beaunix/car-crash-detector

Aegis-Traffic-Sentinel

Car crash detection over short traffic video clips using MobileNetV2 combined with a temporal attention mechanism (Attention MLP) and a binary classifier. This Space loads the trained checkpoint from beaunix/car-crash-detector and runs it on ZeroGPU.

How it works

1. The uploaded video (max 45 seconds) is opened and 16 frames are sampled uniformly across its full duration (not a sliding window).
2. Each frame is resized to 224x224 and normalized with ImageNet statistics, matching the training pipeline exactly (no letterbox padding here).
3. All 16 frames are processed in a single GPU forward pass: MobileNetV2 extracts per-frame features, temporal attention pools them, and the MLP classifier (with an internal sigmoid) outputs a single crash probability for the clip.
4. The report shows the crash probability as a HUD gauge and a metrics summary table.

Model performance (test set)

Metric	Value
Accuracy	0.9829
F1	0.9828
Precision	0.9661
Recall	1.0000
AUC-ROC	0.9950
Threshold	0.50

Notes

• Videos longer than 45 seconds are rejected to protect ZeroGPU quota.
• The model's classifier head already includes a sigmoid activation; the Space uses its output directly as a probability.

License

CC BY-NC-ND 4.0 (Attribution - NonCommercial - NoDerivatives). See https://creativecommons.org/licenses/by-nc-nd/4.0/