POSE TO PROTECT: FEDERATED SKELETON-BASED ANOMALY DETECTION FOR PRIVACY-CONSCIOUS VIDEO SURVEILLANCE

Author

Dina Famouri, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2025

Degree Name

Master of Science

Department

Electrical and Computer Engineering

First Advisor

Sayeh, Mohammad

Second Advisor

Imteaj, Ahmed

Abstract

In an era of increasing concern over data privacy, especially in surveillance applications, this thesis explores a privacy-preserving approach to crime detection using pose-based features. Rather than analyzing raw video footage, which often raises ethical and legal issues, our method uses OpenPose to extract 2D skeletal keypoints from surveillance videos. These pose sequences serve as an abstract yet informative representation of human activity.To classify pose sequences as either normal or criminal behavior, we employ a two-layer LSTM model capable of capturing both short and long-range temporal patterns. The study also compares two training paradigms: centralized learning, where all data is collected and trained on a single server, and federated learning, where multiple clients train locally and share only model updates. This federated setup helps preserve user privacy by keeping raw data decentralized. Experimental results show that skeletal features are highly effective for recognizing anomaly behavior, and the LSTM model performs well across both setups. Notably, the federated learning approach achieves performance comparable to the centralized model while significantly improving data privacy. This research demonstrates the viability of combining pose-based representations with federated learning for secure and effective crime detection, offering a practical solution for real-world surveillance systems operating under privacy constraints.