People counting using visible and infrared images

• Published in: Neurocomputing (Amsterdam) Vol. 450; pp. 25 - 32
• Main Authors: Filipic, Joaquín, Biagini, Martín, Mas, Ignacio, Pose, Claudio D., Giribet, Juan I., Parisi, Daniel R.
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 25-08-2021
• Subjects: Convolutional neural network; People counting; Visible and infrared images; People counting; Convolutional neural network; Visible and infrared images
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2021.03.089

We propose the use of convolutional neural networks (CNN) for counting and positioning people given aerial shots of visible and infrared images. Our data set is entirely made of semi-artificial images created from real photographs taken from a drone using a dual FLIR camera. We compare the performance between the CNNs using 3 (RGB) and 4 (RGB + IR) channels, both under different lighting conditions. The 4-channel network responds better in all situations, particularly in cases of poor visible illumination that can be found in night scenarios. The proposed methodology could be applied to real situations when an extensive data bank of 4-channel images is available.