Human Activity Recognition using Convolutional Neural Network
Meenal Dugar · 2023
This study delves into Human Activity Recognition (HAR) using wearable inertial measurement units (IMUs), aiming to enhance classification accuracy in systems reliant on a single sensor. By leveraging convolutional neural networks and denoising autoencoders, the research seeks to outperform previous benchmarks and diminish dependence on manually curated features. Data from a single IMU, capturing seven activities like - running, jumping, walking, falling, sitting, standing, lying; was segmented using a time-window approach. Data augmentation techniques were employed to address class imbalances. The optimized 8-layer CNN model achieved an F1-Score of 0.972, notably improving accuracy for the underrepresented ’falling’ activity. The research’s implementations are made available online.