Rapid Object Depth Estimation From Position-Referenced EMI Data Using Machine Learning

Abstract

State-of-the-art methods for localization and detection of small metallic objects using electromagnetic induction (EMI) sensing usually struggle due to the strong correlation between the intrinsic parameters of the object and the object’s depth. In this article, we present a machine-learning-based approach for rapid estimation of metallic object depth from line-scan EMI data. The 1-D-convolutional neural network (1D-CNN), trained on a simulated dataset, takes advantage of metal detector (MD) spatial response to extract features from which depth is inferred. Experimental evaluation using a mono-coil pulse induction MD and an electromagnetic (EM) tracking system was performed under laboratory conditions on a large dataset containing arbitrarily oriented objects of different sizes, shapes, and materials. The nonlinear least-squares (NLSs) inversion was employed as the benchmark method for comparison. The proposed solution shows superior performance over NLS at depths >10 cm. From two passes of MD over the object at depths within the range of 2.5–15 cm, our method yields a median absolute error (MedAE) on the order of millimeters.