SkyScan is a Hardware/Software demonstrator of a collection technique to create labeled datasets using a tip and cue method while requiring minimal human involvement. The tip and cue method leverages multiple sensors – one to “tip” the system that something of interest is nearby and another to “cue” it to to take a photograph.
The SkyScan system is made up of into two sections: the Capture Kit and the Dataset Pipeline. The Capture Kit is a tripod-mounted pan-tilt-zoom camera which is directed by a low-cost companion computer equipped to receive position broadcasts from passing aircraft. It stores both imagery and the unique identifier of the aircraft concurrently. Following collection of data, the Dataset Pipeline is a series of tools and techniques to convert the imagery and metadata into a rich, labeled dataset. This pipeline demonstrates how to “bootstrap” a machine learning model, going from no imagery to a usable multi-type classifier.
The design of the system is intended to be flexible, allowing for the addition of other tippers (i.e., other types of RF sensors, cameras, etc.) or the addition of other collection mechanisms (i.e., acoustic, vibration, etc.). This demonstrator seeks to showcase the technique and make available the insights and technical files to recreate our work.