Towards standoff concealed object detection using false-color near-infrared imaging and deep learning: A preliminary study on detecting the presence of metal and PVC pipes under clothing

Abstract

Detecting concealed hazardous objects beneath clothing is crucial for security; however, existing sensing modalities may pose safety, privacy, or operational limitations. Millimeter wave (MMW) and terahertz (THz) imaging systems are effective at close range, but suffer from reduced spatial resolutions at longer standoff distances. Thermal infrared methods achieve higher spatial resolutions than MMW and THz imaging at longer standoff distances but struggle to detect objects if their temperatures closely match the human body or surrounding environment. In contrast, near-infrared (NIR, 780 to 1400 nm) does not require temperature contrasts and can achieve higher spatial resolutions than thermal imaging systems. We propose a concealed object detection approach using false-color images generated from hyperspectral NIR data with wavelengths between 1090 and 1210 nm in combination with fine-tuned YOLO deep learning models for classification. We evaluate our approach on images of human subjects concealing metal and PVC pipes underneath clothing taken at a distance of 7 m, simulating a concealed person-borne improvised explosive device (PBIED). 2880 trials are conducted to examine combinations of different image processing methods, model variants and sizes, and training hyperparameters. The most efficient configurations are identified through post hoc statistical analyses on test set results using the area under the receiver operating characteristic curve (AUC) metric. The most effective image processing methodology yields a median AUC of 0.9301; YOLO11 models employing the proposed processing pipeline yield a median of 0.9427, of which YOLO11 XL models have a median of 0.9535 and a maximum of 0.9856.