| Popis: |
Search and Rescue (SAR) operations in hazardous and difficult-to-reach locations can be significantly enhanced by introduction of unmanned aerial vehicles (UAVs). Processing of aerial images acquired by the camera mounted on UAV with the goal of person detection is challenging task due to very small footprint, varying textures, and shapes of the potential target and of various types of background. Although some classical approaches exist, lately the use of convolutional neural networks has made significant improvements in object detection performance. Precision and recall of these state-of-the-art methods have been concurrent to the results of visual inspection conducted by humans. Therefore, we comment and suggest possible further improvements with particular focus on introduction of context, image sequences and real-time processing |
