Federated Learning for Robust People Detection in Decentralized Surveillance Systems
Saif Ismael, Dinah Waref, Mohammed A.‐M. Salem · 2024
In recent years, machine learning has made significant progress. Federated learning is one of the new techniques of machine learning, but the estimated costs of implementing it stand in the way of its widespread adoption. The infrastructure computation resources and the skills needed to deploy these systems are often considered expensive, especially for small and medium-sized businesses. In this paper, an automated federated learning architecture for decentralized surveillance systems is proposed, addressing the demand for efficient and accurate surveillance systems while ensuring cost-effectiveness. For this research, the YOLOv8n model for real-time people detection and Autodistill for model distillation and data labeling are used. Federated learning allows for the training of the global model through collaboration of the participating systems, which optimizes detection accuracy and system efficiency while retaining data privacy. After thorough evaluation, the proposed architecture demonstrates better results in accuracy, efficiency, and scalability in comparison to traditional systems, and these results are comparable with other recent work discussed in our background section. For instance, our system achieved an accuracy of 74.8% and a 33.9% increase in consistency on all datasets in only 20 Federated learning rounds on highly heterogeneous data. Furthermore, real-world deployment feasibility is explored, with hardware and computational restraints taken into consideration; a comprehensive analysis for the feasibility study is done.