Title: Human action recognition using spatio-temporal skeletal data

Authors: Awadhesh Kumar Srivastava; K.K. Biswas

Addresses: Uttarakhand Technical University, Post Office, Chandanwadi, Prem Nagar, Sudhowala, Dehradun (Uttarakhand), India; KIET Group of Institutions, Ghaziabad-Meerut Road (NH-58) Muradnagar-201206, Distt. Ghaziabad, India ' Bennett University, Plot Nos 8-11, TechZone II, Greater Noida 201310, Uttar Pradesh, India

Abstract: Human action recognition from video is an important task with multiple challenges like cluttered background, luminance, occlusions, etc. Availability of depth sensor like Kinect makes the action recognition task a bit easy but it brings new challenges in terms of computation cost and noise. We present a novel, computationally economical but effective method for human activity recognition using skeleton data. We consider the relative changes in body parts positions to recognise the activity in the video and propose 'sum of temporal differences of joint-pair-distances' (STD) as feature descriptors. Further, we show that using random forest as a classifier with these features can produce better accuracies compared to various recent state of the art methods. We establish this by experimenting with publicly available MSR-action 3D dataset and MSR-daily activity datasets. The results show that proposed method archives accuracies of 93.9% in former dataset while 87% in latter dataset.

Keywords: surveillance; tracking; RGB video; human gesture; activity recognition; depth data; skeleton data; MS-kinect.

DOI: 10.1504/IJAIP.2024.140090

International Journal of Advanced Intelligence Paradigms, 2024 Vol.28 No.3/4, pp.253 - 271

Received: 23 Jun 2018
Accepted: 19 Nov 2018
Published online: 24 Jul 2024 *

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