Fall detection with a wearable camera using traditional and deep learning algorithms

Helmus, Mark (2021) Fall detection with a wearable camera using traditional and deep learning algorithms. Master's Thesis / Essay, Computing Science.

Preview	Text mCS_2021_HelmusMA.pdf Download (27MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (95kB)

Abstract

A fall can have severe consequences to one's life, even leading to death if no adequate help is provided. This help can only be provided if the fall is detected. In this research, two traditional algorithms and one deep learning algorithm are proposed to automatically detect falls from videos filmed with a wearable camera. The traditional algorithms used are dense trajectories (DT) and improved dense trajectories (IDT). The compressed video action recognition algorithm (CoViAR) is the deep learning algorithm used. The algorithms are trained and tested on the data set showing different actions, filmed with the camera mounted to the waist or neck. In total, the data set consists of 1459 videos. After optimizing the hyperparameters of the classifiers, CoViAR outperforms the traditional algorithms on each problem. When detecting falls using videos filmed with cameras mounted to either the neck or waist, CoViAR has a ROC-AUC of 99.82% and accuracy of 97.95%. The traditional algorithms can be used to detect falls as well. A ROC-AUC of 97.58% and 98.47% for respectively DT and IDT is obtained. Further, no evidence is found that a specific camera mounting point is preferred over another. Evaluating the results, the three selected algorithms can be used to detect falls from videos filmed with a wearable camera. The deep learning algorithm has a slight edge over the traditional algorithms, since it has better performance and does not need to compute optical flow explicitly.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Azzopardi, G. and Wang, X. and Talavera Martinez, E.
Degree programme:	Computing Science
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	07 Jan 2021 10:12
Last Modified:	07 Jan 2021 10:12
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/23782

Actions (login required)

View Item