Human Detection in Images Captured by UAVs in Natural Environments

Unmanned Aerial Vehicles (UAV) are nowadays used for many different purposes. Particularly, they are very useful in search and rescue operations in situations where the area to explore is either big or difficult to access. Images collected through UAVs could be processed by deep learning-based models to automatically detect human presence and guide rescue teams to specific locations. To this end, this paper employs Convolutional Neural Networks (CNN) to detect humans on images of the HERIDAL dataset. Original images are preprocessed and divided into smaller patches to keep the original resolution when feeding them to the CNN and common data augmentation techniques for images are employed in order to increase the variety of images in the training set. The results showed a very good capability of the CNN models tested when detecting humans, achieving AU-ROC values of 0.95.