One of the primary concerns we have had as we inspect and attempt to improve our initial results is that for many frames, our keypoint detection algorithm isn’t able to detect too many keypoints. This creates problems because any inaccuracy in a point’s location will significantly drag the plane of best fit away from what it should be. With more points, any inaccurate points are likely to be diluted, allowing for a more accurate calculation.

To address this, we experimented with a variety of different other keypoint detection algorithms implemented by opencv to observe whether any perform better for our use-case. We also tried using the bounding boxes themselves as keypoints, using the corners and centerpoints of lines between the corners as keypoints.

We then combined the results, appending the results of multiple algorithms into the list.

While there are still some antennas that are comparatively lacking in keypoints, over the course of several images from the video, we hope that they will all be able to gather a significant number of keypoints.