A Comparative Study on Sensor Displacement Effect on Realistic Sensor Displacement Benchmark Dataset

Abstract

Activity Recognition (AR) research is growing and plays a major role in various fields. The approach of using wearable sensors for AR is well-accepted, as it compensates the need to install cameras in image processing approach which can lead to privacy violation. Using wearable sensors can suffer from one disadvantage – sensor displacement. There have been a number of research which studies sensor displacement problem. However, the conclusion cannot be made as which classifier is better than another in recognizing displacement data, as the prior experiments were performed under different conditions and focused on different parts of the body. This work aims to evaluate recognition performance of different algorithms – SVM, C4.5, and Naïve Bayes – on ideal-placement and displacement data on whole body activities, by adopting REALDISP dataset to make such evaluation. The accuracy of all algorithms on ideal placement data was above 90%, where SVM yielded the highest accuracy. Displacement data were tested against classification models constructed from ideal-placement data. The results shows that there was a dramatic drop in recognition performance. The accuracy of all algorithms on displacement data was between 50-60%, and C4.5 could handle displacement data the best.