---
license: agpl-3.0
tags:
  - image-classification
  - driver-monitoring
  - traffic-safety
  - phone-use
  - distracted-driver
  - yolov11
  - ultralytics
  - onnx
pipeline_tag: image-classification
library_name: ultralytics
---

# YOLOv11s-cls · driver phone-use binary classifier

Binary classifier that predicts whether a **driver is holding / using a phone**
(talking or texting, left or right hand) from a cropped driver RGB image.

Part of the **ktk-studio** traffic-violation analytics stack (DeepStream 9.0 +
Triton + B200).

## Summary

| | |
|---|---|
| Architecture | YOLOv11s-cls (Ultralytics) |
| Input | 224×224 RGB |
| Output | logits over 2 classes: `no_phone`, `phone` |
| Parameters | 5.4 M |
| GFLOPs | 12.0 |
| Weights | `best.pt` (PyTorch, 11 MB) / `best.onnx` (21 MB, opset 19) |
| Val top1 | **99.91 %** at epoch 21 (30 epochs total) |

## Training data

Source: [gymprathap/Driver-Distracted-Dataset](https://huggingface.co/datasets/gymprathap/Driver-Distracted-Dataset) (State-Farm-style 10-class).

Original 10 classes were collapsed to binary:
- `phone` = `c1` (texting right) ∪ `c2` (talking right) ∪ `c3` (texting left) ∪ `c4` (talking left)
- `no_phone` = `c0` (safe) ∪ `c5` (radio) ∪ `c6` (drinking) ∪ `c7` (reaching) ∪ `c8` (hair/makeup) ∪ `c9` (passenger)

| Split | `phone` | `no_phone` | Total |
|---|---|---|---|
| train | 7 870 | 11 196 | 19 066 |
| val (15 % holdout) | 1 386 | 1 972 | 3 358 |

## Usage

### Ultralytics

```python
from ultralytics import YOLO
model = YOLO("best.pt")
r = model("driver_crop.jpg")
print(r[0].probs.top1, r[0].names[r[0].probs.top1])
```

### ONNX Runtime

```python
import cv2, numpy as np, onnxruntime as ort
sess = ort.InferenceSession("best.onnx", providers=["CUDAExecutionProvider"])
img = cv2.cvtColor(cv2.imread("driver_crop.jpg"), cv2.COLOR_BGR2RGB)
img = cv2.resize(img, (224, 224)).astype(np.float32) / 255.0
x = np.ascontiguousarray(img.transpose(2, 0, 1)[None])
logits = sess.run(None, {"images": x})[0][0]
print(["no_phone", "phone"][int(logits.argmax())], float(logits.max()))
```

## Intended use

- Real-time phone-use violation flagging on road-traffic video after car
  detection + tracking.
- Run on the driver region of a detected car bbox (typically top-left half for
  right-hand-drive / top-right half for left-hand-drive, or the full driver
  crop from an in-cabin camera).

## Out-of-scope / limitations

- Trained on a single-source in-cabin dataset; domain gap on windshield-through
  views is not measured. Fine-tune on target footage for production use.
- Binary — does not distinguish "talking" vs "texting"; collapses holding+using
  into a single `phone` class.
- Confuses with other close-to-ear gestures (adjusting hair, scratching face)
  rarely — not tested at adversarial scale.

## License

AGPL-3.0 (inherits Ultralytics YOLOv11 weight license).