Time-CNN and Stacked LSTM for Posture Classification

In current times, when work-from-home and remote learning have become part of our life, maintaining healthy lifestyle is very essential. The posture in which we are sitting or standing may affect our health, if not corrected consciously. In this paper, we propose a novel deep-learning model to classify nine different sitting and standing postures using data from accelerometers. Three tri-axis accelerometers are placed on the back of a person to record the data. A combination of time convolutional neural network (tCNN) and stacked long short-term memory (LSTM) is proposed for classification of the considered postures from the raw accelerometer data acquired from seven subjects. The performance of the proposed model is compared with that of individual tCNN and two different configurations of stacked LSTM models in terms of loss and classification accuracy. The combined network outperforms the other models, giving an average accuracy of 99.77%.