Characteristics Analysis of RODNet ConfMap for MMW Radar Target Detection

Unlike the Constant False-Alarm Rate (CFAR) based MMW radar target detection methods, RODNet ( A Real-Time Radar Object Detection Network ) is based on Convolutional Neural Networks ( CNNs ), and directly learns the radar target scattering signatures from the original range-azimuth ( RA ) radio frequency image sequence. Although this is a big advantage to keep more useful information, the generated confidence map (ConfMap) characteristics of predicated proximal pedestrian targets is unknown. It leads to a missed detection problem in the dense pedestrian scene. In this paper, we analyze the characteristics of ConfMap and the limitations of RODNet. The relationship among ConfMap value distribution, occupied grid spatial distribution and target number is analyzed. Through the CRUW dataset, the target detection experiment of dense pedestrian scene is carried out, and is used to support our analysis.