Visual Semantic Context of Semantically Segmented Aerial Images

This paper proposes a descriptor for the position and heading of aerial vehicles out of down-looking images using the concept of the visual semantic context aided by semantic segmentation and semantic labelled map which can be utilized during aerial navigation. The study presents the derivation of the visual semantic context from the given image while also conducting a feasibility analysis of the visual semantic context by presenting the corresponding error characteristics using both information- and heuristic-based residual metrics over the two contexts. The analysis using semantically segmented aerial images and the semantic labelled map indicates that the proposed concept shows distinct numeric features in a local sense and that it can be utilized as a position and heading fixing tool if associated with exhaustive search and/or data assimilation methods. The local uniqueness of the proposed context also implies that it is possible to use the concept as a validity index for a given aerial image when combined with a filtering paradigm.