Deep Learning Frameworks for Suspicious Activity Detection

Thesis Overview

Supervisors:

• Prof. Dr. Amitabha Chakrabarty, Professor, Dept. of Computer Science & Engineering, BRAC University
• Prof. Dr. Md. Golam Rabiul Alam, Professor, Dept. of Computer Science & Engineering, BRAC University

Automated detection of violent altercations, theft, and erratic behaviors in public surveillance streams is vital for municipal safety and proactive threat mitigation. This thesis systematically investigates spatiotemporal deep learning paradigms, examining the architectural trade-offs between dense temporal modeling and edge compute constraints.

Key Contributions & Results

1. Rigorous Multi-Dataset Benchmarking: Conducted comprehensive empirical evaluation of state-of-the-art spatiotemporal models under identical evaluation metrics across 5 public datasets: UCF-Crime, XD-Violence, RWF-2000, UBI-Fights, and SCVD.
2. Nano3D Edge Architecture Formulation: Designed and trained Nano3D, reaching 88.7% overall classification accuracy, 0.947 Area Under Curve (AUC), and 7.19 ms inference latency with only 486,000 parameters (1.86 MB file footprint).
3. Spatiotemporal Attribution via Grad-CAM: Integrated gradient-weighted class activation mapping extended across temporal video slices, confirming that feature representations localize accurately onto actors and objects undergoing anomalous interaction.

View Code & Benchmarks on GitHub

Recommended citation: S. M. Mohiuddin Khan Shiam. (2026). "Deep Learning Frameworks for Suspicious Activity Detection." Academic Thesis, Department of Computer Science & Engineering, BRAC University. Supervised by Prof. Dr. Amitabha Chakrabarty & Prof. Dr. Md. Golam Rabiul Alam.
Download Paper