visaitech/vehicle-mixed-traffic-detection

Visaitech Mixed-Traffic Vehicle Detection Dataset (v0.1)

Dashcam frames annotated for pedestrian / 2-wheeler / 3-wheeler / 4-wheeler detection in South Asian mixed traffic, a class taxonomy general-purpose COCO-trained detectors don't cover (COCO has no concept of an auto-rickshaw or motorcycle-vs-bicycle-as-one-class "2-wheeler" grouping tuned for how this traffic actually mixes on the road).

This is an early v0.1 release: 293 annotated frames from 6 source videos, published now so the data and annotation format are public and versioned. The full anonymized raw clip archive (203 clips, ~12GB, unannotated) ships alongside it in videos/, with annotations for it to follow in future releases (see Limitations below).

Sample frames

One mid-clip frame from six of the raw clips in videos/, exactly as published: faces and license plates pixelated, timestamp band blacked out (see Privacy and anonymization below).

Motorcycles sharing the lane with cars (clip_040)	Stopped congestion at close range (clip_015)
Cargo rickshaws lining a bright street (clip_120)	Multi-lane arterial, mixed vehicles (clip_140)
Dense bazaar street, 2- and 3-wheelers (clip_160)	Squeezing past a truck at close range (clip_129)

Classes

id	class	description
0	pedestrian	Person on foot
1	2-wheeler	Motorcycle / scooter / bicycle
2	3-wheeler	Auto-rickshaw / three-wheeled vehicle
3	4-wheeler	Car / van / bus / truck

Occlusion is not a separate class. Partially-visible boxes (tagged ... half by the original annotators) are folded into their base class, with the flag preserved per-box in attributes.csv (occluded column) for anyone who wants to filter or weight by visibility.

Dataset structure

images/{train,val}/*.jpg   # source video id prefixed, e.g. V003__frame.jpg
labels/{train,val}/*.txt   # YOLO format: class x_center y_center width height (normalized)
data.yaml                  # Ultralytics-ready config
classes.txt                # class id -> name, one per line
attributes.csv             # image, class, occluded (per-box)
videos/clip_001.mp4 ...    # 203 anonymized raw dashcam clips (~12GB, unannotated),
                           # flat sequential naming, ~2 minutes each
samples/clip_XXX.jpg       # preview frames from six of the clips (see Sample frames)

Directly usable with Ultralytics:

from ultralytics import YOLO
model = YOLO("yolo11n.pt")
model.train(data="data.yaml", epochs=100, imgsz=640)

Statistics

• 293 images (640x480), 1,506 boxes, split by whole source video (not by frame) to avoid near-duplicate leakage between train/val.
• Split: 239 train / 54 val images.

class	train	val	total	occluded
pedestrian	30	9	39	0
2-wheeler	731	128	859	121
3-wheeler	185	32	217	56
4-wheeler	351	40	391	153
total	1297	209	1506	330

Boxes per source video: V001=377, V002=136, V003=325, V004=262, V005=333, VEHICLES=73.

Collection methodology

Frames were extracted from vehicle-mounted dashcam recordings and manually annotated with bounding boxes using VoTT (Visual Object Tagging Tool). Annotation JSON was normalized from VoTT's per-project tag conventions (which varied in capitalization/spacing across annotation sessions, e.g. "2 Wheeler" / "2 wheeler" / "PERSON") into the canonical 4-class taxonomy above.

Privacy and anonymization

All published frames are anonymized. Detected regions are pixelated rather than Gaussian-blurred, since blur can be partially reversible. Labels are unaffected: pixelation does not move bounding boxes.

• People: the primary redaction signal is a general COCO-pretrained person detector run at a low threshold (favoring recall), with the head region of every detected person (top ~32% of the box, expanded upward and sideways) pixelated. Dedicated face detectors were evaluated first but perform poorly on this footage's haze, glare, and compression, so a face detector runs only as a supplementary layer on top of the person-based pass.
• License plates: an automatic plate detector runs at a deliberately low confidence threshold, but in this footage detector confidence does not cleanly separate legible plates from illegible ones. So on top of the automatic pass, every frame went through a manual audit: a detector sweep at near-zero confidence with each plate-like candidate reviewed by eye, crop-level checks of large vehicles, and an OCR sweep over vehicle crops of the output as a final net. The 80 hand-marked plate boxes this audit produced are pixelated unconditionally.
• Burned-in timestamps: the camera's timestamp overlay (exact date and time of recording) is blacked out. The blackout covers only the measured region the timestamps occupy, which clips 35 of the 1,506 annotation boxes by at most ~11% of their height; those pixels were already partly covered by the timestamp text itself.
• One advertisement poster on a truck showing a person's photo and phone number was also manually pixelated, even though it is public commercial signage.

Known limitations

• Small v0.1 sample: 293 frames from 6 videos, out of a much larger (~12GB, 203-clip) raw footage archive that is not yet annotated. The full anonymized archive is included in this repository under videos/ so others can extract and annotate frames themselves. Expect this dataset to grow in future releases.
• Daytime only: all footage in this release was recorded in daylight hours (the 6 annotated videos span 10:14 to 16:19). Night clips were recorded during the same pilot but excluded from the release: in low light the camera falls to roughly 6 fps with long exposures, so frames are dominated by motion blur and headlight bloom and contain no annotatable vehicle shapes. This release should not be assumed to generalize to low-light conditions.
• Class imbalance: pedestrian is only ~2.6% of boxes and has zero occluded examples, so expect weaker detector performance on this class until more data is added.
• Playback speed is mixed across the raw archive: the dashcam stamps a 20 fps header on every clip regardless of its true capture rate, which varies from roughly 6 fps at night to 15 fps in daylight. Most of the anonymized archive is written at a per-clip real rate measured from the burned-in timestamps (before they are blacked out), so those clips play at real-world speed; the first ~100 clips processed kept the camera's 20 fps header and play 1.3-3.3x faster than real time. Frame content is identical either way, only the playback timing differs, so this does not affect training on extracted frames.
• Anonymization is best-effort, not a guarantee: the combination of automated detection and full manual audit described above was verified by visual review, but redaction of real-world footage can never be proven exhaustive. Pixelated head and plate regions also mean a small fraction of annotated-object pixels are destroyed; boxes are kept at their original geometry, and models trained on this data will see pixelation artifacts inside some boxes. Measured effect: a YOLO26n baseline trained on the original frames scores mAP50 0.719 on the original validation images but 0.629 on this anonymized copy of them.

License

CC BY 4.0: free to use, share, and adapt, including commercially, with attribution to Visaitech.

Citation

@dataset{visaitech_vehicle_mixed_traffic_2026,
  title        = {Visaitech Mixed-Traffic Vehicle Detection Dataset},
  author       = {Visaitech},
  year         = {2026},
  version      = {0.1},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/visaitech/vehicle-mixed-traffic-detection}
}