Object Depth Estimation From Line-Scan EMI Data Using Machine Learning

Abstract

In this paper, we present a novel approach to metallic object depth estimation using a pulse induction metal detector in combination with an electromagnetic tracking system. A dipole approximation model is used for modeling the spatial response of the metal detector, while 1D-convolutional neural network is employed for depth estimation. The proposed algorithm is experimentally validated in laboratory conditions. Given a single horizontal pass over a metallic object placed within the range (−10.5, −2.5) cm and (−1,1) cm for the $\boldsymbol{z}$ and $\{\boldsymbol{x},\boldsymbol{y}\}$ coordinates, respectively, the algorithm estimates the depth of the object regardless of its shape, size, and material properties with a mean absolute error $< \mathbf{4}.\mathbf{5}\ \mathbf{mm}$.