Title: Human pose estimation based on region refined network

Authors: Minghui Wu; Pintong Zhao; Guangjie Zhang; Huifeng Wu

Addresses: School of Computer and Computing Science, Zhejiang University City College, Hangzhou, 310015, China ' School of Computer and Computing Science, Zhejiang University City College, Hangzhou, 310015, China; School of Computer, Zhejiang University, Hangzhou, 310027, China ' School of Computer and Computing Science, Zhejiang University City College, Hangzhou, 310015, China; School of Computer, Zhejiang University, Hangzhou, 310027, China ' Institute of Software and Intelligent Technology, Hangzhou Dianzi University, Hangzhou, 310018, China

Abstract: With the development of detection technology based on deep learning, the keypoint detection of the human body has gradually formed a theoretical system based on convolutional neural networks. In this context, the keypoints of the predicted and real values in the model follow the statistical deviation law of neighbourhood, and an improved model for refined prediction of keypoints is proposed. The original one-stage detection network is transformed into a two-stage end-to-end detection network. The detection error of the typical model in the keypoint neighbourhood is reduced, and the AP of the model on the COCO2017 dataset has an increase of 1 percentage point. This paper will detail in the improvement work the structure and parameters of the network, the training and prediction process of the network, the network supervision and loss function, and finally the experimental results on the COCO2017 dataset.

Keywords: human pose estimation; fully convolutional neural networks; region refined network.

DOI: 10.1504/IJIITC.2020.110281

International Journal of Intelligent Internet of Things Computing, 2020 Vol.1 No.2, pp.145 - 156

Received: 21 Nov 2019
Accepted: 30 Jan 2020

Published online: 25 Sep 2020 *

Full-text access for editors Access for subscribers Purchase this article Comment on this article