Human activity recognition using inter-joint feature fusion with svd

Abstract

In this paper, a multi-feature descriptor for human activity recognition (HAR) was presented. The joints of the human skeleton were extracted from RGB images by using OpenPose to develop a robust multi-feature descriptor. Three features which are joint-joint angle, joint-joint horizontal distance, and joint-joint vertical distance were calculated. For the ease of computational cost, the singular value decomposition (SVD) was performed. In order to obtain singular values representing one full cycle of activity without information loss, the matrix sizes were equalized by zero paddings and row shifting. The singular values obtained from SVD form the final descriptor. The authors evaluated the performance of the proposed method on the well-known KTH and Weizmann datasets. The experimental results showed that the proposed descriptor gives out state-of-the-art results in human action recognition. © 2021 Elsevier B.V., All rights reserved.