<![CDATA[Body Mass Index (BMI) is a widely used indicator for evaluating body proportions and individual health status. Traditional measurement methods require direct physical contact and are time-consuming. Previous studies have utilized image-based BMI measurement using controlled backgrounds. However, this limitation restricts their application in real-world settings. Moreover, no studies have combined background-independent segmentation with a camera-parameter-based conversion from pixels to physical units, leaving accurate, background-agnostic BMI estimation an open problem. This research addresses this gap using digital image processing on front-view and side-view human body images. The proposed methods include body segmentation using the U²-Net deep learning model, height estimation using pinhole camera projection approach that accounts for camera parameters (focal length, sensor size, camera distance), and weight estimation based on body volume modeling using an elliptical tube slice geometry. The body is treated as a stack of elliptical cross-sections whose width and depth are read from the front- and side-view silhouettes and summed into a total volume. Calibration and correction factors were determined through numerical optimization to obtain optimal global values. The dataset consists of 120 body images collected from 20 subjects. Experimental results show that the system achieves an accuracy of 96.2% (MAE 6.0 cm, RMSE 7.6 cm, MAPE 3.8%) for height estimation and 82.3% (MAE 9.7 kg, RMSE 13.2 kg, MAPE 17.7%) for weight estimation. BMI category classification accuracy reaches 60%. The use of U²-Net enables body segmentation without requiring a controlled background, making the system more flexible and applicable for real-world environments.]]>
Copyrights © 2026
