After we have the output of the range-velocity-acceleration matched filter bank, we still don't have a list of space debris detections. We need to use some machine learning algorithms to cluster out results. I tried K-means and some other techniques, but ended up with a physics based model, which compares expected range and velocity with measured range and velocity, relative to some trial point. This works remarkably well. I also needed to use the CLEAN algorithm to remove the effects of the range-doppler ambiguity (range smearing next to strong targets).
Here's a labeled plot with before and after detections.
![](https://blogger.googleusercontent.com/img/b/R29vZ2xl/AVvXsEhZUe_IfwosP4fxnBYzzowXK1KTx1F7pFZCFA7Z-7u9Tu-BXvBLrcoAT1PTZ0qmYzhOg23FfLULAigeVAImKrrIaasOQQG2Z4WH-rrU-HDaqUWA78OBKrz-TMFKo6r-HQ-1_QVOZOKaB94/s400/im-000004-0.png) |
Matched filter output, time on y-axis, and range on the x-axis. Colored points mark detections. |
![](https://blogger.googleusercontent.com/img/b/R29vZ2xl/AVvXsEhu5HzTCQUTejaqN-46aty-eix8YdJ7rRovX9ZnDZnmrojddYAxi9bizQbaDXDe7j9yH0b5wEMldEU702jTIVB73xNxeAPbjEOOWrBBm19_hqhrVRHs9dZSJ-whj_X8699ruUia4uXcytU/s400/im-000004-1.png) |
The original matched filter output. It seems that I've missed one weak echo. |
Comments
Post a Comment