ABSTRACT
The advancement of object detection has revolutionized numerous fields, including security surveillance, search and rescue missions, semantic segmentation, and autonomous driving. However, detecting humans or objects in images taken by Unmanned Aerial Vehicles (UAVs) presents unique challenges such as variations in poses, differences in scales, adverse weather conditions, and environmental interferences. This study introduces a novel method tailored for human detection in aerial imagery, with a specific focus on optimizing performance for search and rescue missions. Our method involves training the Efficient DET deep neural network employing a recently compiled high-resolution aerial database named HERIDAL. To the best of our knowledge, our proposed approach achieves a remarkable accuracy of 93.29% mean Average Precision (mAP), surpassing existing techniques. We thoroughly compare our method with the systems employed by Croatian Mountain Search and Rescue (SAR) teams (IPSAR) and the HERIDAL database paper, which depends on salient feature extraction. Consistently surpassing both benchmarks, our method demonstrates exceptional efficacy in detecting humans from UAV-captured images, highlighting its superior performance.
