Final year project (Systems Engineering, UNILAG). A pretrained pose estimator turns
video frames into skeletons; a classifier reads a short window of skeletons and says
fall or not fall. Three classifier types are compared on unseen people and timed on
edge hardware.

Repo: ~/code/school/skeleton-based-fall-detection. Proposal: project-details.md in the repo.

Read in order:

  1. 01-research-basics: aim vs objectives, what an evaluation protocol is, subject-wise splits and data leakage, the metrics (sensitivity, specificity, G-mean, F2, false alarms per hour) with worked numbers
  2. 02-pose-estimation: what a pose estimator outputs, MediaPipe’s 33 landmarks and visibility, VIDEO mode and timestamps, the first real finding (36% vs 61% detection)
  3. 03-preprocessing: gap filling, hip-centred torso-scaled normalisation, sliding windows and the window-count formula, labelling windows, what not to normalise
  4. 04-python-and-shell: uv, the src/ layout and python -m, writing files safely with .part + rename, commands used so far

Each file ends with In your own words: questions to answer in writing. Your answers go
in under the question, then the sharpened version.

See also (general vault): ai-ml · linux

↑ Back to Projects — Learning Logs