cGAN-Based High Dimensional IMU Sensor Data Generation for Enhanced Human Activity Recognition in Therapeutic Activities
Authors:
Mohammad Mohammadzadeh,
Ali Ghadami,
Alireza Taheri,
Saeed Behzadipour
Abstract:
Human activity recognition is a core technology for applications such as rehabilitation, health monitoring, and human-computer interactions. Wearable devices, especially IMU sensors, provide rich features of human movements at a reasonable cost, which can be leveraged in activity recognition. Developing a robust classifier for activity recognition has always been of interest to researchers. One ma…
▽ More
Human activity recognition is a core technology for applications such as rehabilitation, health monitoring, and human-computer interactions. Wearable devices, especially IMU sensors, provide rich features of human movements at a reasonable cost, which can be leveraged in activity recognition. Developing a robust classifier for activity recognition has always been of interest to researchers. One major problem is that there is usually a deficit of training data, which makes developing deep classifiers difficult and sometimes impossible. In this work, a novel GAN network called TheraGAN was developed to generate IMU signals associated with rehabilitation activities. The generated signal comprises data from a 6-channel IMU, i.e., angular velocities and linear accelerations. Also, introducing simple activities simplified the generation process for activities of varying lengths. To evaluate the generated signals, several qualitative and quantitative studies were conducted, including perceptual similarity analysis, comparing manually extracted features to those from real data, visual inspection, and an investigation into how the generated data affects the performance of three deep classifiers trained on the generated and real data. The results showed that the generated signals closely mimicked the real signals, and adding generated data resulted in a significant improvement in the performance of all tested networks. Among the tested networks, the LSTM classifier demonstrated the most significant improvement, achieving a 13.27% boost, effectively addressing the challenge of data scarcity. This shows the validity of the generated data as well as TheraGAN as a tool to build more robust classifiers in case of imbalanced and insufficient data problems.
△ Less
Submitted 14 February, 2024; v1 submitted 15 February, 2023;
originally announced February 2023.
