lxlegend99
Posts: 4
Joined: Thu Apr 18, 2019 2:23 am

Object Detection

Tue May 07, 2019 4:30 am

I am currently working on a project build an object detection program for the drone to detect cracks on the wall of the buildings. May I know whether YOLO object detection can be used in the Raspberry Pi OS ?

blimpyway
Posts: 184
Joined: Mon Mar 19, 2018 1:18 pm

Re: Object Detection

Sun May 12, 2019 8:23 am

A time ago I have downloaded and compiled darknet as described here https://pjreddie.com/darknet/yolo/

On a Pi 3 A+, having only 512MB memory, it could only run the tiny variants of yolo v2, v3

It worked but much slower than what I expected from some youtubes where it got ~1fps

PS: I assume you have to train your own nn for wall cracks.

See also this recent post: viewtopic.php?f=41&t=240071

User avatar
omegaman477
Posts: 148
Joined: Tue Feb 28, 2017 1:13 pm
Location: Sydney, Australia

Re: Object Detection

Tue May 14, 2019 7:33 am

YOLO is a pretty big task for the RPI, as mentioned, RAM is a factor. But I think that YOLO is overkill for such an application. Look into the various machine vision libraries/applications around, most should run well on a RPI. I am thinking about basic edge detection approach.
..the only thing worse than a stupid question is a question not asked.

blimpyway
Posts: 184
Joined: Mon Mar 19, 2018 1:18 pm

Re: Object Detection

Thu May 16, 2019 8:12 pm

Yeah makes more sense to just make a big batch of photos of the building on SD then process them on a properly sized CPU/GPU combo for your problem.

If video link image is good enough you can use a beefy pc/laptop on ground for object detection in real time.

But this does not seem a problem with real time response requirement.

User avatar
omegaman477
Posts: 148
Joined: Tue Feb 28, 2017 1:13 pm
Location: Sydney, Australia

Re: Object Detection

Fri May 17, 2019 7:34 am

Was thinking about this again, as I do. Cracks in an external concrete wall may be difficult to detect due to the lighting condition. I think the idea of taking the images and then on ground processing is good. I would also consider a couple of cameras, to get a aka 3D depth perspective, and maybe a 3rd that is IR only, looking for the colder cracks?

Taking a series of photos under natural light, then one with an onboard strobe, would maybe reveal more detail.

I remember a project I worked on for Satellite photo detection of forestry dieback. We overlaid IR, UV, Vis images to get the data we needed.
..the only thing worse than a stupid question is a question not asked.

User avatar
Gavinmc42
Posts: 3899
Joined: Wed Aug 28, 2013 3:31 am

Re: Object Detection

Fri May 17, 2019 8:19 am

I suppose it depends on the crack sizes, small ones will be hard to detect unless high res image data is used.
Most AI vision stuff in real time is around 300x300 pixels.

Could a florescence spray help, it would tend to collect in the cracks and shows up under UV LED spotlights.
Very easy to detect that, pretty much a standard practice.
Spray cans on drones next to walls, "Officer, I am just checking for cracks" :lol:
UV graffiti? A new art form?
Those Highlighter pens are pretty good, cut them open and squeeze out the goop into water :D
Yellow ones are best :lol:

I have used Ximea Hyperspectal USB cameras with Pi's but not for fluorescence, yet.
I'm dancing on Rainbows.
Raspberries are not Apples or Oranges

Return to “Other projects”