Machine Learning for Radar Imaging

Abstract

Radar Systems find widespread use in the technological world ranging from old existing applications in military vehicles to modern implementations in mobile phones and self-driving cars. Radars perform detection by sweeping a beam to locate the object similar to a blind person swinging their cane. But this process has hardware limitations and is slow and expensive. A Radar System makes use of an antenna array in which finding the direction-of-arrival (DoA) angle problem is of significant interest because real-time estimation of the DoA can lead to a faster and more efficient object detection by the radar. Existing methods based on subspace estimation such as MUSIC or based on rotational invariance such as ESPRIT first generate a forward relation by mapping the antenna source positions to the output signals and then try to perform the inverse mapping to find the DoA. In a practical scenario, these methods are time-consuming because they require complex calculations on finding the spatial covariance matrix and they can be inaccurate because of array imperfections which lead to faulty forward mapping. This project aims to make use of a Machine Learning (ML)- based model for DoA estimation because ML models are built on data and do not rely on the forward mapping between the array geometry and the output signal. They instead try to find the inverse mapping by making input-output pairs between the output signal vectors and the source locations by means of classification. Hence, they can be potentially faster and more accurate which is explored in this project. The project specifically focuses on the specialized form of ML, which is Deep Learning and makes use of Deep Neural Network (DNN) framework with a one-vs-all classification method to which can eliminate the shortcomings of other ML methods that do not perform well when the data has a lot of variability. The focus is also on introducing simulated array imperfections such as mutual coupling and errors in the position of sensors and observing the robustness and performance of the model in dealing with these imperfections. Simulations are performed to demonstrate that the proposed method does indeed handle the generalization of data and array imperfections problem satisfyingly and possibly providing real-time implementation performance.