IMU Data Based HAR Using Hybrid Model of CNN & Stacked LSTM

Abstract

Human activity recognition (HAR) keeps the track of various activities in different domains including health, security, sports and robotics. Inertial sensors – specifically, gyroscopes and accelerometers – can be used for these purposes and can now also be found in the accelerometer chip or inertial measurement unit of common smartphones and smartwatches. The widespread use of these smart devices necessitates the acquisition of activity data for HAR systems. Unlike previous studies, we propose the hybrid model of CNN & Stacked-LSTM which is an integration of convolutional neural networks (CNNs) and long-short term memory networks (LSTMs) to complete the task. Behind our design, lies the idea that the integrating architecture is capable of capturing the spatial features as well as temporal features of sequential data which are equally important for the prediction of the future values. We evaluate the model in benchmark dataset UCI-HAR to demonstrate its comparative performance with the existing work. Our proposed model achieves an accuracy of 93.4%, displaying its effectiveness. We trust that the architecture can be used in various social sectors attached with HAR.