Measurement methods for body fat (%) assessment from 3D Kinect scans

Abstract

Body composition is a better health indicator than just scale weighting. Previous research internally conducted at Philips showed that a model based on full 3D body representation can accurately measure body fat(%) (RMSE = 2.22%) on pregnant women. The current study investigates the plausibility of performing body fat assessment from single 3D depth-maps, thus an incomplete3D representation of human body. A Kinect v2 device was used for data acquisition and two predictive models based on hand crafted features extracted from the point cloud data have been developed. A Lasso regression lead to a 3 parameter fat prediction model with ad j 􀀀 R2 = 0.72 and RMSE = 8.02%. A multivariate linear regression with a stepwise elimination routine resulted into a predictive model with adj-R2 = 0.60, and RMSE = 9.85%.