People counting based on CNN using IR-UWB radar

Abstract

People counting serves a vital role in sensing applications. Impulse radio ultra-wideband (IR-UWB) radar, which has strong penetration and high-range resolution, has been extensively applied to detect and count people. Current signal processing methods that rely on IR-UWB radar include three basic steps: the removal of the direct current (DC) component, bandpass filtering and clutter signal removal. An environment-dependent threshold is manually established to select effective peaks for counting. However, the steps that are employed to obtain cleaner signals may also eliminate significant information. In this paper, a novel approach using convolutional neural networks (CNNs) is proposed. This data-driven method learns and directly obtains features from radar data and analyzes them to automatically produce results. It addresses the challenge of counting people in various complex scenes, in which signal processing methods are inadequate. A series of experiments are conducted in the Caffe platform; the results indicate that: (i) some signal processing approaches are harmful rather than beneficial when a CNN is employed; (ii) the proposed method has considerably good accuracy and stability in narrow spaces which have great interference on the signal processing methods; and (iii) the results reach 99.9% accuracy for queue counting.