Attitude Estimation for Posture Detection in eHealth Services

Abstract

We investigate the influence of two attitude estimation methods for human posture detection. In the context of ADL (i.e. Activities of Daily Living) we analyze inertial sensors to reveal uncomfortable situations. Quantifying postures such as standing up, walking, lying down or sitting may feature people autonomy and well being. We report comparisons between two main attitude estimation strategies. Our experimental protocol uses a precise ground truth obtained from two annotators. The dataset involves 9 participants and provides 50 various data sequences. We discuss the obtained promising results, analyze advantages and limits when using attitude estimation in this context.

Learning personalized ADL recognition models from few raw data

“…A comparison is presented in Table 2. On average, the GRU accuracy of 0.742 is lower than the 0.807 achieved in [43]. Nevertheless, our model is not tuned for each user and no specific expert rule is applied.…”

“…Finally, each training starts with a learning rate of 0.01 which decreases by a factor 10 if the loss does not diminish during 10 epochs. In a previous paper using this dataset, Makni et al [43] compared two attitude device estimation algorithms and used expert rules and Kalman filters in order to estimate the individual postures. Our first experiment consists in reproducing these experiments using only machine learning and no expert a priori rules.…”

Personalized Posture and Fall Classification with Shallow Gated Recurrent Units

“…A comparison is presented in Table I. On average, the GRU accuracy of 0.742 is lower than the 0.807 achieved in [13]. Nevertheless, our model is not tuned for each user and no specific expert rule is applied.…”

“…We chose to concentrate on two datasets on which results have already been published so we can make comparisons. We trained our GRU models on a dataset called Postures used by the authors of [13]. Unlike this article, we trained personalized neural network models and so did not define expert rules to classify the postures.…”

“…Dataset: Makni et al [13] compared two attitude device estimation algorithms and used expert rules in order to estimate the individual postures. Our first experiment consists in reproducing these experiments using only machine learning and no expert a priori rules.…”