HARDeep: design and evaluation of a deep ensemble model for human activity recognition. (2nd June 2023)

• Record Type: Journal Article
• Authors: Subramanian, R. Raja`; Vasudevan, V.
• Is Part Of: International journal of innovative computing and applications. Volume 14:Number 3(2023)
• Page Start: 155
• Page End: 166
• Publication Date: 2023-06-02
• Subjects: activity recognition -- deep ensemble model -- unmanned aerial vehicles -- UAV -- scene stabilisation -- pretrained models -- HARDeep -- human detection -- YOLO -- fog computing; Evolutionary computation -- Periodicals; Neural networks (Computer science) -- Periodicals; Genetic programming (Computer science) -- Periodicals; Biologically-inspired computing -- Periodicals; Swarm intelligence -- Periodicals; Quantum computers -- Periodicals; 006.3
• Languages: English
• ISSNs: 1751-648X

Abstract

With the emergence of smartness in various fields including medical science, forensics and security, remote monitoring of human activities has gained more interests in research. The ambulatory health monitoring services includes monitoring the activities of mentally challenged and elderly people. In this research paper, we propose a novel framework for activity recognition from video sequences captured from static cameras and those captured from UAVs. The proposed framework, named HARDeep, consists of three models: an optional scene stabilisation model for UAV captured video sequences, a human detection model leveraging YOLOv3, and, to extract the set of video frames containing humans, an activity recognition model leveraging the ensemble of three deep learning models: GoogleNet, ResNet-50, and ResNet-101. HARDeep is evaluated against three datasets including Hollywood2, KTH and the UCF-ARG dataset, consisting of video sequences captured from UAVs. The recognition accuracies are compared with the various inference models leveraging wide learning paradigms.