Abstract
To successfully adhere to flight plans, aerial vehicles must keep track of their location in 3D space, which is usually reliant on external references such as GNSS which are susceptible to interference. To develop self-reliant onboard positional localization, a workflow using 360-degree panoramic images in an image-based localization system using a Deep Convolutional Neural Network is proposed. 360-degree panoramic images have the advantage that they take into account visual information from all angles. Model performance is also enhanced by generating synthetic data from a 3D model of the region of interest created via photogrammetry techniques. The performances of different training configurations are compared, and the configuration with mixed real and synthetic data exhibits the highest performance, an approximately 10 to 15 percent improvement over using solely real data. Additional image augmentations also further reduce the localization error by 8 to 15 percent.