dr-pod/data-dive-campus-cams

UIUC Campus Cams (Apr 23 – Apr 29 2026)

Frame-level archive from three public UIUC YouTube live streams, captured at 1 Hz (after a brief startup ramp; see per-camera intended_interval_s). Paired with per-frame metadata (timestamp, brightness, capture status).

Capture window: 2026-04-23 05:27:28 UTC → 2026-04-29 23:41:52 UTC (~6.76 days, ~1.55M frames total across the three cameras, 489 hourly WebDataset shards).

Collected as part of an academic research project on UIUC class-release pedestrian surges and bus delays at signalized intersections (code: github.com/drPod/data-dive). Three companion archives released alongside this one:

• dr-pod/data-dive-mtd-gtfs-rt — CUMTD GTFS-Realtime feeds at 5 s polling.
• dr-pod/data-dive-google-directions — 8 campus OD pairs polled at 60 s.
• dr-pod/data-dive-cu-weather — KCMI hourly NWS observations (covariate).

Cameras

id	intended interval	framing	primary use
quad	1 s	Foellinger-terrace view of Main Quad X paths	pedestrian density + trajectory across the central campus X
alma	1 s	Alma Mater statue and the visitor plaza	per-visitor dwell-time and visit-count behavior at a campus landmark
morrow	1 s	Morrow Plots + Wright Street corridor	MTD bus-pass detection (for vehicle speed/trajectory) + sidewalk flow

All three cameras sample at 1 Hz after 2026-04-23T10:50Z. Earlier hours in the archive use the rates documented in each camera's camera_meta.json → capture.interval_change_log. Filter on the per-frame intended_interval_s field in the metadata to align analyses.

Per-camera location, bearing, and framing notes are in metadata/{cam}/camera_meta.json.

Format

Shards are WebDataset-compatible tars, one per camera per UTC hour:

data/{cam}/{YYYY-MM-DD}/{cam}-{YYYY-MM-DD}-{HH}.tar

Each tar contains pairs:

{cam}_{YYYY-MM-DD}_{HH-MM-SS}.jpg   — 1280x720 JPEG (q:v 5)
{cam}_{YYYY-MM-DD}_{HH-MM-SS}.json  — capture metadata

Per-frame metadata schema

{
  "ts": "2026-04-23T08:37:05+00:00",
  "ok": true,
  "bytes": 80802,
  "path": "/.../cam_frames/quad/2026-04-23/08-37-05.jpg",
  "camera_id": "quad",
  "intended_interval_s": 10,
  "resolution": "1280xauto",
  "brightness_mean": 0.2431,
  "attempts": 1,
  "err": null
}

brightness_mean is mean pixel luminance in [0,1] (downsampled to 1x1 grayscale by ffmpeg). Useful for filtering out night frames.

Example analyses this dataset supports

• Pedestrian density vs transit delay (our primary research use): correlate Quad person counts against MTD bus traversal times on adjacent corridors during class-release windows.
• Statue dwell-time behavior: for the Alma camera, detect individual visitors, track how long they linger, and plot dwell-time distributions by time-of-day, day-of-week, weather, and academic-calendar events (e.g., parents' weekend, graduation).
• Vehicle speed/trajectory from consecutive frames (Morrow): 1 s sampling with a bus visible for ~10–15 s per pass gives 10+ frames per passage — enough to estimate speed from pixel displacement and validate GTFS-Realtime position accuracy.
• Weather / daylight signal extraction: the per-frame brightness_mean field is a free diurnal / cloud-cover covariate.
• Action recognition on the Quad: cyclists and skateboarders crossing at ~5 m/s are visible for ~15 s, so 1 s sampling captures the full trajectory.

Known limitations

• Faces: pedestrians are visible but at 1280-wide resolution and camera distances typical of campus overhead views, faces are roughly 15–25 px tall — not identifiable. No individual-tracking labels are included.
• Night frames: low contrast; person-detection recall on these frames is poor.
• Alma framing: the camera points at the statue, not the adjacent street. Framing was documented before use, not pivoted.
• Morrow corridor: construction equipment (ARC site) is in the mid-ground through Spring 2026.
• Capture gaps: streaming failures occasionally drop frames. The index.jsonl records every attempt (success or failure) with ok and err fields.

Loader (WebDataset)

import webdataset as wds
from huggingface_hub import HfFileSystem

fs = HfFileSystem()
shards = fs.glob("datasets/dr-pod/data-dive-campus-cams/data/quad/2026-04-25/*.tar")
url_list = [f"https://huggingface.co/datasets/dr-pod/data-dive-campus-cams/resolve/main/{p[len('datasets/dr-pod/data-dive-campus-cams/'):]}" for p in shards]
ds = wds.WebDataset(url_list).decode("rgb").to_tuple("jpg", "json")
for img, meta in ds:
    # img is HxWx3 numpy, meta is the per-frame dict
    ...

Suggested citation

Darsh Poddar (2026). UIUC Campus Cams (Apr 23 – Apr 29 2026).
Hugging Face. dr-pod/data-dive-campus-cams.

License

Source streams are public UIUC YouTube live broadcasts. This archive is released for non-commercial research use. Downstream users should verify YouTube ToS and UIUC broadcast policy before commercial or redistribution use.

MTD attribution

This dataset was collected alongside an MTD GTFS-Realtime archive. Per the MTD API Terms of Use, any use of MTD data must include the following identification:

Data provided by the Champaign-Urbana Mass Transit District (CUMTD). For more information, visit mtd.dev.