Atsutake Kosuge

Atsutake Kosuge

Auto-generated profile, based on publicly available information

Atsutake Kosuge

Graduate School of Engineering, University of Tokyo, Tokyo, Japan

Research Focus & Metrics

Academic Career Length

7Years (2020-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
18 May 2022 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
AKAtsutake Kosuge
Atsutake Kosuge
SSSatoshi Suehiro
Satoshi Suehiro
MHMototsugu Hamada
Mototsugu Hamada
TKTadahiro Kuroda
Tadahiro Kuroda
This article presents a millimeter wave (mmWave) imaging radar-based real-time multiclass object recognition system for advanced driver-assistance system (ADAS) applications related to construction machinery. While mmWave radar has the advantage of being able to detect signals even in the optically harsh, dark, and dusty environments in which construction machinery is often used, the radar has two orders of magnitude lower resolution than cameras. Since the distance from the radar increases, object features vary, resulting in making it difficult to classify and detect the location of multiple objects. To address this issue, an mmWave-you only look once (YOLO) architecture is proposed that enables highly accurate object classification and location recognition by applying different detectors to each distance data. To provide precise labels for the radar data semiautomatically, a camera-radar cooperative data annotator is also developed. Using the radar only, a real-time (46.6 ms) object classification and location detection of six class objects is achieved. The accuracy of our system is 84% [mean average precision (mAP)], which is slightly higher than that of the red, blue, green (RGB)-camera-based system (78%). In addition, since the radar can acquire similar data of target objects against any background, mmWave-YOLO can detect objects in a variety of scenes with only a small variation of the training dataset, resulting in a lower training cost. Experiments confirmed that it can detect objects in outdoor scenes even when it is trained only with indoor scene data.

10 December 2020 in IEEE Transactions on Instrumentation and Measurement
Verified
Journal Article
Closed access
VCViviana Crescitelli
Viviana Crescitelli
AKAtsutake Kosuge
Atsutake Kosuge
TOTakashi Oshima
Takashi Oshima
In this paper, we present POISON: an end-to-end Deep Neural Network (DNN) method to estimate the human pose in real-time and under variability of lighting conditions. A lightweight version of the open-source library OpenPose is used to extract human keypoints from RGB and Infra-Red (IR) images and the contribution of each image is combined by a fusion step. We propose a method to fuse these two types of information, using a fusion strategy followed by a refinement stage DNN, which aims to identify complex relationships between extracted keypoints. Finally, the combined information is used to infer the human pose. The experimental results validate that POISON improves the overall performance of conventional single-camera methods by a factor 1.79x. On a custom dataset of RGB/IR pair images captured in challenging low-light conditions, POISON correctly detects up to 86% of human keypoints, outperforming conventional single-camera methods, which can detect up to 48% of keypoints. We also provide a comparison of POISON against current state-of-the-art image fusion methods based on DNN techniques. We perform extensive quantitative evaluation and show that POISON outperforms existing approaches for the task of estimating the human pose in challenging lighting environments. POISON runs at 17fps on Intel Core i7 CPU with NVIDIA GeForce GTX-1080 GPU dedicated to DNN operations, making the solution suitable for real-time performance.

11 March 2020 in IEEE Transactions on Industrial Electronics
Verified
Journal Article
Closed access
AKAtsutake Kosuge
Atsutake Kosuge
KYKeisuke Yamamoto
Keisuke Yamamoto
YAYukinori Akamine
Yukinori Akamine
TOTakashi Oshima
Takashi Oshima
The conventional picking robots suffered from low picking throughput due to a large amount of computation of the object-pose-estimation algorithm which is called iterative-closest-point (ICP) algorithm. This work presents an FPGA-based ICP accelerator, which achieves 11.7-times-faster object-pose estimation by a picking robot compared with the state-of-the-art technique based on K-D-tree k-NN search and 4-core CPU. To accelerate the ICP, both algorithm-level and hardware-level techniques have been proposed and developed. The former is a hierarchical-graph-based k-NN search enabling simultaneous acquisition of plural neighboring points. The latter is a sorting-network-based circuit implemented on an SoC-FPGA. In addition, dynamic structural reconfiguration between the two key functionalities (graph generation and nearest neighbor search) is explored by utilizing the partial reconfiguration capability of FPGA to save required hardware resource. Experiments of the proposed FPGA-based ICP accelerator using Amazon Picking Challenge data sets have confirmed that the object-pose estimation by ICP takes only 0.72 seconds at the power consumption of 4.2 W.

20 September 2021 in IEEE Journal on Emerging and Selected Topics in Circuits and Systems
Verified
Journal Article
Closed access
AKAtsutake Kosuge
Atsutake Kosuge
MHMototsugu Hamada
Mototsugu Hamada
TKTadahiro Kuroda
Tadahiro Kuroda
A reconfigurable field-programmable gate array (FPGA)-based wired-logic deep neural network (DNN) accelerator is presented. High energy efficiency of 16 nJ/classification (Modified National Institute of Standards and Technology: MNIST) is achieved due to the wired-logic architecture. Each neuron in the neural network consists of combinational circuits only, and all the neurons are implemented on an FPGA. Intermediate data are never stored in memory or registers and are transmitted to the output stage by passing through only neuron cells. The latency and power required for memory access can be minimized, enabling low power and high throughput operation. A critical technical issue is reducing hardware resources because all neurons need to be implemented on an FPGA, where hardware resources are limited. Two core technologies have been developed to minimize the required hardware resources: (1) A neural network with a small number of neurons in which weight values of all synapses are fixed to a common value, and (2) a small neuron cell circuit consisting of an adder and look-up table containing an activation function. By fixing all weight values to a certain common value, processing can be simplified from multiply-accumulate operations to just additions, and the hardware resources required for each neuron are minimized. An experiment with the MNIST dataset using a 28-nm FPGA confirmed power consumption of 0.16 W and latency per inference of 100 ns (16nJ/classification). With the same recognition accuracy, the power efficiency is 45.6 times higher than that of the conventional state-of-the-art binarized DNN accelerator with a digital ASIC.

2 December 2022 in IEEE Journal of Solid-State Circuits
Verified
Journal Article
Closed access
KSKota Shiba
Kota Shiba
MOMitsuji Okada
Mitsuji Okada
AKAtsutake Kosuge
Atsutake Kosuge
MHMototsugu Hamada
Mototsugu Hamada
TKTadahiro Kuroda
Tadahiro Kuroda
A 0.7-pJ/bit, 8.5-Gb/s/link inductive coupling interchip wireless communication interface for a 3D-stacked static-random access memory (SRAM) has been developed in a 7-nm FinFET process. A new physical placement method that allows coils to be placed over off-the-shelf SRAM macros with small magnetic field attenuation, together with the use of synchronous communication using Manchester encoding and a clocked comparator to enable the detection of small-swing signals, achieves a 26% reduction in SRAM die area compared to through-silicon via (TSV)-based stacking. Interchip communication at 0.7 pJ/bit, 8.5 Gb/s/link was confirmed using test chips. A 4-hi 3D-stacked SRAM module using the proposed interface achieves a 1.2-TB/s/mm $^{2}$ area efficiency, representing a two orders-of-magnitude improvement over the state-of-the-art 3D-stacked SRAM.

30 September 2022 in IEEE Transactions on Circuits and Systems I: Regular Papers
Verified
Journal Article
Closed access
AKAtsutake Kosuge
Atsutake Kosuge
TKTadahiro Kuroda
Tadahiro Kuroda
This paper presents an overview of proximity wireless communication (PWC) technologies, their principles, design guidelines and practical applications. In particular, two different applications of PWC are reviewed. One is PWC between stacked chips. Both communication distance and coupler size are several tens of microns. Area and energy efficient design techniques are introduced. Another is PWC between module boards. Both communication distance and coupler size are several millimeters. Energy and area efficient practical designs are introduced for mobile and industrial machinery applications.