Complex Human Activities Recognition Based on High Performance 1D CNN Model

Abstract

Human activity recognition (HAR) is an emerging scientific research field that has wide area of applications in different fields such as healthcare, social-sciences and human-computer interaction etc. In many cases, humans perform very complex physical activities that needs to be tracked in order to improve well-being, quality of life and health. In this study, a method for complex HAR based on One dimensional (1D) CNN model using tri-axis accelerometer sensor data was proposed. The sensor data was collected from a smartwatch for three complex human activities which are studying, playing games and mobile scrolling. 1D CNNs provides high accuracy as well as less computational complexity in performing HAR. The proposed 1D CNN model was trained and optimized on a self-prepared dataset in Python. The adapted model provides an accuracy of 98.28 %. A preliminary study shows that the proposed model could effectively recognize the intended activities as a baseline for extending future work in the HAR area.