Weakly Supervised Semantic Segmentation for Range-Doppler Maps

Deep convolutional neural networks (DCNNs) have been successfully applied for object detection and semantic segmentation of radar range-Doppler (RD) maps. However, training a DCNN requires many annotated examples that are costly and difficult to create. In this work we present a method that reduces significantly the manual effort involved in the annotation of RD maps to train a DCNN for segmentation. A 40 times reduction in manual labelling effort is achieved because the annotation of each RD map includes only the class of the objects instead of drawing a polygon around the corresponding cells. The localization of the objects is performed by tracing back from the output to the input of a classification neural network. Experimental results show that our approach achieves robust localization performance in complex real-world urban scenarios as observed with a low-cost automotive radar. Furthermore, we show that our approach performs similarly to DCNNs that are trained with a publicly available dataset in which localization information is provided.