Automated GCP detection

Project description

Main Research Question:
How can GCP marker detection be integrated into the existing software of MRR in a robust, sustainable and scalable way?

Sub-Questions:
1. What is the ideal GCP marker for detection?
2. How can GCP marker detection be done in a robust way?
3. How does the data need to be prepared for GCP marker?
4. How can the data best be used to detect future GCP marker?

Context

MRR drones is a company that specializes in photogrammetry for construction projects. The use for that photogrammetry is to do volume measurements of pits or mounds, to see how much fill is needed or how much stuff is in a mound.

The ground control point markers then, are to increase the accuracy of those photogrammetry models. Those markers are normally hand marked in the software which takes a lot of time. Thus, MRR drones wants this process to get automated thus leading to this research.

This led to a choice in marker type, a pipeline design, a data preparation method, and proof of concepts for 2 of the proposed models.

With these results, an advice was given on how to proceed with development of this automation project.

Results

For automatic detection of GCP models in a robust way, the following pipeline should be built: tiling -> filtering -> detection -> identification.

The method of tiling that has been devised can deal with varied sizes of images thus, in theory, working for any possible drone used to survey an area.

As for filtering a small model can be created to filter out images without GCP markers, although the right data needs to be collected for the accuracy to be high enough.

The detection model that seems to be the best is YOLOv5x, thus this should be used. But again to get the accuracy high enough different data has to be collected.