Ground-to-Satellite Image-Based Geo-Localization

Join Zoom Meeting
https://zoom.com.cn/j/68525327077?pwd=eFJSOHBGVnMrWXNFVG0wUThIWkVkdz09

Meeting ID: 685 2532 7077
Password: 004285


Abstract:

Image-based geo-localization is widely used in many applications due to its high accuracy and low cost. Over the past ten years, ground-to-satellite image-based geo-localization (G2S-IBL) has gradually drawn more attentions, since the satellite imagery becomes easily accessible and has the full coverage of the earth. G2S-IBL is often solved as cross-view image matching/retrieval, where the geographic location of the query ground-view image is found by matching it with a large database of geo-referenced satellite images. This problem remains very challenging due to the lack of common features in the query and database images with drastic change in viewpoint. The huge difference in the appearances of two views also causes traditional image descriptors-based matching to fail. In this thesis, we try to find effective and efficient approaches to solve the cross-view image matching. Based on the cross-view matching results, we propose a Markov localization framework to solve the localization task on a moving vehicle. For the cross-view image matching, we leverage on the recent success of deep learning to propose the CVM-Net and LFT-Net. CVM-Net generates the global descriptor of the input image with rotation invariance. It can be applied in the case that the orientation of ground-view image is unknown. LFT-Net generates the global descriptor of the input image incorporating the orientation information. LFT-Net achieves higher recall provided that the orientation of ground-view image is known. Our experimental results show that our proposed networks significantly outperform state-of-the-art approaches. Based on the CVM-Net, our proposed Markov localization framework can continuously localize the vehicle which is running in the real world within a small error.