engleski

Supporting Search and Rescue Activities with Deep Learning

In the case of search and rescue (SAR) activities, time is a vital element influencing probability of a successful scenario (i.e., when a missing person is found alive) and any delays can result in dramatic consequences. Using airborne tools – such as unmanned aerial vehicles (UAV) equipped with cameras – to survey the environment and collect evidence about the position of the victim, can provide a crucial support to SAR operations, by helping responders to focus their search efforts in the right location as quickly as possible, whilst avoiding hazards. However, the task of person detection and classification in images and video sequences is very complex and has not yet been solved in its generality. Recently, deep convolutional neural networks (CNNs) have achieved astonishing results in image classification and object detection. In addition, by removing the final layer that outputs class scores, a pre-trained CNN can be utilized as a generic feature extractor. This paper aims to assess the effectiveness of lower-layer CNN features for detecting lost individuals in a wilderness outdoor environment using aerial imagery. Support vector machine (SVM), which is a discriminative classifier, is used to classify the feature vectors learned by the convolutional network. Our evaluation on a representative image dataset yields advantageous experimental results that merit further exploration.

search and rescue ; unmanned aerial vehicles ; convolutional neural networks ; support vector machines

nije evidentirano