A Reconfigurable Hardware Architecture for Graph Convolution Network in Action Recognition

Tsung Han Tsai, Tzu Chieh Chen

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Graph convolution neural network (GCN) has gained more attention in recent years. Data types like human skeleton joints which are in non-Euclidean space are suitable for GCN. However, due to the high computational complexity and data sparsity of GCN, it is common to have high latency or low power efficiency in CPU or GPU platforms. Therefore, dedicated hardware accelerators are critical for such tasks. In this paper, we proposed a highly parallelized and flexible architecture for the GCN layers of the Spatial-Temporal Graph Convolutional Networks (ST-GCN) model, which is a classic model and widely used in human action recognition. The accelerator also has high scalability due to our proposed method. Compared with the hardware implementation on ST-GCN, the proposed method reduces the latency by up to 39.5% and improves 1.46x on power efficiency.

Original languageEnglish
Title of host publication2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2023
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages1075-1078
Number of pages4
ISBN (Electronic)9798350300673
DOIs
StatePublished - 2023
Event2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2023 - Taipei, Taiwan
Duration: 31 Oct 20233 Nov 2023

Publication series

Name2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2023

Conference

Conference2023 Asia Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA ASC 2023
Country/TerritoryTaiwan
CityTaipei
Period31/10/233/11/23

Keywords

  • accelerator
  • FPGA
  • graph convolutional neural network
  • hardware
  • reconfigurable architecture

Fingerprint

Dive into the research topics of 'A Reconfigurable Hardware Architecture for Graph Convolution Network in Action Recognition'. Together they form a unique fingerprint.

Cite this