---
license: cc-by-4.0
tags:
- autonomous-vehicles
- danger-detection
- driving
- safety
task_categories:
- object-detection
language:
- en
---

# PRECOG-Labels: PhysicalAI-AV Danger Labels

**Author:** Nikhil Upadhyay | MSc Business Analytics | Dublin Business School
**Project:** [PRECOG-AV](https://github.com/TrazeMaG/PRECOG-AV)

## Overview

Physics-based danger labels for **298,326 clips** from the
[NVIDIA PhysicalAI-AV](https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles) dataset,
mined using Time-to-Collision (TTC) analysis of 3D obstacle parquets.
The first danger labelling at this geographic scale: **25 countries**.

**14,192 danger clips identified (4.76% positive rate)**

## Files

| File | Rows | Description |
|------|------|-------------|
| `danger_labels.csv` | 298,326 | Binary danger flag and min TTC per clip |
| `precog_dataset.csv` | 298,326 | Full metadata: country, split, hour, sensor availability |

## Geographic Split

| Split | Countries | Clips |
|-------|-----------|-------|
| Train | 20 countries | 275,573 |
| Val | Austria, Finland, Portugal | 16,151 |
| Test | Greece, Bulgaria | 6,602 |

## Key Findings

- Danger rates vary **4.7x across countries** — Italy 9.9% vs Estonia 2.1%
- Peak danger window: **13:00 to 15:00** across all countries
- Mean danger window: **11.18 seconds** across 2,177 labelled test clips

## Citation

```bibtex
@misc{upadhyay2026precog,
  title  = {PRECOG: Proactive Risk and Environmental Cognition for Autonomous Vehicles},
  author = {Upadhyay, Nikhil},
  year   = {2026},
  url    = {https://github.com/TrazeMaG/PRECOG-AV}
}
```