Incept-LSTM: deep learning model for action recognition

Abstract Physical action recognition systems are among the computationally efficient human-machine interface systems that are necessary for human well-being. The various real-life applications such as health assistance, elderly care, personal care, make these systems more useful in today's scenario. In this paper we propose one such system based on deep learning and motion sensors for physical action recognition. A hybrid architecture that combines inception-based CNN and LSTM named as Incept-LSTM is implemented and experimented on UCI-HAR and WISDM datasets. In this work, additionally, we have acquired weakly labelled datasets from 19 participants who were engaged in 16 distinct activities while wearing wearable belts on their chests. It is observed that the proposed algorithm performed very well on all the three datasets. The proposed model obtained 96.3% accuracy on UCI-HAR, 99.4% on WISDM and 99% on locally collected weakly labelled dataset.