Year: 2018
Type: conference-paper
Abstract: Most studies on wearable sensing assume that each sensor is correctly placed on the body, fixed to a pre-determined position at a pre-determined orientation.This is not practical and feasible in many applications where elderly, disabled, or injured people need to place the sensor units on their own, especially for wireless and small sensor units.It is a considerable improvement to make wearable systems robust against the placement orientations of the sensors, and further, to allow the sensors to be placed at any orientation.For this purpose, we propose a transformation based on Singular Value Decomposition (SVD) that removes the absolute orientation information from the sensor data.We apply this transformation in the preprocessing stage of the standard human activity recognition scheme using multiple publicly available datasets, classifiers, and cross-validation techniques, and achieve an average accuracy that is only 7.56 % lower than the reference approach with fixed sensor orientations.The most common method that is an alternative to the proposed transformation is taking the Euclidean norm of 3D data vectors, which obtains 13.50 % lower accuracy than the reference approach.We show that randomly oriented sensors cause a reduction of 21.21 % in the activity recognition accuracy when no transformation is applied for orientation invariance; hence, the standard system cannot handle incorrectly oriented sensors.On the other hand, the proposed approach allows users to place the sensor units at any orientation on their body with an acceptable reduction in the accuracy, outperforming the common Euclidean norm approach.
Source: Lirias
Authors: Aras Yurtman, Billur Barshan
Institution: Bilkent University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2964726279