Localization
Background
Localization as it applies for the AUVSI competition refers to determining the location and orientation of the targets on the groun.
Overview
Location
From a theoretical point of view, finding the location of the targets should be fairly straightforward. Using the metadata of the original images (the location, heading, and altitude of the plane when the picture was taken) along with the camera's resolution, and the extracted location of the targets (from saliency), we can determine the coordinates of the target.
A pixel in an image can be directly converted to its equivalent representation in feet as long as the altitude, field of view, and number of pixels is known.
Orientation
Determining the orientation of the object is more difficult due to the fact that it requires identifying what the actual target is. In order to extract orientation of targets, a CNN can/is being used with one of the classes being the orientation for the target.
Current Implementation
Location
There is a work in progress implementation to determine the target geolocation in localization.
Orientation
A CNN in transfer is being used to identify the orientation of the targets.
Future Work
- Improve the current implementation that determines the location of targets. (testing and fix existing issues)