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.