コロキアムB発表

日時: 9月16日(水)5限(16:50~18:20)


会場: L1

司会: Chen Na
馬越 圭介 M, 2回目発表 ユビキタスコンピューティングシステム 安本 慶一, 岡田 実, 藤本 まなと, 松田 裕貴
title:Construction and evaluation of Personal Identification system using gait vibration
abstract: In recent years, many research has been approached to estimate location and person identification by using floor vibrations that occur when a user walks (i.e., gait vibrations). We propose that the gait vibrations may enable low-cost services such as elderly monitoring and crime prevention systems to be realized. Most of the existing systems are difficult to accurately capture low-level gait vibrations occurring at a distance from the sensor, and there are several problems to overcome in order to achieve accurate person identification, such as the need to utilize multiple sensors and the need to get a large number of steps. To solve these problems, we developed a device that can accurately capture low-level gait vibrations by using a source follower circuit and a piezoelectric sensor, and proposed a new system that can identify a person with a small number of steps at the start of walking (about one to three steps). This system uses machine learning to construct a highly accurate person identification model for thirteen features detected from the gait vibration obtained by a piezo sensor. In this study, two evaluation experiments were conducted to verify the effectiveness of the device and the proposed system. In the evaluation of the circuit performance, it was verified by using a high-performance audio analyzer, and it was confirmed that the proposed circuit has lower noise and wider dynamic range than the conventional circuit. To evaluate the proposed system, we asked six subjects to walk back and forth along a linear floor in a smart home owned by Nara Institute of Science and Technology and collected their gait vibrations, and evaluated the accuracy of person identification using five different machine learning models. The result showed that the average F-value when person identification was conducted by using an optional step from all steps was a high value of 70.8%. Also the average F-values are 63.1%, 75.9% and 87.1% for the first, second and third steps at the start of the outward journey and return journey.
language of the presentation: Japanese
発表題目: 歩行振動を用いた個人識別システムの構築と評価
発表概要: 近年,ユーザが歩行した際に発生する床振動(=歩行振動)を用いて位置推定や個人識別を行う研究が行われている.歩行振動を用いることで低コストで高齢者見守りや防犯システムなどのサービスが実現できる可能性がある.既存システムの多くは,センサから離れた場所で発生する信号レベルの低い歩行振動を正確に捕捉することが困難であり,高精度な個人識別を実現するには,複数のセンサを利用する必要があることや数多くの歩行振動を観測する必要があるなど,いくつかの課題が残されている.本研究では,これらの課題を解決するため,ピエゾセンサにソースフォロワ回路を組み合わせることで信号レベルの低い歩行振動を正確に捕捉できるデバイスを開発し,歩行開始時の少ない歩数(一歩から三歩程度)で個人識別を可能とする新たなシステムを提案する.本システムは,1つのピエゾセンサによって観測されるユーザの歩行振動から検出される13種類の特徴量に対し,機械学習を用いることで高精度な個人識別モデルを構築している.本研究は開発デバイスと提案システムの2点においての有効性を検証するため,2つの評価実験を行なった.回路性能評価では,高性能オーディオアナライザーを用いて検証した結果,提案回路が従来回路よりも低ノイズ化されており,ダイナミックレンジが広いことが確認された.さらに,提案システムの評価検証のため被験者6人の協力のもと,奈良先端大が所有するスマートホーム内において,直線経路を往復で歩行してもらい,歩行振動を収集した.5つの異なる機械学習モデルを用いて個人識別精度を評価した結果,全歩数から任意の一歩を用いて個人識別した場合の平均F値は,70.8%と高い値となった.また,往路・復路の歩行開始時の一歩目のみ,一歩目と二歩目,一歩目から三歩目の3パターンを用いて個人識別した場合の平均F値は,それぞれ63.1%,75.9%,87.1%という結果が得られた.
 
CHOI HYUCKJIN D, 中間発表 ユビキタスコンピューティングシステム 安本 慶一, 岡田 実, 諏訪 博彦, 藤本 まなと
title: Crowd Estimation using WiFi CSI by Machine Learning: Crowd Counting and Cluster Localization
abstract: Various approaches for crowd estimation have been newly attempted by many researchers in the world. Especially, Wi-Fi crowd estimation techniques have achieved a great advance since people became to be able to access to WiFi channel state information. WiFi CSI, which is a channel estimation information on physical layer of RF signal, contains relatively large amount of available data compare to other methods. We can obtain amplitude and phase information of every OFDM subcarriers from a single WiFi packet, which means high possibility to get enough data to properly extract features from data sequence. In this study, we are ultimately aiming to discover meaningful features for crowd estimation system, not only for crowd counting but also estimating the cluster location in certain space or area. So far, we use the CSI curves as a basic source of data processing and feature extraction, which is a data curve made by amplitude values of all subcarriers. We suppose crowd estimation system which is able to accomplish both crowd counting and positioning could be realized by machine learning using combination of state-dependent features structure-dependent features from CSI curves.
language of the presentation: Japanese
 
有元 遼 M, 2回目発表 知能システム制御 杉本 謙二, 岡田 実, 小蔵 正輝(客員准教授)
title: Relaxation of Conditions in the Design of Feedback Error Learning Control System
abstract: Feedback error learning (FEL) with a two-degree-of-freedom (2DOF) structure is a model for volunteer motion proposed from a biological perspective and has turned out to be effective in control system design. In FEL, we achieve accurate tracking by stabilizing the closed-loop system with a fixed feedback (FB) controller and tuning the parameters of a feed-forward (FF) controller, respectively. Although the applicability of this method is limited by the strictly positive real (SPR) condition, a new filter design method is proposed to relax the condition by focusing on the feedback error signal. In this report, we apply the proposed method to a flexible inverted pendulum and verify the effectiveness of the method with a numerical example.
language of the presentation: Japanese