An integrated solution for human activity recognition in smart indoor environments using embedded wearable devices and deep learning

Abstract Over the past years, the interest in using wearable and mobile devices for smart healthcare and living solutions has led to intensive research on Human Activity Recognition (HAR). While many research studies on HAR solutions have achieved significant practical results, there is still room for further optimization and improvement in terms of equipment cost, user-friendliness, and identification accuracy. This study presents an innovative Internet of Things (IoT)-based HAR solution that leverages wearable devices and deep learning (DL) techniques to recognize daily indoor activities. The device in this solution integrates an inertial measurement unit (IMU), a clock module, and a ZigBee communication module for efficient data transmission and activity monitoring. Meanwhile, a model combining the capabilities of a multibranch convolutional neural network (CNN) and bidirectional gated recurrent unit (BiGRU) is proposed for automatic feature extraction and activity classification across five activities (running, jogging, sitting, upstairs, and downstairs). We demonstrate the effectiveness of our approach on our local human activity dataset, achieving a recognition accuracy of 97.44%. Furthermore, we show that this solution could gather data and give post-activity feedback, enabling smart indoor healthcare and living applications to overcome the limitations of current solutions and drive HAR solutions toward improved efficiency and reliability.