스마트 깔창을 이용한 앙상블 학습 기반의 Open-Set gait identification 시스템 = An open-set gait identification system based on ensemble learning using a smart insole

국문 초록 (Abstract)

사람의 보행은 인간의 보편적인 신체활동인 동시에, 개개인의 고유한 생물학
적 특징 정보를 내포하고 있기 때문에 의료 진단이나 헬스 케어 등의 분야에서
생체 인식을 위한 태스크를 수행하는데 있어 가치 있는 정보를 제공해줄 수 있
다. 이를 토대로, 사람의 보행 정보에 딥러닝 기법을 적용하여 보행 인식을 시도하
는 연구들이 이어져 왔다. 그러나 기존 연구들의 경우, 대부분, 테스트 데이터셋에
대한 정답이 훈련 데이터셋에 존재하는 closed-set identification 문제 해결에만 집
중하였다. 그래서, 본 논문에서는 모델 학습 시, 한 번도 본 적 없는 데이터에 대해
인식을 해야 하는 open-set identification 문제를 해결하기 위해, 앙상블 딥러닝 네
트워크 구조의 특징추출기와 OSVM(One-class SVM)을 이용한 분류기로 구성된 방
법을 제안한다. 제안하는 방법은 크게 보행 데이터를 전처리하는 단계와 특징 추출
단계, 분류 단계로 구성되어 있다. 웨어러블 기기로부터 취득한 보행 데이터는 학
습에 적합한 형태로 바꿔주기 위하여 가우시안 필터를 사용하여 단위 걸음으로 구
분한 뒤, 동일한 길이로 전처리를 해주었다. 특징 추출 단계에서는 CNN, RNN, and
self-attention 모듈로 이뤄진 앙상블 딥러닝 네트워크가 입력된 데이터를 잠재공간
(Latent space) 상의 embedding vector로 mapping시키는 역할을 한다. 마지막으로
분류 단계에서는 one-class support vector machine (OSVM)을 사용하여 데이터를
분류한다. 실험 데이터는 총 40명으로부터 취득하였으며, Train 20명, Validation 10명,
Test 10명으로 나눠주었다. 실험은 Train과 Validation 데이터셋으로 모델을 학
습시킨 뒤, Test dataset으로 모델의 성능을 확인하는 실험을 수행하였다. 실험 결
과, 본 논문에서 제안한 모델이 open-set gait identification 문제를 해결하는데 있
어 우수한 성능을 보여주는 것을 확인하였다.

다국어 초록 (Multilingual Abstract)

Human gait is a general human physical activity, and at the same time
contains information about each individual's unique biological characteristics,
it can provide valuable information in performing tasks for biometric
identification in fields such as medical diagnosis and health care. As a
result, studies have continued to attempt gait identification by applying
deep learning techniques based on human gait information. However, in the
case of existing studies, most of them focused on solving the closed-set
identification problem in which the label for the test dataset exists in the
training dataset. Therefore, in this paper, to solve the open-set
identification problem that needs to recognize data that has never been
seen before when training a model, we propose a method consisting of an
ensemble deep learning network and a classifier used by OSVM(one-class
SVM). The proposed method is composed of a pre-processing step of the
gait data, a feature extraction, and a classification step. In order to change
the gait data obtained from the wearable device into a form suitable for
learning, a Gaussian filter was used to divide the gait data into unit steps,
and then pre-processed to the same length. In the feature extraction stage,
an ensemble deep learning network composed of CNN, RNN, and
self-attention modules plays a role in mapping the input data to an
embedding vector in the latent space. Finally, in the identification step, the
data is classified using a one-class support vector machine (OSVM).
Experimental data were acquired from a total of 40 people, and divided
into 20 trainees, 10 validation people, and 10 test people. For the
experiment, after training the model with the Train and Validation datasets,
an experiment was performed to check the performance of the model with
the Test dataset. As a result of the experiment, the ensemble deep
learning network using a classifier used by OSVM showed excellent
performance in solving the open-set gait identification problem.

목차 (Table of Contents)

• 목 차
• 국문초록 ·····················································································································ⅰ
• 목 차 ·····················································································································ⅱ
• List of Tables ············································································································ⅳ
• List of Figures ···········································································································ⅴ
• Ⅰ. 서론 ························································································································1
• Ⅱ. 관련연구 ················································································································4
• 2.1 합성곱 신경망 네트워크(CNN) ·····································································4
• 2.2 순환 신경망 네트워크(LSTM) ·······································································6
• 2.3 셀프 어텐션 네트워크(Self-Attention) ························································9
• Ⅲ. 본론 ······················································································································11
• 3.1 Feature Extractor Using Ensemble Deep Learning ······························11
• 1) 합성곱 기반의 멀티 모달 딥러닝 네트워크 ······································14
• 2) LSTM 기반의 멀티 모달 딥러닝 네트워크 ········································16
• 3) 셀프 어텐션 기반의 멀티 모달 딥러닝 네트워크 ····························17
• 3.2 Classifier for Open-Set Gait Identification ·············································19
• 1) Embedding vector ····················································································19
• 2) Classifier Using OSVM ············································································21
• Ⅳ. 실험 ······················································································································22
• 4.1 실험 환경 ·······································································································22
• 4.2 데이터 취득 및 전처리 ···············································································22
• 4.3 데이터셋 Split 및 평가지표 ········································································25
• 4.4 Loss function for Feature extractor ························································27
• 4.5 실험 결과 ·······································································································29
• Ⅴ. 결론 ······················································································································33
• 참고문헌 ·····················································································································35
• Abstract ······················································································································39
• 감사의 글 ···················································································································41