Analysis of Computer Vision Application in The Context of SAR Missions

Abstract

This work includes a general feasibility and performance analysis of some image classification models for Search And Rescue (SAR) operations. Since SAR missions are typically high-wear and high-risk situations for both victims and SAR crew, some teams around the world have been seeking to use technology to speed up rescues and reduce damage. The Croatian Mountain Rescue Service (CMRS) and other SAR teams have been using Unmanned Aerial Systems (UAS) to obtain bird's eye captures of the search area, contrasting with typical low-altitude manned flies. This paper uses images from the HERIDAL dataset to train, validate, and test models. We used two different neural network architectures and eight different training parameters. The accuracy above 98% was achieved, but it does not necessarily mean that the models are appropriate for use in real life, so several considerations were made.