Human Activity Recognition with Anomaly Prediction for E-Health Systems using Lightweight AI

摘要

Computing devices should really be mindful of human actions in order for flourishing human-assistive systems to be used effectively in medical services and human-robot collaboration operations. It is necessary to create a trustworthy real-time action recognition mechanism for the uninterrupted and error-free functioning of such smart devices. This can only be done with simple, intelligent methods that make advantage of all-around sensors. With both the connection of time series information in mind, this work presents a fresh way to information architecture for improved extracting features. Using the assistance of a specially created iOS application, the movement data were gathered using a smartphone. In order to identify significant both spatial and temporal elements from the signals, data from eight operations were structured into single and double channels. The Fourier and wavelet domains were used to depict raw data alongside the time domain. Other CNN models that were employed to meet the deep-learning classification of the tasks fared worse than a deep neural network that utilized a double-channeled time-domain input. Several public datasets were used to further examine this strategy, and improved results were obtained. Ultimately, the learned model's applicability was evaluated in real time on a PC and a smartphone, yielding encouraging results.