People counting using visible and infrared images

"We propose the use of convolutional neural networks that consider as input four channels images (RGB+IR) for counting and positioning people in images. Our data set was made of images based on photographs taken from a drone using a dual FLIR camera. Comparison between 3 (RGB) and 4 (RGB+IR) channels are studied for different lightning conditions. The four channel network performs better in all situations, particularly in cases of poor visible illumination that can be found in real night scenarios. The average precision of this network on a testing data set (independent from the training one) is approximately 1 cm in nding the positions of pedestrians (from 15 and 30 m altitude images) and 0.0001% in the relative counting error." Proyecto final Ingeniería Informática (grado) - Instituto Tecnológico de Buenos Aires, Buenos Aires, 2020