Multiple Walking People Classification with Convolutional Neural Networks Based on Micro-Doppler

Abstract

Classification of multiple walking people is researched based on radar micro-Doppler features in this paper. An architecture of deep convolutional neural networks without pooling layer is designed to extract the inherent features of micro-Doppler and complete the classification automatically without specific feature selection. The pooling layer is not used in the convolutional neural networks in order to preserve more subtle micro-Doppler features to improve the classification accuracy. The radar data of different types of pedestrians including one, two and three walking people are collected in the outdoor environment. Then the deep convolutional neural networks is trained with a small data set and the average accuracy of 95.55% is achieved.