File size: 1,720 Bytes
e81efb3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
---
license: other
license_name: rtgs-internal
license_link: LICENSE
task_categories:
- object-detection
- image-classification
tags:
- cctv
- helmet-detection
- anpr
- india
- motorcycle
pretty_name: Andhra Pradesh RTGS CCTV helmet + plate datasets
size_categories:
- 10K<n<100K
---

# CCTV Datasets for helmet detection + ANPR

Training and evaluation data used by [vivekvar/helmet-v5](https://huggingface.co/vivekvar/helmet-v5) and [vivekvar/helmet-v4](https://huggingface.co/vivekvar/helmet-v4).

Source: Andhra Pradesh RTGS CCTV feeds (public road cameras). All crops and frames are from motorcycle traffic scenes.

## Folders

| Folder | Contents | Purpose |
|---|---|---|
| `merged_v3/` | YOLO-format dataset (`data.yaml` + `train/valid/test`) | Bike + rider detection training |
| `clean_merged_data/` | Cleaned / deduped crop set | Base training data for v4 |
| `extra_khadatkar/` | Additional labeled crops (source: Khadatkar junction) | Augment v4 helmet classifier |
| `extra_learning/` | High-quality hand-checked samples | Active-learning batch |
| `extra_nckh/` | Additional labeled crops (source: NCKH junction) | Augment v4 helmet classifier |

## Use

`merged_v3/data.yaml` is ready for Ultralytics YOLO:

```bash
yolo detect train data=merged_v3/data.yaml model=yolo11l.pt imgsz=1280 epochs=50
```

Crop folders use directory-per-class layout compatible with `torchvision.datasets.ImageFolder`.

## Notes

- Labels are a mix of human-verified and model-pseudo-labeled (helmet-v4 self-labels). The `extra_*` sets skew higher quality.
- Cameras used: GNT_TMS_269, GNT_TMS_273 and other Andhra Pradesh RTGS nodes.
- Not included: raw video clips — too large and proprietary. Frames are extracted only.