Southeast Asia helmet-cam
Continuous forward-facing video from working motorbike drivers across Southeast Asian cities: dense mixed traffic, alleys, markets, monsoon rain and night riding. Each clip has one metadata row covering place, time, conditions, capture settings and consent.
Request access
The figures and rows on this page come from synthetic sample metadata, not the production dataset.
- 429
- hours of video
- 5,000
- clips
- 420
- drivers
- 10
- cities
- 5
- countries
- 2.5 TB
- total size
Hours by city
- Jakarta132 h
- Ho Chi Minh City57 h
- Surabaya45 h
- Bandung40 h
- Bangkok40 h
- Hanoi35 h
- Medan26 h
- Denpasar23 h
- Manila19 h
- Kuala Lumpur12 h
Hours by time of day
- morning118 h
- afternoon100 h
- midday94 h
- evening73 h
- night44 h
Hours by weather
- clear213 h
- overcast101 h
- rain85 h
- heavy rain22 h
- haze8 h
Metadata columns
One row per clip, stored as Parquet. Video is MP4 (H.264).
| Column | Type | What it holds |
|---|
| clip_id | Utf8 | Stable identifier for the clip. |
| driver_id | Utf8 | Pseudonymous driver identifier. Use it to split train and test by driver. |
| country | Utf8 | Country where the clip was recorded. |
| city | Utf8 | City where the clip was recorded. |
| captured_at | Timestamp(ms, "UTC") | Start of recording, UTC. |
| duration_s | Float32 | Clip length in seconds. |
| fps | Int32 | Frames per second. |
| width | Int32 | Frame width in pixels. |
| height | Int32 | Frame height in pixels. |
| codec | Utf8 | Video codec. |
| time_of_day |
Sample rows
| avg_speed_kmh | captured_at | city | clip_id | codec | consent_version | country | driver_id | duration_s | faces_blurred | fps | has_imu | height | plates_blurred | road_type | size_bytes | start_lat | start_lon | time_of_day | video_uri | weather | width |
|---|
| 17.06 | 2026-04-18T23:37:57.753Z | Ho Chi Minh City | clip_c49ad551 | h264 | driver-consent-v1 | Vietnam | drv_00345 | 332.23 | true | 60 | false | 1440 | true | residential | 835288606 | 10.81 | 106.64 | night | s3://datachad-lake/datasets/sea-helmet-cam/clips/ho-chi-minh-city/clip_c49ad551.mp4 | overcast | 2560 |
| 28.57 | 2026-03-17T20:05:49.366Z | Surabaya | clip_f02ad2de | h264 | driver-consent-v1 | Indonesia | drv_00305 |
Load it in PyTorch
Filter the Parquet metadata first, then decode only the matching clips. The loader reads from local disk or S3.
import pyarrow.compute as pc
from torch.utils.data import DataLoader
from helmet_cam import HelmetCamClips
clips = HelmetCamClips(
"s3://datachad-lake/datasets/sea-helmet-cam/metadata/",
filter=(pc.field("city") == "Jakarta") & (pc.field("weather") == "rain"),
num_frames=16,
size=(224, 224),
)
loader = DataLoader(clips, batch_size=8, num_workers=8, shuffle=True)
for batch in loader:
video = batch["video"] # uint8, [8, 16, 3, 224, 224]