When Clutter Reduction Meets Machine Learning for People Counting Using IR-UWB Radar

Abstract

People counting provides key information 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 require to establish an environment-dependent threshold manually. Due to the high sensitivity of the IR-UWB radar, the wide diversity of scattered waveforms would bring false alarms. Clutter reduction serves a vital role in signal processing steps to obtain the signal reflected only from the target, while it may also eliminate significant information. In this paper, data-driven solutions based on two machine learning algorithms, the random forest and convolutional neural network (CNN), are proposed to address the challenge of counting people with complex changing scatters. These data-driven methods learn from selected features from radar signals or directly obtain features from radar data and analyze them to automatically produce results. A series of experiments are conducted in the Orange and Caffe platform, and the results indicate that: (i) In data-driven solutions, clutter reduction methods are harmful rather than beneficial for data analysis, verified by discussing four representative clutter reduction methods. (ii) Random forest classification for selected time-domain features in radar signals before complex clutter reduction reaches 91.5% accuracy in testing environment. (iii) CNN provides an automatic counting solution learning directly from radar data.