← Back
Solo researcher

Synthetic Data for Fall Detection

PythonPyTorchJupyterVAENumPy

Problem

Fall detection models need large, diverse training datasets, but real fall footage is scarce and sensitive to collect. Models trained on small datasets fail to generalize.

What I Built

Final year research project exploring variational autoencoders for generating synthetic fall detection training data. Pipeline covers frame preprocessing, chunk preparation, VAE training, and synthetic sequence generation.

Tech Decisions

  • Variational Autoencoder for learning latent representations of fall sequences from video frames
  • Multi-stage pipeline: frame preprocessing, chunk preparation, model training, and generation
  • Jupyter notebooks for reproducibility and iterative experimentation

Outcome

Built a VAE pipeline that generates synthetic fall sequences from real video frames for data augmentation.