Accurate People Counting Based on Radar: Deep Learning Approach

Abstract

In this study, a novel radar-based people counting (PC) method is presented using the deep learning (DL) approach. The DL algorithm is a great tool that enables the automatic formation of the optimal features; however, it must be utilized carefully, considering the domain knowledge to prevent the concerns of learning unnecessary information, followed by overfitting. To address the problem and successfully apply the DL framework to the radar-based PC, we propose three novel solutions. First, we establish the preprocessing pipelines to transform the raw signals into a suitable form for network inputs. Second, a network architecture is newly proposed considering the radar signal characteristics and PC application. Finally, we propose several data augmentation strategies to artificially increase the size of training data. It was observed from experiments using real measured data that the proposed DL-based PC approach outperforms the conventional PC methods.