Detection of humans in drone images for search and rescue operations

Abstract

Object detection has solved many problems in different applications like monitoring security, search and rescue operations, semantic segmentation, autonomous driving and so on. Despite this huge success rate in normal ground captured images, it is still a challenging task to detect humans or any other objects from the UAV(Unmanned Aerial Vehicle) captured images due to a few challenges like pose and scale variations, weather conditions, artefacts like people wearing hats, varying attitude and camouflaged environment. In this paper, we propose a novel approach for the detection of humans in aerial images, for search and rescue operations. This method explains how to train the existing high-resolution aerial database of HERIDAL. The EfficientDET deep neural network is trained using a newly generated database to solve the human detection problem. To the best of our knowledge, the proposed method has achieved the best accuracy of 93.29% mAP compared to all existing methods. The proposed method has been compared to the system used by Croatian Mountain search and rescue (SAR) teams (IPSAR) and also with the state-of-art proposed HERIDAL database paper which is based on extracting the salient features, which has slightly worse result compared to the results of this paper.