ゼミナール発表

日時: 09月25日(月)3限(13:30-15:00)


会場: L1

司会: 河合 紀彦
藤原 聖司 1651093: 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, 清川 清, 荒川 豊, 諏訪 博彦, 藤本 まなと
title: Implementation and Evaluation of Activity Recognition System using Analog-PIR-Sensor
abstract: Thanks to the development of ubiquitous information technology in recent years, smart home applications such as home appliance control, adult care sercice, and so on are attracting attention. To realize these applications, we have to develop a system which recognizes various human activities accurately and cheaply. There are many studies on the activity recognition in a smart home which uses a wearable device or a camera. We also have proposed an activity recognition technique in a smart home by utilizing digital-output-PIR sensors, door sensors, and power meters. However, the study has an unsolved issue: we cannot distinguish similar activities happening at the same place, for example, ``reading'' and ``using a smartphone'' while sitting on a sofa. In order to cope with this challenge, we introduce ALPAS: analog-output-PIR-sensor-based activity recognition technique which recognizes the different activities in the same place. Our technique recognizes user's activity by utilizing machine learning with frequency components of the sensor's output as features. However, because the number of features used in ALPAS is 1000 for each analog PIR-sensor, a large capacity memory is required. To reduce the number of features, we select a part of the sensing data. We call the starting point of the selected data as starting frequency (SF) and the ending point as ending frequency (EF). We searched SF and EF using a grid search, and evaluated the recognition accuracy. We evaluated the proposed technique in a smarthome testbed. In the evaluation, five participants performed four different activities while sitting on a sofa. As a result, we achieved F-Measure: 59.6 % without reducing the number of features. In addition, we achieved F-Measure: 63.9 % when the EF is 1.4 Hz, and F-Measure: 50 % or lower when the SF is 9.9 Hz or higher.
language of the presentation: Japanese
発表題目: アナログ出力焦電型赤外線センサを用いた行動認識システムの実装と評価
発表概要: 近年のユビキタス情報技術の発展に伴い,省エネ家電制御や高齢者見守りサービスなど,生活を支援するサービスへの応用が期待されている. これらのサービスの実現には,居住者の様々な屋内生活行動を高精度かつ安価に認識する必要がある. スマートホーム内における生活行動を認識するウェアラブルデバイスやカメラを用いた手法が多数提案されている.また,これまでにドア開閉センサや焦電型赤外線センサ, 消費電力センサを利用した生活行動認識手法が開発されている. しかし,これらの先行研究では,居住者の位置とその行動が1対1で対応しない位置非依存行動を高精度で認識できないという問題があった. そこで,本研究ではアナログ出力焦電型赤外線センサとそのセンサを用いた行動認識システム(ALPAS:AnaLog-output-Pir-sensor-based Activity recognition System)を開発し, 人の身振りの大きさなどユーザの動作の違いを検出することで,この問題を解決する. 提案手法では,人の動作の違いによりアナログ赤外線センサの周波数成分に違いが生じることに着目する.周波数成分を特徴量として用い,機械学習を用いて位置非依存行動を認識する. また,使用する特徴量の数を削減するために,本研究では特徴量の選定を行う.特徴量として用いる周波数帯としてstarting frequency (SF)とending frequency (EF)を定義し, 行動認識に最適なSFとEFをgrid searchによって探索し,認識率を評価する. 提案手法の有効性を評価するため,スマートホーム設備(1LDK)にて計5名の被験者にソファに着座した状態で4種類の異なる行動を行う実験を実施した. その結果,平均F値が59.6 %でユーザの行動を認識可能であることを確認した.加えて,EFが1.4Hzのとき平均F値は63.9 %に,SFが9.9Hz以上のとき平均F値が50 %を下回ることを確認した.
 
川上 蓮也 1651035: M, 2回目発表 ロボティクス 小笠原 司, 清川 清, 高松 淳, 丁 明
title: Refinment of disparity map by using Semantic Segmentation
abstract: Recently, there has been a demand for a vision sensor for quickly obtaining three-dimensional position information in fields such as automatic driving.Especially, since stereo cameras can obtain both RGB and distance informattion , they are used in various robots. However, the accuracy of distance estimation in stereo cameras is not yet sufficient. In this research, I propose a method to combine semantic segmentation to solve this problem.
language of the presentation: Japanese
 
豊島 健太 1651079: M, 2回目発表 ロボティクス 小笠原 司, 清川 清, 高松 淳, 丁 明
title: Development of a touch-care robot that touches humans
abstract: Physical contact is a very important means of communication. "Touch care" accompanied by physical contact has the effect that leads to the healthy growth of infants and the extension of healthy life span. Long-term care for the elderly is expected to be carried out by the robot in the future. Touch care by a robot is an important factor in carrying out quality care. In this research, we propose a touch-care system with a robot arm and report on explanation and results of a preliminary experiment to stroke.
language of the presentation: Japanese
 
山 亘 1661019: D, 中間発表 ロボティクス 小笠原 司, 清川 清, 高松 淳, 丁 明
title: Analysis of Deformable Object Manipulation Performed by a Human using Egocentric RGB-D Videos
abstract: The relationship between human hand activity and manipulated objects helps robots learn how to handle objects. Since the appearance of an object and the environment influence grasp and motion planning, we use an egocentric RGB-D camera to get images from the same point of view as a robot which has eyes on its head. This presentation shows a study on hand segmentation in egocentric RGB-D images. Hand segmentation using deep neural networks require a very large number of manually labeled training data to deal with person-specific hand appearance. We automatically generate hand labels from motion and depth information in a calibration gesture video using another network. Generated images of hand labels are then used for training an appearance network.
language of the presentation: Japanese
発表題目: 一人称視点RGB-D映像を用いた人の柔軟物操作の解析
発表概要: 人の把持及び手の動作と対象物の関係を理解することはロボットの動作生成に有用である.一人称視点映像を用いることにより,ロボットでの実行時と同一視点からの操作物体の状態と把持箇所,動作の関係を解析できる.本発表では,一人称視点RGB-D画像中から手領域を検出する手法について検討する.ディープニューラルネットで個人差などの色情報の違いに対応するため,大量に必要である手領域ラベルのついた学習データを自動的に生成する.キャリブレーションジェスチャ映像の動きと深度の情報を用いて手領域ラベルを生成するネットワークを前段に構成し,後段のネットワークで色情報からの手領域検出を学習する.
 

会場: L2

司会: 進藤 裕之
ADLIZAN BIN IBRAHIM 1651128: M, 2回目発表 サイバーレジリエンス構成学 門林 雄基 ☆, 笠原 正治, 小林 和真(客員), 河合 栄治(客員)
title: DDoS Attack Mitigation Mechanism using Entropy-based Approach in sFlow for Software Defined Network (SDN) Environment
abstract: Although a lot of security researches regarding SDN have been done over the years, DDoS attack, which is one of the main challenges to traditional Internet, still posed a huge threat to this maturing architecture. At the moment, to mitigate DDoS attacks, most current SDN controllers are able to perform anomaly detection algorithm based on the flow recorded feature of OpenFlow, then further analyze on these collected data and detect whether DDoS attacks had happened or not. However, this may only work well in small network environments because in large-scale networks, these collection and statistic processes may in turn overload the SDN controller. Plus, if quick detection is needed, it means that we need to increase the number of collected data and this consequently will further aggravate the overhead. Thus, in this project, we are utilizing packet sampling technology (sFlow) to relieve the overhead, combining it with entropy detection mechanism to counter the tradeoff issue related with sampling rate and detection accuracy of packet sampling technology via adaptive threshold.
language of the presentation: English
 
松井 琢朗 1651098: M, 2回目発表 インタラクティブメディア設計学 加藤 博一, 向川 康博, Christian Sandor, 武富 貴史
title: Fence Removal Based on Curvelet Transform for Diminished Reality
abstract: Diminished reality is a technique for visually removing real objects from captured images. Pre-captured background based methods and image inpainting based methods have been proposed to generate diminished images. In contrast to the previous methods, we propose a method for achieving diminished reality using curvelet transform which is one of frequency conversion. Especially, we focus on automatic fence region removal from captured images. In our method, the fence region is removed by optimization based on sparsity of curvelet coefficient. As a preprocessing of the fence removal, the fence region is automatically detected using convolutional neural network (CNN). In this presentation, we will show detection results and removal results.
language of the presentation: Japanese
発表題目: Curvelet変換を用いた網状物体の除去による隠消現実感
発表概要: 隠消現実感は取得した画像から不要な物体を視覚的に除去する技術である.これまでに,除去後の背景を事前に計測しておく手法や画像修復に基づく手法が提案されている.本研究では,フェンスのような網状物体の除去を目的とし,これらの物体を自動で検出,除去することが可能な手法を提案する.提案手法では,網状物体の写り込んだ画像と網状物体をマスキングした画像のペアを用いてCNNを訓練し,そのCNNを用いて網状物体を検出する.除去については周波数変換の一種であるCurvelet変換のスパース表現を最適化することで行う.本発表では,CNNを用いた検出とCurvelet変換を用いた網状物体の除去結果を示す.