Forum

Author Topic: Non-coded cross targets automatic detection failure  (Read 3192 times)

jansawicki

  • Newbie
  • *
  • Posts: 20
    • View Profile
Non-coded cross targets automatic detection failure
« on: July 11, 2018, 12:36:38 PM »
Hello,

I have a problem with automatic detection of non-coded cross targets. The targets are 50x50cm black/white quadrants, and the pictures (JPGs) were taken with DJI drones (P3A and P4P, same issue) from altitude ranging between 70-120m (GSD 2-4cm).

I have A LOT of test pictures and as I can see, only about 20% of targets are successfully detected but I don't see any relation between target appearance/resolution and detection effectiveness. Increasing the detection tolerance does not help. I've tried applying a median filter on the images, increasing the contrast, and some other tricks, but none of them worked.

I attach link to google drive folder in which i share 3 sample photos captured with Phantom 4 Pro from 90m above ground level:
https://drive.google.com/open?id=1FdzNZsNBRt1hkvMIKCd1np92v1hqmQdU

Can anyone please tell me why these targets fail to be detected? My guessings are either JPEG compression noise or improper exposure. I think that this size of targets should be ok (20-40px) .

Any advise on how to improve automatic detection (during capturing photos or in post-processing)?

I would be pleased even if there was many false markers detected, but with correct ones among them.

Greetings

Dave Martin

  • Full Member
  • ***
  • Posts: 170
    • View Profile
Re: Non-coded cross targets automatic detection failure
« Reply #1 on: July 11, 2018, 05:52:30 PM »
Jansawicki,

Can I just check - have you aligned the images before attempting detection?

Looking quickly at the DJI_000x images you linked to, to me it looks like a pair of white squares laid corner-to-corner on the ground, rather than black and white squares with maximum contrast - can you post a closer image or describe the actual targets?

Also, re
..... I have a problem with automatic detection of non-coded cross targets. The targets are 50x50cm black/white quadrants,... GSD 2-4cm ..... I think that this size of targets should be ok (20-40px) ........
if your GSD is at the 4 cm level, your 50cm x 50cm targets are only 12.5 px

Dave



jansawicki

  • Newbie
  • *
  • Posts: 20
    • View Profile
Re: Non-coded cross targets automatic detection failure
« Reply #2 on: July 11, 2018, 08:59:25 PM »
Dave,

Thank for your reply.

Yes, i did the initial photo alignment (I've tried various accuracy settings from low to high, I don't know if it matters).

The targets were plastic boards containing 2 white and 2 black squares but as you said, the black squares are not perfectly black in the pictures (and the white ones aren't perfectly white either). To increase the contrast during photo capture, we added some dark margin around one of the targets (using dark adhesive tape) - no effect. I find it hard to increase the contrast further without any postprocessing.

You can download another set of photos (taken with Phantom 3 Adv from 50m AGL, GSD around 2cm):
https://drive.google.com/open?id=1wYmVpuSdtsKpBXeW9TbHjESR8v_akxRB

This set of photos covers 2 targets (with and without dark adhesive tape around it) - although both targets seem pretty similar, only one is being detected automatically (the one away from the railroad, without adhesive tape margin).
These pictures are generally much darker, but the black areas on the targets are still just dark grey, so I don't know if it's possible do achieve higher contrast.

The photos were taken using Pix4D Capture app, which I belive automatically sets all exposure parameters and doesn't add any color corrections/filters to the images.

3DWinter

  • Full Member
  • ***
  • Posts: 103
    • View Profile
Re: Non-coded cross targets automatic detection failure
« Reply #3 on: September 20, 2022, 04:14:21 AM »
Hello Jansawicki,
Have you learned anything new regarding more reliable automatic non-targets best practices?
Thanks

Kiesel

  • Sr. Member
  • ****
  • Posts: 332
    • View Profile
Re: Non-coded cross targets automatic detection failure
« Reply #4 on: September 20, 2022, 01:00:00 PM »
3DWinter,

it works quit well unless your targets are not too small and you haven't structures which looks similar to targets, for example cobblestones.

In another post Alexey has written "The size of non-coded targets actually are not so critical (at least there are no max limits), but I suggest to have at least 4-5 pix radius for them." But perhaps this was meant for the type circle and you have to double this for the non-coded cross.


You can also use targets of different sizes in one project to overcome the size limitation.



Best regards

Kiesel 
« Last Edit: September 21, 2022, 12:36:07 PM by Kiesel »

3DWinter

  • Full Member
  • ***
  • Posts: 103
    • View Profile
Re: Non-coded cross targets automatic detection failure
« Reply #5 on: September 20, 2022, 10:14:03 PM »
Thanks for that.
Do you know what is "Tolerance" and "Maximum Residual (pix)"? And, do you adjust these parameters?
Thanks

Kiesel

  • Sr. Member
  • ****
  • Posts: 332
    • View Profile
Re: Non-coded cross targets automatic detection failure
« Reply #6 on: September 21, 2022, 08:53:44 AM »
3DWinter,

I use higher tolerance values (these are dimension less) if not all markers are detected at the first time for standard values (2). But before I use any higher tolerance value I erase all already detected markers, because otherwise I get double markers with the same number in the marker list. 
The maximum residual is only avaiable for non-coded targets (and not for coded targets). This value is in pixel and to my understanding it means the tolerated maximal residual of the detectected circle center or cross section.

Hope it helps

Kiesel
« Last Edit: September 21, 2022, 02:07:07 PM by Kiesel »