Camouflaged Object Detection in Multispectral Imagery for Surveillance Applications Using YOLO Variants

The proposed study investigates the efficacy of advanced YOLO (You Only Look Once) variants, specifically YOLOv5, v7, v8, and v9, in detecting concealed camouflaged objects within multispectral imagery. Utilizing the multispectral camera, “Altum PT,” a comprehensive dataset of multispectral images was captured and preprocessed through various techniques, including fusion, false coloring, and pan sharpening. The performance of these YOLO models was evaluated through extensive experiments in various environmental contexts. The results demonstrate that YOLO variants, especially when combined with Slicing Aided Hyper Inference (SAHI), exhibit robust detection capabilities, with precision ranging from 93.9% to 98.2% and recall varying from 80.2% to 98.1%. It was found that although performance metrics are affected by architectural and training modalities, these models are promising for surveillance and security purposes. The proposed SAHI implementation helped overcome the problem of false alarms and thus improved the prediction results. The proposed research is an improvement over current methods used to detect concealed objects and indicates the capability of deep learning to improve surveillance. Resources: https://github.com/iamabhi1007/YOLO-Variants-Multispectral-Imagery-Enhanced-Camouflage-Object-Detection.git.