Quicker Orthomosaics

24ab54_ac551b2b16984b19914ad65eb0cfe1fa.jpg

Idea:

Can you reduce processing or flight time for a mosaic by flying an "overview flight" from high altitude, processing it at low quality, exporting the resulting mosaic, and using that mosaic as a "trainer" image for the primary low altitude, high quality, second flight? And ss this a way to speed up processing time for data sets where geo-tagging is not available?

Test:

To test the idea I used Google Earth to simulate a drone flight. I first made a box and then drew guideline grids to define the area of interest. Thee resulting 129 screenshots (aerial images) were batch trimmed to remove stationary features (google earth logo, nav tools, etc.). Note that no geo-tagging was applied to the screenshots. This was then repeated at a much higher altitude for the "overview flight". The resulting pass was comprised of 23 images. This processed in under 5 minutes and resulted in the mosaic to the right. 

I then added the overview mosaic as a image to the low altitude primary data set. Unfortunately, the SFM software failed to match the high altitude mosaic to any of the low altitude images. I think this may be because:

  • The software can't match ‘point of view’ images to an orthographic image.

  • The scales of images at each altitude are too far apart for SFM feature detection to work.

  • The resolution of the high altitude mosaic was too low to match features from the low altitude data set.

Below is a side by side comparison of the low altitude and high altitude orthomosaics.

24ab54_8975e8e8f1b443ff90a689f477b0623a.jpg

Conclusions: 

  • Taking images at higher altitude allows less flight time.
  • Injecting a complete high level mosaic did not help as the software could not match images to it.
  • The high level model does not take long to process and could be used for detailed 3D flight planning with "recent time" obstacle avoidance.

 
Future work:

  • Time comparison between low altitude mosaics with and without high altitude context mosaics. 
  • 3D accuracy test: is there a difference between the 3D accuracy of the model when using context images? Requires ‘good’ DEM data to compare with (LIDAR).
  • Figure out which failure hypotheses is correct and figure out how to get low altitude pictures to match features with the high altitude mosaic.

The next post in this series will address at least one if not all of these points. 

Tim Paszalek