A Hybrid Deep Learning Neural Network for Recognizing Exercise Activity Using Inertial Sensor and Motion Capture System
Sakorn Mekruksavanich, Ponnipa Jantawong, Anuchit Jitpattanakul · 2023
The recognition of human activities holds great significance in the fields of AI and human-computer interaction studies. Human activity recognition (HAR) finds extensive application, including in behavioral assessment, gesture identification, fall detection, and gait evaluation. Recent strides in wearable technology and AI have showcased the potential to accurately classify movement events for HAR, achieved by analyzing time-series data collected from wearable sensors. This study focuses on deep learning networks, particularly CNN and LSTM, to extract spatial and temporal features from wearable sensor data for efficient exercise activity classification. A comprehensive assessment compared two sensor data types: 3-axis linear acceleration from IMUs and 3D acceleration from MOCAP markers. Findings highlight comparable accuracy between IMUs and MOCAP in exercise identification. The hybrid CNN-LSTM demonstrated high efficacy, achieving remarkable F1-scores of 98.85% for IMU and 98.23% for MOCAP data.