PhD Studentship in Non-linear Aspects of Through-Wall Radar Imaging and Object Characterisation and Tracking

Applications are invited for a 3-year fully-funded PhD studentship, available September 2017.

The work will develop imaging methods for the detection, identification and tracking of obscured targets inside buildings. This includes through-wall observation of unseen activity for hostage situations, disaster relief, and detection of illegal activity. Low-frequency radar offers one modality which can be applied to both the through-wall and ground-penetrating cases. However, applying an imaging method based on a linearization of the inverse scattering problem (such as synthetic aperture radar) may result in poor or uninterpretable reconstructions. The data may be highly non-linear due to multiple scattering, and so solving the linearized inverse problem can create severe artefacts, or misplace and distort objects.

This Studentship will support DSTL’s ongoing research and development activities in multi-static low-frequency radar for through-wall imaging, by developing algorithms and techniques to solve this non-linear inverse problem. Such algorithms will effectively incorporate a-priori knowledge, for example of the building structure and interior, as well as the scattering phenomena observed in through-wall radar.
This project concentrates on the identification, classification and tracking of objects in the above described scenarios. The main assumption will be that data are obtained over a time range in which objects are potentially moving through the scene. Mathematically this will amount to solving a time-dependent inverse problem, a task being closely related to data assimilation problems often found in weather prediction, reservoir characterization, or even robotics. We will apply variants of the Kalman filter approach, in particular the so-called Ensemble Kalman Filter, for predicting and analysing movement in the scene, based on data obtained in each time step from solving an inverse problem of Maxwell’s equations in frequency domain. For the latter one we will start out using iterative full nonlinear inversion techniques, but possibly truncated at a low iteration number based on given time constraints for the inversion. In addition, novel tailor made regularization techniques will be applied to this problem making use of either level set techniques for shape characterization or sparsity approaches for efficient representation of the scenes, which will speed up the tracking task. This approach will enable us to understand the impact of various factors of the inversion process on speed, accuracy and uncertainty of the obtained information.