Real-Time Smart Surveillance and Enforcement for Dust throw Detection and Identity Recognition Using YOLO 12 and SA - FaceXNet

• Published in: Signal, image and video processing Vol. 19; no. 12
• Main Authors: Tiruvikraman, V, Selvakumar, D, Vijayakumar, P
• Format: Journal Article
• Language: English
• Published: Heidelberg Springer Nature B.V 01.12.2025
• Subjects: Attention; Face recognition; Machine learning; Object recognition; Real time; Surveillance
• ISSN: 1863-1703; 1863-1711
• DOI: 10.1007/s11760-025-04582-x

Urban littering poses persistent environmental and civic challenges, particularly in the context of smart city management. Although deep learning has improved object and face recognition in surveillance, existing systems typically address these tasks in isolation and fail to associate specific public behaviors with individual identities. Manual monitoring remains common and ineffective for timely enforcement. This study proposes an integrated, real-time surveillance framework that detects littering behavior and identifies violators using standard CCTV feeds. A fine-tuned YOLOv12 model detects actions and triggers facial recognition only when a violation occurs. The recognition module–SA-FaceXNet–is built on a ConvNeXt-inspired backbone enhanced with Selective Attention (SANet) and Efficient Channel Attention (ECA) for robust feature extraction. Spatio-temporal tracking is employed to associate actions with detected faces. Trained on diverse, large-scale datasets, the system improves face detection accuracy by 12.5% over YOLO-based baselines and boosts recognition accuracy by 9.8% compared to ResNet variants. The system achieves 92.4% end-to-end efficiency in real-time scenarios.