Time-Frequency Based Feature Extraction For Automated Classification Of Bed Occupancy Using Accelerometer Data

Abstract

The accurate detection of bed occupancy is crucial in healthcare applications. In this study, we apply time-frequency analysis using Wavelet decomposition to analyze 3-axis accelerometer data and predict bed occupancy. We tested various wavelet filters, decomposition levels, and segment sizes to identify optimal parameters to maximize the prediction accuracy for bed occupancy. Accurate segmentation followed by extraction of features in the transform domain for each segment ensures accurate detection of transitions between "in-bed" and "not-in-bed" states. Results show that combining wavelet transform with ensemble models like the Random Forest Classifier (RFC) significantly improves detection accuracy. The highest accuracy achieved was 0.98952 for Track 1 (classification for pre-chunked data) and 0.93214 for Track 2 (classification for streaming data).