A YOLOv8n model fine-tuned on Bangladeshi roadside traffic footage to detect and classify 9 vehicle and pedestrian types common in South Asian urban environments. Designed for real-time inference — 3.4 ms per image on GPU, ~15 FPS end-to-end in the streaming pipeline.

Evaluation Results

Evaluated on 4,736 validation images (52,850 instances).

Class	Images	Instances	P	R	mAP@50	mAP@50-95
All	4736	52850	0.701	0.713	0.762	0.528
Bike	2135	3173	0.774	0.781	0.851	0.510
Bus	3895	10408	0.787	0.866	0.900	0.657
Car	4281	12627	0.824	0.900	0.934	0.699
Cng	3126	5509	0.823	0.845	0.911	0.652
People	2874	7647	0.799	0.764	0.863	0.546
Rickshaw	3849	11964	0.804	0.850	0.904	0.623
Truck	218	243	0.430	0.494	0.474	0.375
Mini-Truck	979	1136	0.568	0.681	0.690	0.504
Cycle	139	143	0.497	0.231	0.330	0.183

Note: Truck, Mini-Truck, and Cycle have significantly fewer training instances (143–1136) compared to other classes (3000–12000+), which explains their lower recall and mAP. Performance on these classes will improve with more annotated data.

Inference speed (per image): 0.1ms preprocess · 3.4ms inference · 0.4ms postprocess

Model Details

Property	Value
Base architecture	YOLOv8n
Layers (fused)	73
Parameters	3,007,403
GFLOPs	8.1
Input size	480 × 480 px
Precision	FP32 (CPU) / FP16 (CUDA)
Tracker	ByteTrack
Framework	Ultralytics 8.x

Classes

ID	Class	Description
0	Bike	Motorcycle / motorbike
1	Bus	Full-size passenger bus
2	Car	Passenger car / sedan / SUV
3	Cng	CNG auto-rickshaw (3-wheel, compressed natural gas)
4	People	Pedestrian
5	Rickshaw	Human-powered cycle rickshaw
6	Truck	Goods truck / lorry
7	Mini-Truck	Small covered van / pickup
8	Cycle	Bicycle

Intended Use

Real-time traffic monitoring from fixed roadside cameras
Vehicle counting and classification
Speed estimation with multi-object tracking (ByteTrack)
Traffic density and flow analysis dashboards

This model is optimised for Bangladeshi and similar South Asian traffic environments where CNGs, rickshaws, and cycle rickshaws are prevalent — classes typically absent from Western traffic datasets.

Limitations

Trained on Bangladeshi urban roads; performance may degrade on highways or in other countries with different vehicle types
Truck, Mini-Truck, and Cycle detection is weaker due to limited training data (see evaluation table)
Speed estimates rely on a fixed pixel-to-metre heuristic (PX_TO_METER = 0.05) calibrated for a typical roadside camera height — recalibrate for different mounting heights
Low-light and heavily occluded scenes will reduce detection confidence

How to Use

Inference only

python

from ultralytics import YOLO

model = YOLO("best.pt")
results = model("traffic.mp4", imgsz=480, conf=0.3)

With ByteTrack (counting + speed)

python

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.track(
    source="traffic.mp4",
    tracker="bytetrack.yaml",
    imgsz=480,
    conf=0.3,
    iou=0.5,
    persist=True,
    stream=True,
)

for result in results:
    print(result.boxes)

Full web dashboard

The model powers a real-time FastAPI + WebSocket dashboard with live class filtering, speed estimation, and Chart.js visualisations.

→ abrarCSE29/traffic-detection-yolo

Training

Base weights: yolov8n.pt (ImageNet pre-trained)
Dataset: Bangladeshi Traffic Flow Dataset — Islam, Mohammad Manzurul; Rashid, Mohammad Rifat Ahmmad (2024), Mendeley Data, V2, doi:10.17632/h8bfgtdp2r.2
Validation set: 4,736 images · 52,850 instances across 9 classes
Framework: Ultralytics YOLOv8

Citation

If you use this model or dataset, please cite:

bibtex

@misc{abrar_hameem_2026,
	author       = { Abrar Hameem },
	title        = { traffice-detection-best (Revision 0567f6e) },
	year         = 2026,
	url          = { https://huggingface.co/abrarhameem398/traffice-detection-best },
	doi          = { 10.57967/hf/8409 },
	publisher    = { Hugging Face }
}

@misc{bangladeshi-traffic-dataset,
  author    = {Islam, Mohammad Manzurul; Rashid, Mohammad Rifat Ahmmad },
  title     = {Bangladeshi Traffic Flow Dataset},
  year      = {2024},
  publisher = {Mendeley Data},
  doi       = {10.17632/h8bfgtdp2r.2},
  url       = {https://data.mendeley.com/datasets/h8bfgtdp2r/2}
}