Nano3D: Explainable Ultra-Lightweight Deep Learning Model for Video Anomaly Detection
Published:
Project Overview
Nano3D is an ultra-compact 3D Convolutional Neural Network specifically designed for automated surveillance anomaly detection on resource-constrained edge computing platforms (such as NVIDIA Jetson and embedded IoT hardware).
While standard video anomaly detection frameworks rely on multi-million parameter networks requiring high-wattage server GPUs, Nano3D achieves state-of-the-art detection precision with an astonishingly small 1.86 MB model footprint.
Benchmark Highlights
- Parameter Count: Only 486,000 parameters (~95% fewer than standard 3D Conv baselines).
- Latency & Throughput: 7.19 ms per clip inference latency, operating at real-time speeds (>130 FPS).
- Cross-Dataset Performance: Benchmarked across UCF-Crime, XD-Violence, RWF-2000, UBI-Fights, and SCVD, delivering 88.7% accuracy and 0.947 AUC.
- Spatiotemporal Grad-CAM: Integrated visual explanations showing dynamic heatmaps where violent altercations, fights, or anomalous behaviors are localized in time and space.