M.S. AAI Capstone Chronicles 2024

3

Dataset Summary

The dataset used for this analysis consists of video frames, formatted to a standard 2448-pixel

width by 2048-pixel height and encoded as 8-bit grayscale images. Bounding boxes were manually

labeled for all visible airborne objects in the images and are classified as helicopters, airplanes, birds,

airborne, drones, and flocks as shown in Figure 1. The airborne label contains objects that do not fall

into the other categories such as hot air balloons and ultra lights. The object labels are text datatypes

and any additional information pertaining to the images such as size and distance to the object are

stored as float datatypes. Figure 2 illustrates the distribution of objects in the dataset. It is noted that

the distribution of objects per image is skewed to favor those with only one labeled object as shown in

Figure 3. Although this finding is not a critical deterrent from the goal of the analysis, it highlights the

limitations of applying the results to an autonomous drone operating in a more congested airspace.

Figure 1

Labeled image used for the training dataset

117