Aerial Imagery Object Detection

Abstract

With the growing advancements towards automated understanding of images and videos, the domain of object detection is a keen area of research towards under standing objects on the ground from aerial images. In recent years we have witnessed significant progress but these algorithms are inclined towards detection in natural scenarios and general images. Not much focus has been given to detection in remote or aerial images. The algorithms do not prove to be optimal when applied to aerial images or sequences. This work addresses the difference and challenges in Aerial Im agery Object Detection as compared to general object detection. Several challenges like changes in scale variation, arbitrary orientation of dense objects lead to incorrect identification of instances. We have tried to address some of the challenges like im proving accuracy of small and dense object detection, handling the class-imbalance problem and using contextual information to boost the performance. We have used a density map based approach on the drone dataset VisDrone-2019 [1] accompanied with increased receptive field architecture such that it is able to detect small ob jects properly. Further, to address the class imbalance problem we have picked out the images with classes occurring fewer times and augmented them back into the dataset with rotations. Subsequently, we have used RetinaNet with adjusted anchor parameters instead of other conventional detectors to detect aerial imagery objects accurately and efficiently. Object detection in Aerial Images holds importance as it is used in applications such as surveillance, crowd counting. Thus, detecting the objects accurately and achieving state-of-the-art results can help in improving some of the existing methodologies.