using 3D information of the face. Firstly, head pose estimation (estimation of different head pose angles) is accomplished by the POSIT algorithm. Then, 3D face image data is constructed by using head pose estimation. After that, 2D image and the constructed 3D face matching is performed. Face verification is accomplished by using commercial face recognition SDK. Performance evaluation of the proposed method indicates that the error range of head pose estimation is below 10 degree and the matching rate is about 95%.