| Abstract [eng] |
In this paper, we are looking at how to solve the drone localization problem using neural networks. A drone without GPS must have other ways to learn its position. Drone traveling long distances can‘t rely on linear odometry with hand-crafted solutions because of various inaccuracies. A map-based system that uses neural networks can help drones navigate long distances without significant inaccuracies and allow them to learn their initial position. In this work, we research capsule networks with aerial images. Anchor images are transformed by angle (from -10 to +10 degrees) to simulate drone direction and compass error. Data sets consist of triplets: anchor image, positive image and negative image. In our experiments, we use triplet loss with a margin value of 0.2 to calculate distances between triplets and be able to start the learning process. The trained modified capsule network model improved learning speed and accuracy compared to the original Hinton et al. model and ResNet model. |
