LRCN-Enabled UAV Surveillance System for Suspicious Human Activity Recognition in Smart Cities

Public safety and security remain critical concerns in urban environments. Detecting suspicious activities in densely populated areas poses significant challenges for modern smart cities due to occlusions, limited fixed camera coverage, and the dynamic nature of large crowds. To address this problem, this paper proposes an Artificial Intelligence (AI)-driven unmanned aerial surveillance framework for proactive monitoring and abnormal activity recognition. The system leverages a Long-term Recurrent Convolutional Network (LRCN)-enabled architecture capable of extracting spatiotemporal patterns from aerial video streams, allowing it to detect suspicious behavior with high precision. Three deep learning models are comparatively evaluated: (i) Visual Geometry Group 16 (VGG 16) with Long Short-Term Memory (LSTM) network, (ii) a Motion Influence Map (MIM)-based approach, and (iii) LRCN. Among them, the LRCN model demonstrated superior performance, achieving an accuracy of up to 88% and outperforming the comparative methods in F1-score, Precision, Area Under the Curve (ROC-AUC), and Matthews Correlation Coefficient (MCC). Unlike traditional ground-based CCTV systems, the proposed Unmanned Aerial Vehicle (UAV)-based framework provides wider field-of-view, higher scalability, and improved visibility in dense crowds. Extensive experiments were performed to validate the practicality and robustness of the approach. The empirical findings confirm that the LRCN model effectively identifies and categorizes suspicious activities in real-world, densely populated smart-city environments. Overall, this study presents a novel and scalable aerial surveillance solution that enhances situational awareness, strengthens public-safety infrastructure, and contributes to the development of safer smart cities.