3D-mapping-aided (3DMA) global navigation satellite system (GNSS) positioning that improves positioning performance in dense urban areas has been under development in recent years, but it still faces many challenges. This paper details a new algorithm that explores the potential of using building boundaries for positioning and heading estimation. Rather than applying complex simulations to analyze and correct signal reflections by buildings, the approach utilizes a convolutional neural network to differentiate between the sky and building in a sky-pointing fisheye image. A new skymask matching algorithm is then proposed to match the segmented fisheye images with skymasks generated from a 3D building model. Each matched skymask holds a latitude, longitude coordinate and heading angle to determine the precise location of the fisheye image. The results are then compared with the smartphone GNSS and advanced 3DMA GNSS positioning methods. The proposed method provides degree-level heading accuracy, and improved positioning accuracy similar to other advanced 3DMA GNSS positioning methods in a rich urban environment.
Keywords: GNSS; GPS; autonomous driving; cameras; image segmentation; land application; localization; navigation; urban canyon.