---
library_name: onnx
pipeline_tag: video-classification
tags:
  - onnx
  - video-classification
  - accident-detection
  - dashcam
  - rockchip
  - rk3588
license: mit
---

# Dashcam Collision Detector — Rockchip RK3588 (`resnet18_temporal`)

Causal sliding-window crash detector for the **RK3588 NPU**. Because the NPU has no
3D convolutions, this is a per-frame 2D-CNN + a small temporal head, deployed as
**two ONNX graphs** (convert the backbone to INT8 RKNN; run the head on the CPU):

| file | shape | runs on |
|---|---|---|
| `backbone.onnx` | `[1, 12, 112, 112]` → `[1, 512]` | NPU (INT8 RKNN) |
| `temporal_head.onnx` | `[1, 16, 512]` → `[1]` | CPU |
| `rockchip.meta.json` | — | inference config |

- **Input:** RGB + temporal diffs at lags [1, 2, 4] (12 channels), 16-frame window
- **Decision rule:** threshold `0.95`, `3` consecutive windows

## Results (held-out videos, full-video streaming eval)

| metric | value |
|---|---|
| detection rate | 34.8 % |
| false-alarm rate | 15.0 % |
| mean localization error | 0.34 s |
| operating threshold | 0.95 |

Window-level ranking on validation windows (incl. full-timeline negatives): AUC 0.894, AP 0.528.

## Usage

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

repo = "akhra92/dashcam-collision-rockchip-resnet18-motion"
bb   = ort.InferenceSession(hf_hub_download(repo, "backbone.onnx"))
head = ort.InferenceSession(hf_hub_download(repo, "temporal_head.onnx"))
meta = json.load(open(hf_hub_download(repo, "rockchip.meta.json")))

T, C = meta["window_frames"], meta["feat_dim"]
feats = np.zeros((1, T, C), np.float32)               # fill from per-frame backbone
frame = np.random.randn(*meta["frame_shape"]).astype("float32")
feats[0, -1] = bb.run(["feat"], {"frame": frame})[0][0]
logit = head.run(["logit"], {"feats": feats})[0]
```

See `deploy/rockchip/` (convert_rknn.py, infer_rknn.py) in the source repo for the
INT8 conversion and streaming inference.