Nano3D: An Explainable Ultra-Lightweight Deep Learning Model for Surveillance Video Anomaly Detection

Abstract & Contributions

Real-time surveillance anomaly detection on resource-constrained edge hardware is traditionally bottlenecked by massive spatiotemporal neural networks that require heavy GPUs. Nano3D solves this dilemma by presenting an ultra-lightweight, explainable 3D Convolutional Neural Network specifically engineered for edge inference without sacrificing detection sensitivity.

• Extreme Parameter Efficiency: Formulated with only 486K parameters and a compact 1.86 MB disk footprint—orders of magnitude lighter than standard 3D ResNet, C3D, and I3D architectures.
• Ultra-Fast Edge Latency: Achieves 7.19 ms per-clip inference latency, easily enabling real-time processing (>130 FPS) on embedded edge accelerators.
• Benchmarked Across 5 Public Datasets: Evaluated under rigorous and uniform testing conditions across UCF-Crime, XD-Violence, RWF-2000, UBI-Fights, and SCVD, achieving 88.7% accuracy and 0.947 AUC.
• Spatiotemporal Visual Interpretability: Integrated spatiotemporal Grad-CAM saliency mapping to generate bounding region activations, attributing anomalous behavior to specific moving entities and human interactions across time frames.

Code & Artifacts

The full implementation, pre-trained weights, and evaluation pipelines are openly available:

Explore Nano3D GitHub Repository

Recommended citation: S. M. Mohiuddin Khan Shiam, et al. (2026). "Nano3D: An Explainable Ultra-Lightweight Deep Learning Model for Surveillance Video Anomaly Detection." Under review in Scientific Reports, Nature Portfolio.
Download Paper