---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: objects
    struct:
    - name: bbox
      list:
        list: float32
        length: 4
    - name: category
      list:
        class_label:
          names:
            '0': accident
            '1': non_accident
  splits:
  - name: train
    num_bytes: 766998045.75
    num_examples: 2122
  - name: validation
    num_bytes: 126687458
    num_examples: 316
  - name: test
    num_bytes: 120640133
    num_examples: 325
  download_size: 1014077565
  dataset_size: 1014325636.75
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
license: cc0-1.0
task_categories:
- object-detection
tags:
- object-detection
- computer-vision
- cctv
- traffic
- safety
- surveillance
- accident-detection
- low-resolution
pretty_name: cctv-traffic-accident-object-detection-dataset
size_categories:
- 1K<n<10K
---

# CCTV Traffic Accident Object Detection Dataset

## Overview
This dataset contains low-resolution CCTV traffic surveillance images annotated with bounding boxes for **traffic accident localization**.  
Each image includes regions corresponding to:
- **Accident**: vehicles or regions directly involved in a collision or traffic incident
- **Non-accident**: visible vehicles or regions not involved in the incident

The dataset is designed to support **fine-grained object detection under real-world surveillance constraints**, such as compression artifacts, motion blur, and fixed camera viewpoints.

---

## Dataset Composition

- **Total images**: 1,485 unique CCTV frames  
- **Expanded images**: ~3,000 images after offline augmentation  
- **Resolution**: 640 × 640 (letterboxed)  
- **Annotation format**: COCO-style bounding boxes  
- **Classes**:
  - `accident`
  - `non_accident`

Each image may contain **multiple objects**, including both accident and non-accident regions, reflecting realistic traffic scenes.

---

## Annotation Logic

Bounding boxes are defined as follows:

- **accident**  
  Regions corresponding to vehicles or areas directly involved in a traffic collision or incident.

- **non_accident**  
  Vehicles or regions present in the scene but **not involved** in the accident event.

This design enables **context-aware accident localization**, where the model must distinguish involved and uninvolved objects within the same frame.

---

## Data Source

The original dataset was published on **Roboflow Universe** as the  
**“Accident and Non-accident label Image Dataset”**.

- Original source:  
  https://universe.roboflow.com/accident-and-nonaccident/accident-and-non-accident-label-image-dataset

This Hugging Face version provides:
- cleaned category structure
- standardized splits (train / validation / test)
- consistent class indexing
- ready-to-use object detection schema

---

## Preprocessing & Augmentation

The following preprocessing and augmentation were applied **offline** during dataset creation:

- Resize to 640×640 with padding
- Horizontal flip (50% probability)
- Random brightness adjustment (±20%)

⚠️ **No additional augmentations are applied at training time by default.**  
Users are encouraged to account for existing augmentations when designing training pipelines.

---

## Intended Use

This dataset is suitable for:

- Traffic accident localization
- CCTV-based surveillance analysis
- Safety and incident monitoring systems
- Research on object detection under low-quality visual conditions
- Benchmarking transformer-based detectors (e.g., DETR-style models)

---

## Limitations

Users should be aware of the following limitations:

- Small-to-medium scale dataset
- Fixed camera viewpoints
- Low-resolution and compressed CCTV imagery
- Limited geographic and environmental diversity

As such, models trained on this dataset should not be assumed to generalize broadly without further validation.

---

## License

This dataset is released under the  
**Creative Commons Zero v1.0 Universal (CC0-1.0)** license.

The original dataset was designated as **Public Domain** by its authors.

---

## Citation

If you use this dataset in academic research, please cite the original dataset:

```bibtex
@misc{
  accident-and-non-accident-label-image-dataset_dataset,
  title = { Accident and Non-accident label Image Dataset Dataset },
  type = { Open Source Dataset },
  author = { Accident and Nonaccident },
  howpublished = { \url{ https://universe.roboflow.com/accident-and-nonaccident/accident-and-non-accident-label-image-dataset } },
  url = { https://universe.roboflow.com/accident-and-nonaccident/accident-and-non-accident-label-image-dataset },
  journal = { Roboflow Universe },
  publisher = { Roboflow },
  year = { 2023 },
  month = { dec }
}