Deep learning research, 2026
AnomLite
A hybrid model for multiclass crime detection in surveillance video, small enough to run on constrained hardware.
- Role
- Solo developer
- Type
- Deep learning research, 2026
- Status
- Source only
- Accuracy
- 79%
- ROC AUC
- 0.97
- Macro F1
- 0.75
- Parameters
- ~11M
- Problem
Anomaly detection on surveillance footage usually means a model too heavy to run anywhere near the camera, so the video has to travel to the compute instead.
- What I built
MobileNetV2 handles the spatial features and an LSTM handles the temporal ones, which keeps the whole network to roughly 11 million parameters. Trained on the UCF-Crime dataset across 14 classes, from abuse and arson through to normal video.
- Result
79 percent accuracy, 0.97 ROC AUC and a 0.75 macro F1, at a size that suits real-time deployment on constrained devices.
- Stack
PyTorch, MobileNetV2, LSTM