Optimum signal duration for Human Activity Recognition based on Deep Convolutional Neural Networks

Human Activity Recognition (HAR) stands as a pivotal technique within pattern recognition, dedicated to deciphering human movements and actions utilizing one or multiple sensory inputs. Its significance extends across diverse applications, encompassing monitoring, security protocols, and the development of human-in-the-loop technologies. However, prevailing studies in HAR often overlook the integration of human-centered devices, wherein distinct parameters and criteria hold varying degrees of importance compared to other applications. Notably, within this realm, curtailing the sensor observation period assumes paramount importance to safeguard the efficiency of exoskeletons and prostheses. This study embarks on the optimization of this observation period specifically tailored for HAR using Inertial Measurement Unit (IMU) sensors. Employing a Deep Convolutional Neural Network (DCNN), the aim is to identify activities based on segments of IMU signals spanning durations from 0.1 to 4 seconds. Intriguingly, the outcomes spotlight an optimal observation duration of 0.5 seconds, showcasing an impressive classification accuracy of 99.95%. This revelation holds immense significance, elucidating the criticality of precise temporal analysis within HAR, particularly concerning humancentric devices. Such findings not only enhance our understanding of the optimal observation period but also lay the groundwork for refining the performance and efficacy of devices crucially relied upon for aiding human mobility and functionality.