Deep Neural Networks for Ortophoto-Based Vehicle Localization

Navigation of unmanned vehicle especially using orthophoto is a topic of active research. This paper is dedicated to study of different methods of orthophoto-based localization methods. For this task new dataset was created. It consists of pairs of ground level and bird’s eye view images collected on vehicle test site of the technology contest Up Great “Winter City”. Different deep network approaches to localization were used: 1) embedding-based, 2) based on synthesis of bird’s eye view using Pix2pix conditional generative adversarial network and masked cross-correlation in map subwindow. The second approach has demonstrated good applicability for the proposed dataset. Mean absolute error of localization on known scenes reached 1 m. The average total time of bird’s eye view generation and subsequent localization is from 0.1 s to 0.2 s. This is an acceptable quality for the task solution and its further use as part of the navigation systems of unmanned vehicles.