Enhancement of Range Accuracy for Closely Located Targets in FMCW Radar Sensor With CNNs
This study proposes to enhance the range accuracy of radar sensor through a deep learning approach. As object-detection needs in automotive, industrial, and medical applications evolve, radar sensors are becoming increasingly important for their cost-effectiveness and robust performance across varied environmental conditions. Despite these advantages, radar’s measurement accuracy often falls, especially when measuring closely spaced objects. Conventional frequency modulated continuous wave radar range estimation relies on detecting peaks in the range profile obtained via Fourier transform (FT). However, FT suffers from the limited bandwidth as well as the existence of sidelobes in the transformed domain. Therefore, the range profile is merged or distorted when target peaks have insufficient separation. By applying a convolutional neural network (CNN) to the radar range profile, we aim to improve range estimation accuracy when two targets are closely spaced so that peaks are merged or distorted. The proposed method shows a reduction in mean absolute error compared to traditional techniques and other models, enhancing the radar’s capability to precisely measure the range of targets. The results demonstrate that the CNN-based approach can improve sensor range accuracy by an average of 21.2%, suggesting potential applications for analyzing multiscatterer targets.