A Rotation & Scale Invariant CNN Model to Detect Humans in Disaster Area from Aerial View

Human detection in aerial images is incredibly challenging because of the characteristics aerial images possess. One of the main applications of human detection in aerial images is for search and rescue operations after a disaster. An efficient human detection approach will hugely optimize the search and rescue operation, which helps reduce the loss of lives due to disasters. A custom dataset has been made for this study using an Unmanned Aerial Vehicle (UAV) or drone. The dataset is made to simulate real-life disaster scenarios. Various types of preprocessing are done on the dataset considering the wide range of aerial images, i.e., different scales, orientations, and postures of the human body cover different areas in aerial images. A new convolutional neural network (CNN) based model has been proposed for classification in this paper. The model has been made both scaling and rotation invariant. A comparative study has been done among 12 pretrained models and compared to the proposed model. You Only Look Once (YOLOv3) has been used for transfer learning for the task of localization. The proposed classification model has shown satisfactory results with accuracy as high as 94.18%.