Utilization of Postural Transitions in Sensor-based Human Activity Recognition

Sensor-based human activity recognition (HAR) has gained tremendous attention due to numerous applications that aim to monitor the movement and behaviour of humans. However, the occurrence of transitions between activities gives rise to many problems in HAR as they can affect the performance of the recognition system by causing fluctuations in the prediction. This paper proposes an HAR system using groups of similar postural transitions, discrete wavelet transform (DWT) and bidirectional long short-term memory (BiLSTM) to deal with the postural transitions, thereby improving the accuracy of the system. In this system, the transitions which have similar patterns are grouped into same groups, after that the essential features are extracted by using DWT before being fed into the BiLSTM network for activity classification task. Our experiment results indicate that the proposed model achieves competitive performance compared to a non-transition model.