Cyber–Physical Surveillance with Adaptive Graph Neural Networks for Person Re-Identification

Abstract

Person re-identification (Re-ID) is a vital component of cyber–physical surveillance systems, enabling precise and real-time tracking of individuals across multiple camera viewpoints. However, conventional deep learning models often face limitations in handling occlusions, varying viewpoints, and dynamic environmental changes. To overcome these challenges, we propose a novel Adaptive Graph Neural Network (AGNN) framework designed for robust and scalable person re-identification within cyber–physical systems (CPS). The AGNN constructs a spatiotemporal graph representation that effectively captures both appearance-based and structural relationships of individuals across different camera feeds. To enhance feature discriminability, we introduce a Dynamic Feature Aggregation (DFA) mechanism that adapts to varying environmental conditions. Furthermore, an Edge-Weighted Attention Module (EWAM) is incorporated to emphasize crucial relational dependencies, improving the model's resilience to noise and ambiguity. Experimental results on benchmark Re-ID datasets show that our AGNN framework surpasses existing state-of-the-art methods in accuracy, robustness, and real-time efficiency. This makes it a promising solution for intelligent surveillance, smart city monitoring, and other security-critical CPS applications.