YOLOv9s-based surveillance of human and animal activities near optical fiber infrastructure

This study presents an intelligent deep learning-based approach for anomaly detection in fiber optic infrastructure using the YOLOv9s object detection model. The objective is to improve surveillance efficiency by accurately detecting humans and animals interacting with pole-mounted fiber optic structures in real time. A custom dataset was developed by capturing and annotating real-world video footage from diverse locations and environmental conditions. High-resolution images were extracted and labeled using CVAT, ensuring high-quality annotations across object types and activities. Model training was conducted in PyTorch, incorporating data augmentation techniques such as brightness adjustment, flipping, noise injection, and geometric transformations to enhance robustness and adaptability. YOLOv9s was evaluated against YOLOv8s and YOLOv5s using standard metrics. It achieved a precision of 86.2%, recall of 88.7%, [email protected] of 90.4%, and [email protected]:0.95 of 75.1%, while maintaining a low inference time of 59.3 milliseconds. These results demonstrate YOLOv9s’s superior balance between accuracy and computational efficiency. The system supports real-time anomaly detection, making it highly suitable for smart city surveillance and telecom infrastructure monitoring. This work contributes a reliable, scalable solution for enhancing the automation and security of networked environments.