TriPath3DNet: an efficient real-time model for multi-class classification in real-life surveillance videos of fixed duration.

This paper presents TriPath3DNet, a novel, efficient, and interpretable 3D CNN architecture designed for real-time, multi-class classification of short, motion-triggered surveillance video clips under challenging real-world conditions-including occlusion, variable lighting, adverse weather, and subtle low-motion anomalies such as loitering or zone intrusion. TriPath3DNet integrates three complementary temporal pathways-short-term motion, long-term context, and Temporal Difference Encoding (TDE)-into a lightweight ResNet3D (R3D)-18 backbone to jointly model transient dynamics and sustained activities. Evaluated on four datasets-including the newly curated Virat1-RC, Virat2-RC, UCF-Crime, and our proprietary In-House Dataset (IHD)-TriPath3DNet achieves state-of-the-art or near state-of-the-art performance, with up to 95.37% accuracy, 99.42% AUC, and an inference latency of 129 to 137 ms per 50-frame clip (approx 2.6 ms per frame) on an 11 GB GPU, using only 33.46 M parameters. Notably, it outperforms both CNN- and transformer-based baselines-including MViTv1, MViTv2, and VideoSwin-by significant margins, especially on anomaly-dense benchmarks like UCF-Crime, where most vision transformers struggle. While MViTv2 achieves slightly higher accuracy on IHD (91.87% vs. 89.46%), TriPath3DNet delivers substantially better AUC (98.07% vs. 90.96%), indicating superior calibration for critical anomaly detection. Ablation studies confirm that each temporal branch contributes meaningfully to performance, and Grad-CAM visualizations demonstrate spatially precise and temporally coherent attention maps. By aligning architectural design with edge-cloud deployment constraints and the operational realities of industrial surveillance, our work bridges the gap between academic research and real-world video analytics.