e-hir GAN

e-hir GAN augments micro-Doppler radar datasets using generative adversarial networks (GANs) to increase training data for human motion classification and related biomedical analyses.

Key Features:

• Data Augmentation with GANs: e-hir GAN employs GANs to generate synthetic micro-Doppler signatures to enlarge training datasets for deep learning.
• Data Collection: Authentic micro-Doppler radar data were collected for seven distinct human activities, each yielding 144 samples.
• Training Process: GANs were trained for 3,000 epochs with generated outputs saved every 300 epochs and best images selected for use.
• Data Synthesis: Each original dataset was augmented to produce 1,472 synthesized spectrograms of size 64 × 64.
• Computational Environment: Training and data generation used GPU drivers, CUDA libraries, cuDNN, Anaconda, Keras-GPU, SciPy, Pillow, OpenCV, Matplotlib, and Git.
• Improved Classification Accuracy: Augmented datasets were used to enhance deep neural network training, improving human motion classification accuracy.

Scientific Applications:

• Human Motion Classification: Provides increased training data for classifying human activities from micro-Doppler radar signals.
• Musculoskeletal Disease Diagnosis: Supports motion-analysis-based approaches for diagnosing musculoskeletal diseases.
• Rehabilitation Therapy Development: Enables augmentation of datasets used to develop and evaluate rehabilitation therapies.
• Fall Detection: Enhances datasets for detecting falls from radar-based motion signatures.
• Energy Expenditure Estimation: Facilitates estimation of energy expenditure from augmented motion spectrograms.
• Biomechanical Research and Safety Monitoring: Augmented data support biomechanical studies and safety-monitoring applications.

Methodology:

Collected authentic micro-Doppler radar data for seven activities (144 samples each); trained GANs for 3,000 epochs saving outputs every 300 epochs and selecting best images; generated 1,472 synthesized 64 × 64 spectrograms per original dataset; training used GPU drivers, CUDA, cuDNN, Anaconda, Keras-GPU, SciPy, Pillow, OpenCV, Matplotlib, and Git.