SuhxsReddy/cati-singapore-dataset

CATI Singapore Expressway Traffic Dataset

Real-time vehicle detection data collected from Singapore's 90 LTA traffic cameras using CATI (Context-Aware Traffic Intelligence) — a novel FiLM-conditioned YOLOv11 detector that adapts to environmental conditions in real time.

Dataset Description

This dataset contains per-camera vehicle detection results collected continuously from Singapore's Land Transport Authority (LTA) expressway camera network. Each record captures a full detection sweep of a single camera including vehicle counts, class breakdown, directional split, and environmental context.

Coverage

Expressway	Cameras
CTE (Central Expressway)	~16
PIE (Pan-Island Expressway)	~20
AYE (Ayer Rajah Expressway)	~8
ECP (East Coast Parkway)	~10
MCE (Marina Coastal Expressway)	~4
TPE (Tampines Expressway)	~8
BKE (Bukit Timah Expressway)	~6
KJE (Kranji Expressway)	~6
SLE (Seletar Expressway)	~6

Data Schema

Column	Type	Description
timestamp	string	Detection timestamp (SGT, UTC+8)
camera_id	int	LTA camera ID
road	string	Expressway code (CTE, PIE, AYE, etc.)
lat	float	Camera latitude (WGS84)
lon	float	Camera longitude (WGS84)
weather	string	Central Singapore weather at time of capture
total_vehicles	int	Total vehicles detected
dir_a	int	Vehicles moving in direction A (tracked via 2-frame IoU)
dir_b	int	Vehicles moving in direction B
car	int	Cars detected
motorcycle	int	Motorcycles detected
bus	int	Buses detected
truck	int	Trucks detected
van	int	Vans detected
lorry	int	Lorries detected
conf_threshold	float	Detection confidence threshold used
iou_threshold	float	NMS IoU threshold used
imgsz	int	Inference image size (px)
model_version	string	Model checkpoint identifier
dir_a_label	string	Human-readable direction A label (e.g. "towards Woodlands")
dir_b_label	string	Human-readable direction B label (e.g. "towards City")
is_ramp	bool	True if camera is likely on an entry/exit ramp (proximity heuristic)

Versioning note: Filter by conf_threshold, iou_threshold, and imgsz to isolate consistent collection windows. Exclude is_ramp=True rows for mainline-only throughput analysis.

Model

Detections are produced by CATI — a novel architecture that injects FiLM (Feature-wise Linear Modulation) layers into YOLOv11s, conditioning the backbone on real-time environmental metadata:

• Weather condition and temperature (data.gov.sg)
• Time of day (cyclical encoding)
• Camera GPS position (sinusoidal positional encoding)
• PM2.5 air quality
• Camera resolution class

Model weights: SuhxsReddy/cati-singapore

How to Load

from datasets import load_dataset
import pandas as pd

ds = load_dataset("SuhxsReddy/cati-singapore-dataset")
df = ds["train"].to_pandas()

# Vehicles by road
print(df.groupby("road")["total_vehicles"].mean())

# Directional flow on CTE
cte = df[df["road"] == "CTE"]
print(cte[["timestamp", "camera_id", "dir_a", "dir_b", "total_vehicles"]])

Collection

• Sweep interval: ~90 seconds (full network scan)
• Detection threshold: confidence ≥ 0.08, IoU ≤ 0.25
• Image source: LTA Datamall Traffic Images API
• Infrastructure: HuggingFace Spaces (CPU), continuous 24/7 collection

License

MIT — data sourced from Singapore's open data APIs under the Singapore Open Data Licence.

Citation

@dataset{cati_singapore_2026,
  author    = {SuhxsReddy},
  title     = {CATI Singapore Expressway Traffic Dataset},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/SuhxsReddy/cati-singapore-dataset}
}