Vehicle pose estimation in WAMI imagery via deep convolutional neural networks

Abstract

Wide Area Motion Imagery (WAMI) are usually taken from unmaned air vehicles at low frame rates, and having very wide ground coverage. These images serve as rich source for many applications like surveillance, urban planing and traffic monitoring. Thus, understanding WAMI imagery exploitation has been gaining more interest recent years. In this paper, we focus on estimating the pose of vehicles in WAMI imagery. The difficulty of this task lies in that a vehicle only occupies a very small low-contrast region with confusing visual appearance in a WAMI image, which raises a serious problem for conventional approaches based on low-level image cues or priors. In this paper, we tackle this problem by adopting deep learning approach, using deep Convolutional Neural Networks (CNN) to learn the pose of vehicles in WAMI images. The proposed deep convolutional network based pose estimation exceeds baseline by 31.5%. Furthermore, we analyzed the effect of different level of context information on the estimation accuracy.