A deep-learning classifier for object tracking from through-the-wall radar data

In this work, a novel data driven object tracking scheme from through-the-wall radar data is presented. The localization task is firstly rewritten in terms of a classification problem which is addressed by using a deep learning approach. This localization and tracking task is assisted by standard inverse problems techniques which are used in order to recover the stationary background and to enhance the performance of the learning method. Together it yields a combined classifier network which is able to estimate very efficiently an almost arbitrary trajectory described by a target moving inside a building. No prior assumption on the dynamical model of the object is needed, which is a big advantage of the proposed scheme compared to standard Bayesian tracking filters. A 2D proof-of-concept study is discussed here which prepares more realistic 3D extensions to be addressed in our future research.