akhra92/dashcam-collision-rockchip-resnet18-motion

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

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.