Person-In-Bed Detection using Frequency Domain Features and GLR-based CuSum

We consider the problem of person-in-bed detection using accelerometer measurements in the segmented as well as streaming setting. For the segmented problem, we identify frequency domain features (4 features for each acceleration coordinate) that can be used to model the in-bed and not-in-bed hypotheses. We estimate the model parameters from the training data and apply the Generalized Likelihood Ratio (GLR) test. Using the same form as the GLR test statistic, we also propose an improvement using quadratic logistic regression. For the streaming problem, we model it as a sequential change detection problem using the models that we obtained for the in-bed and not-in-bed hypotheses and propose a GLRT-based Cumulative Sum (CuSum) algorithm.