A Visual Tracking Deep Convolutional Neural Network Accelerator

DOI

10.2991/amcce-17.2017.87How to use a DOI?

Keywords

Convolutional neural network, visual tracking, accelerator.

Abstract

Convolutional Neural Network (CNN) has been widely used in computer vision problems including image classification, object detection and tracking and semantic segmentation which has obtained state-of-art results. However, CNN is both computation and bandwidth intensive. At present, we usually train and use CNN on CPU, GPU and even GPU clusters which are impractical for embedded platforms. Therefore, we need design CNN accelerator for low power, high computation and high bandwidth processing. There are already some accelerator which is mainly accelerate 2-D convolution, but not high energy-efficiency for 3-D convolution. So we propose an architecture which is effective for large depth 3-D convolution and we support three different scales kernel size. We use a visual tracking CNN algorithm-MDNet to verify our architecture and make no accuracy loss. We evaluate our system using 65nm library which has a small footprint of 2.58 mm2 and 224 mW.

Copyright

© 2017, the Authors. Published by Atlantis Press.

Open Access

This is an open access article distributed under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).

Download article (PDF)

TY  - CONF
AU  - Zhiyong Qin
AU  - Lixin Yu
PY  - 2017/03
DA  - 2017/03
TI  - A Visual Tracking Deep Convolutional Neural Network Accelerator
BT  - Proceedings of the 2017 2nd International Conference on Automation, Mechanical Control and Computational Engineering (AMCCE 2017)
PB  - Atlantis Press
SP  - 493
EP  - 499
SN  - 2352-5401
UR  - https://doi.org/10.2991/amcce-17.2017.87
DO  - 10.2991/amcce-17.2017.87
ID  - Qin2017/03
ER  -