DynaFall experiments

This repository implements the experiment plan in PLAN.md for skeleton-based fall detection.

Structure

data/raw/              Raw videos: data/raw/URFD and data/raw/MCFD
data/poses/            Extracted YOLO pose files
data/processed/        Video-level splits and 32-frame clips
configs/default.yaml   Main experiment config
src/dynafall/          Dataset, features, models, training, evaluation
scripts/               Entry-point scripts
results/               Metrics, checkpoints, tables

Quick smoke test

python scripts/make_synthetic_dataset.py --dataset Synthetic --videos 24
python scripts/prepare_clips.py --dataset Synthetic
python scripts/train.py --dataset Synthetic --method dynafall --epochs 2
python scripts/evaluate.py --dataset Synthetic --method dynafall

Real data workflow

Place videos under:

data/raw/URFD/fall/*.avi
data/raw/URFD/nonfall/*.avi
data/raw/MCFD/fall/*.avi
data/raw/MCFD/nonfall/*.avi

Any common video extension is accepted. Labels are inferred from the parent directory name: fall, falls, 1, positive map to fall; all other directory names map to non-fall.

Then run:

python scripts/extract_pose.py --dataset URFD
python scripts/prepare_clips.py --dataset URFD
python scripts/run_experiments.py --dataset URFD --methods lstm stgcn agcn ctrgcn posec3d tcnte dynafall
python scripts/robustness.py --dataset URFD --methods stgcn agcn ctrgcn posec3d tcnte dynafall
python scripts/aggregate_results.py

Repeat for MCFD. Cross-dataset evaluation:

python scripts/evaluate.py --dataset MCFD --method dynafall --checkpoint results/URFD/dynafall/best.pt --tag trainURFD_testMCFD

Notes

The graph baselines are compact reimplementations designed for small fall datasets and a COCO-17 pose layout. They preserve the paper-level comparison categories: LSTM, ST-GCN-style graph temporal model, two-stream adaptive GCN, CTR-GCN-style channel topology refinement, PoseC3D-style heatmap volume, TCN+Transformer, and DynaFall-GCN.