Fast Vehicle Detection in Satellite Images Using Fully Convolutional Network

Abstract

Abstract. Detecting small targets like vehicles in high resolution satellite images is a significant but challenging task. In the past decade, some detection frameworks have been proposed to solve this problem. However, like the traditional ways of object detection in natural images those methods all consist of multiple separated stages. Region proposals are first produced, then, fed into the feature extractor and classified finally. Multistage detection schemes are designed complicated and time-consuming. In this paper, we propose a unified single-stage vehicle detection framework using fully convolutional network (FCN) to simultaneously predict vehicle bounding boxes and class probabilities from an arbitrary-sized satellite image. We elaborate our FCN architecture which replaces the fully connected layers in traditional CNNs with convolutional layers and design vehicle object-oriented training methodology with reference boxes (anchors). The whole model can be trained end-to-end by minimizing a multi-task loss function. Comparison experiment results on a common dataset demonstrate that our FCN model which has much fewer parameters can achieve a faster detection with lower false alarm rates compared to the traditional methods.