A YOLO26n fine-tune that detects pedestrian / 2-wheeler / 3-wheeler /
4-wheeler in South Asian mixed traffic, trained on the
Visaitech Mixed-Traffic Vehicle Detection dataset (v0.1).
The class taxonomy covers what general-purpose COCO detectors don't: an
auto-rickshaw class, and a single 2-wheeler class for how motorcycles and
bicycles actually mix in this traffic.
This is an early v0.1 baseline trained on 239 images: a working proof
of concept, not a production-grade detector. Expect it to improve as the
dataset grows.
The model wired into a simple proximity risk overlay: safe objects get a
thin green outline, close objects are tagged orange (caution) or red
(critical). The footage is a clip from the raw archive that ships with
the training dataset. A higher-quality video of the same run is in this
repo as demo.mp4, and the overlay code plus a runnable Gradio app live
at github.com/visaitech/mixed-traffic-risk-detector.
Validation split: 54 images / 209 boxes, held out by whole source video
(not by frame) so near-duplicate frames never straddle the train/val
boundary.
The model was trained and evaluated on the original frames. The published
dataset is the anonymized copy (pixelated heads and license plates,
blacked-out timestamp region), which changes object appearance enough to
lower the numbers, so both are reported. If you re-run validation against
the public dataset, expect the second table.
On original (non-public) frames
class
instances
precision
recall
mAP50
mAP50-95
pedestrian
9
0.688
0.333
0.383
0.250
2-wheeler
128
0.653
0.875
0.796
0.502
3-wheeler
32
0.494
0.854
0.764
0.560
4-wheeler
40
0.915
0.875
0.934
0.704
all
209
0.688
0.734
0.719
0.504
On the published anonymized frames
class
instances
precision
recall
mAP50
mAP50-95
pedestrian
9
0.656
0.333
0.251
0.166
2-wheeler
128
0.620
0.750
0.695
0.332
3-wheeler
32
0.538
0.719
0.708
0.509
4-wheeler
40
0.485
0.875
0.862
0.648
all
209
0.575
0.669
0.629
0.414
Training details
Fine-tuned from the official yolo26n.pt COCO checkpoint with
Ultralytics defaults (auto optimizer, lr0 0.01, seed 0), image size 640,
batch 8.
Early-stopped at epoch 105 of a 150-epoch budget (patience 30); best
checkpoint from epoch 75.
Trained on CPU (laptop-class i7-10510U, ~94s/epoch, under 3 hours
total): the dataset is small enough that this was more stable than the
available 2GB-VRAM laptop GPU.
Limitations
Weak on pedestrians: recall is only 0.333. The v0.1 training set has
just 39 pedestrian boxes (~2.6% of all boxes), so this class needs more
data, not more training. Do not rely on this model to find people.
Daytime only: all training footage is daylight (10:14 to 16:19).
Performance on night footage is untested and likely poor.
Single camera setup: all frames come from the same dashcam hardware
and region. Expect domain shift on other cameras, mounting positions, or
countries.
Not a safety component: this is a portfolio/research baseline. Do
not use it for collision avoidance, driver assistance, or any
safety-critical decision.
License
AGPL-3.0. The model is a fine-tune of Ultralytics YOLO26n weights and is
used via the AGPL-3.0-licensed ultralytics package, so the weights are
released under the same license. The training dataset is separately
licensed CC BY 4.0.
Citation
@misc{visaitech_mixed_traffic_yolo26n_2026,
title = {Mixed-Traffic Vehicle Detector, YOLO26n},
author = {Visaitech},
year = {2026},
version = {0.1},
publisher = {Hugging Face},
url = {https://huggingface.co/visaitech/vehicle-mixed-traffic-yolo26n}
}
