YOLOv8-Based Infrared Intrusion Detection and Classification Model Using the OTCBVS Benchmark Dataset

Abstract

Infrared imaging provides a reliable technique of monitoring and detecting intrusions in low-light and harsh environmental situations. This paper presents an optimized deep learning model for infrared intrusion detection and classification utilizing the You Only Look Once version 8 (YOLOv8) framework. The model was developed using the OTCBVS Benchmark Dataset Collection, pre-processed and annotated with Roboflow, and trained in a Google Colab Pro GPU environment. Multiple hyperparameter tuning and picture augmentation procedures were used to enhance detection robustness in a variety of environmental and operational conditions.  The numerical evaluation revealed a precision of 0.999, recall of 0.929, mAP50 of 0.937, and mAP50-95 of 0.70, as against the non-optimized YOLOv8 (baseline YOLOv8) with precision of 0.740, recall of 0.650, mAP50 of o.721, and mAP50-95 of 0.453, respectively. This shows outstanding accuracy and generalization across various test conditions. The smooth convergence of the training and validation loss curves shows the model's stability and lack of overfitting. Visualizing bounding box predictions and confusion matrices proves the efficacy of intrusion target localization and classification. The results establish the proposed YOLOv8-based model as a highly efficient and precise framework for infrared-based intelligent surveillance and smart power infrastructure protection.