Performance Comparison of Moving Target Classification based on Deep Learning

Abstract

Radar target detection is a basic but important process of radar systems, and it is difficult to distinguish and measure targets in real-world environments. Therefore, distinguishing between humans and animals based on radar signals is a difficult task in the field of ground radar. The radar signal processing section uses the in-phase/quadrature- phase (I/Q) matrix radar signal data and geolocation types as inputs and performs binary classification to classify animals and humans. In this radar signal processing, deep learning methods are adopted as feasible solutions. However, there is a limited lack of training data in the real world and a problem with jamming signals, which are adversarial attacks. However, it is difficult to collect a lot of training data in a real-time environment. Reflecting this, we learn only some data from MAFAT Radar Challenge data to compare and analyze the classification performance of conventional methods convolutional neural network (CNN) and CNN-based U-Net and U-Net with residual blocks U-Net (Res- UNet) algorithms.