Convolutional neural network for people counting using UWB impulse radar

Abstract

People counting plays a crucial role in various sensing applications such as in smart cities and shopping malls. In this paper, we propose a data-driven solution that uses a low power ultra-wideband impulse (UWB) radar to count the number of random walking people in an indoor space. A pre-processing signal processing method is applied to clean clutter signals from UWB radar. Instead of the conventional counting methods, which manually extract features and learned from effective data patterns, we investigated deep convolutional neural networks (CNNs) that automatically learn from the data to count the number of people in an indoor space. The CNN model could accurately predict up to 97% accuracy for up to 10 people random walking in an area of 5 × 5 m. The different settings of the CNN models, such as the data input window size, and kernel size in each layer, will be investigated.

Deep Learning for Counting People from UWB Channel Impulse Response Signals

“…A method of counting people in a wide area using IR-UWB radar was introduced in Choi et al [ 11 ]. In Pham et al [ 13 ], the use of convolutional neural networks (CNN) for people counting on IR-UWB was proposed. Human activity recognition using UWB CIR and Wi-Fi CSI was compared in Bocus et al [ 19 ].…”

“…In these methods, statistical methods such as simple thresholding [ 8 ] or analysis of the singular values of the matrix of temporal CIR differences [ 9 ] are used for estimation. While more recent machine learning methods, such as deep neural networks, have been applied to impulse radio (IR)-UWB radar signals [ 13 , 14 , 15 , 16 ], there has been no similar method for UWB transceiver CIR signals. Compared to radar, UWB transceivers have advantages in that they may be used together within a communication system, and that they may have less power consumption.…”