ゼミナール発表

日時: 9月30日(金)2限(11:00-12:30)


会場: L1

司会: 横田 太
大 洋介 1551022: M, 2回目発表 ロボティクス 小笠原 司,佐藤 嘉伸,高松 淳,丁 明
title: Model-free operation of appliances by the service robot using DCNN
abstract: In this paper, we propose a method to estimate the proper operation of appliances using deep convolutional neural networks (DCNN). We use Faster R-CNN to detect objects and perform an ontology analysis of the operating parts of the object (e.g. buttons). We experimentally verify the effectiveness of the proposed method with a robot operating in multi-objects environments.
language of the presentation: Japanese
 
馬場 隆造 1551082: M, 2回目発表 ロボティクス 小笠原 司,佐藤 嘉伸,高松 淳,丁 明

title:Information estimation of the operating object by the measurement of human motion

abstract:In this research, we propose a method to estimate the information of the target object for robot manipulation by measuring human motion. In this presentation, we focus on the motion pushing the box horizontally while human standing with two hands. We calculate the center of gravity and the position of hands from the measured motion data. Using a dynamic model, the pushing force is estimated by keeping the balance at all external force. In the experiment, we estimated the pushing force for various boxes with different weight. The effectiveness of this method was confirmed by comparing the measured and the estimated forces.

language of the presentation:Japanese

 
加藤 大真 1551033: M, 2回目発表 コンピューティング・アーキテクチャ 中島 康彦,岡田 実,高前田 伸也,TRAN THI HONG
title: Development of low complex IEEE 802.11ah Viterbi decoder for IoT application
abstract: To develop IoT sensor, research on low-cost and low-power wireless transceiver is necessary. Inside wireless transceiver, Viterbi decoder (VD) is one of the most complex but important block. VD is used in a receiver for decoding convolutional code from the transmitter and simultaneously correct bit errors. In addition, it is ceaselessly working circuit in a wireless transceiver, therefore it consumes a lot of power. This means that developing small and low power VD is one of the steps to make a wireless transceiver for IoT applications. In this research, we propose low complex VD while keeping the same error correction performance. Our VD employs M-algorithm to reduce amout of Viterbi computation, and also has improved architecture of the circuit. As an ASIC synthesis results, our hardware can reduce 51% of power consumption and 77% of the area.
language of the presentation: Japanese
発表題目: IEEE 802.11ah向け低電力Viterbiデコーダの開発
発表概要: IoT (Internet of Things)で用いられるワイヤレスセンサネットワークにおいては,電力の供給を厳しく制限されるケースが多い.センサの消費電力のうち,通信によって消費される電力が大半を占め,低消費電力かつ小規模な回路が希求されている.本研究では,IoT向けの次世代Wi-Fi規格,IEEE 802.11ah向けの低消費電力な通信機を開発するために,Viterbiデコーダの低複雑度化を行う.Viterbiデコーダはエラー訂正のための重要な要素であるが,既存のものは回路面積が大きく,また電力を多く消費する.我々は既存のViterbiデコーダに対し,Viterbi演算を簡略化するためのMアルゴリズムを導入するほか,回路アーキテクチャを改良した.回路をVerilog HDLで実装し,ASIC合成を行った結果,Packet Error Rate (PER)を維持したまま,既存の回路に対して77%の面積削減,51%の消費電力削減を達成した.
 

会場: L2

司会: 中島 悠太
岩口 優也 1551015: M, 2回目発表 光メディアインタフェース 向川 康博,横矢 直和,舩冨 卓哉,久保 尋之
title:Material estimation based on Depth Distortion of ToF camera
abstract:Recently, Time-of-Flight camera becomes popular. It can get distance by measuring the time of reflected light from object surface. But when measuring the distance to the translucent objects, the optical path is changed due to subsurface scattering. Therefore, Depth Distortion occurs in the result of the distance measurement. In this study, we purpose material estimation using Depth Distortion. Our experiment shows, Depth Distortion correspond to materials and possible for classification.
language of the presentation:Japanese
 
ZENG XINGJI 1551131: M, 2回目発表 光メディアインタフェース 向川 康博,横矢 直和,舩冨 卓哉,久保 尋之
title:Subsurface Scattering Parameter Estimation of Translucent Spherical Object
abstract: Shape acquisition (for instance normal map estimation from image) of translucent object is a difficult problem because of the complicated subsurface scattering. In this study, we aim to estimate the subsurface scattering parameters related to the acquisition of shape. Our key idea is to find the optimal parameters for Directional Dipole Model to the measured data. Unlike most previous work focus on planar surface, our method can also be adapted to spherical surface. By now, we have verified that our method works well in planar surface and simulative spherical surface.
language of the presentation: Japanese
 
秋山 解 1551001: M, 2回目発表 光メディアインタフェース 向川 康博,中村 哲,金出 武雄,舩冨 卓哉,伍 洋
title: Head gesture detection in spontaneous human conversations
abstract: Head gestures take important roles in human communication with various types representing different linguistic functions. However, only few types such as nod and shake were studied for automatic detection, which limits the real applications. This study aims at detecting more varied types of head gestures in spontaneous human conversations, verified on a new dataset built by ourselves. A state-of-the-art face tracker is adopted to extract per-frame head poses, whose temporal dynamics are modeled by a novel feature descriptor proposed by us. With this feature, we are able to generate some reasonable results using the state-of-the-art head gesture detector. Further exploration with deep neural networks are being investigated for possibly improved performance.
language of the presentation: Japanese