README.md

---
library_name: onnx
pipeline_tag: video-classification
tags:
  - onnx
  - video-classification
  - accident-detection
  - dashcam
  - jetson
  - rockchip
license: mit
---

# Dashcam Collision Detector (videomae_base, ONNX)

Causal sliding-window crash detector exported to ONNX. The model scores a short
temporal window of dashcam frames and a downstream rule (`consec` consecutive
detections above `detect_threshold`) decides *when* a collision occurs.

- **Architecture:** `videomae_base`
- **Input:** float32 `[1, 3, 16, 224, 224]` (NCTHW), RGB, ImageNet mean/std normalized
- **Sampling:** 16-frame window at 16 fps
- **Decision rule:** threshold `0.97`, `3` consecutive windows

## Inference metadata

```json
{
  "input_shape": [
    1,
    3,
    16,
    224,
    224
  ],
  "detect_threshold": 0.97,
  "consec": 3,
  "target_fps": 16,
  "window_frames": 16,
  "stride": 3,
  "tolerance_s": 1.0,
  "mean": [
    0.485,
    0.456,
    0.406
  ],
  "std": [
    0.229,
    0.224,
    0.225
  ],
  "arch": "videomae_base"
}
```

## Usage

```python
from huggingface_hub import hf_hub_download
import onnxruntime as ort, numpy as np

path = hf_hub_download(repo_id="akhra92/dashcam-collision-jetson-videomae-hnm", filename="model.onnx")
sess = ort.InferenceSession(path, providers=["CPUExecutionProvider"])
x = np.random.randn(*[1, 3, 16, 224, 224]).astype("float32")
logit = sess.run(["logit"], {"input": x})[0]
```

On-device, the ONNX graph is compiled to a TensorRT engine (Jetson) or an
`.rknn` model (Rockchip). See `deploy/` in the source repo.