simhacker/moollm/urban-safari

Urban Safari — ride map + video sync pipeline

Lifted from a private ebike workflow into a shareable MOOLLM skill. FIT files in, static JSON for a map app out, optional video clips synced to GPS with frame paths drawn over the track.

Part of MOOLLM. See skills/README.md.

Quick start

cd skills/urban-safari
pip install fitparse
 
# Any FIT files in ./rides/
python scripts/build_web_assets.py --trips-dir ./rides --out ./web/data
 
# Bosch Flow iPhone (USB)
pip install pymobiledevice3
python scripts/pipeline.py --sync --trips-dir ./rides --out ./web/data \
  --home-label "Your Home" --home-lat 52.33 --home-lon 4.78
 
# Video sync (needs ffprobe)
python scripts/sync_video.py --videos-dir ./videos --data-dir ./web/data

Synthetic demo (no real GPS): examples/demo/web/data/.

Protocol

SYNC — pull rides from device

BUILD — FIT → web JSON

build_web_assets.py writes manifest.json + per-trip GeoJSON, series, meta. Bosch duplicate rows merged per timestamp (see bosch-fit-quirks.md).

INSPECT — diagnose a FIT

python scripts/inspect_fit.py ride.fit

SYNC_VIDEO — align clips to trips

1. ffprobe reads video creation_time and duration
2. Auto-match picks trip with best wall-clock overlap
3. Interpolate GPS from trip series at each video second
4. Write videos/{id}.track.json + videos/{id}.geojson (frame path)

python scripts/sync_video.py --video clip.mp4 --trip-id my-ride --data-dir ./web/data
python scripts/sync_video.py --videos-dir ./videos --data-dir ./web/data  # auto-match all
python scripts/sync_video.py --video clip.mp4 --offset-s 2.5 --data-dir ./web/data  # clock skew

Calibration trick: film your bike computer clock for one second at ride start; measure offset between video frame and FIT timestamp; pass --offset-s.

MAP_TRANSCRIPT — words on the map, clustered

Requires synced video track (SYNC_VIDEO). Whisper transcribes the clip; each segment gets GPS from the video keyframe at its temporal midpoint; nearby segments merge into readable labels.

python scripts/map_transcript.py --video-id img-0679 --data-dir ./web/data \
  --video ~/Movies/IMG_0679.MOV \
  --cluster-radius-m 40 --cluster-gap-s 20 --min-words 3

Clustering merges segments when within 40 m and 20 s (defaults). Tune for urban density: tighter radius in old city centers, wider on open dike roads.

Or full pipeline with transcription:

python scripts/pipeline.py --sync --videos-dir ./videos --out ./web/data --map-transcripts

Output: videos/{id}.transcript.json + Point GeoJSON for map labels. See DATA-CONTRACT.md.

MATCH — map UI integration

Read DATA-CONTRACT.md. Load manifest.json, fetch trip GeoJSON for all rides, union bounds for overview map. On video play, scrub marker along keyframes[].

Web app sketch (SvelteKit + MapLibre)

1. Static copy of web/data/ → static/safari/data/
2. fetch('/safari/data/manifest.json')
3. Draw all trips[].geojson as layers (color by trip)
4. For selected trip + video: draw video track LineString; <video> timeupdate → marker

Privacy

FIT and video contain exact GPS. Do not commit real tracks to public repos. Use examples/demo/ for public demos; trekify before sharing.

Device adapters

See adapters/README.md. XOSS integration is a natural collaboration with Abraham Moller — his BLE sync lands FITs; this skill builds the shared map + video layer.

See also

• DATA-CONTRACT.md
• bosch-fit-quirks.md
• simulation/examples/urban-safari-ride-game.yml