Vision based navigation musings / Computer Science / Forums

Forums

4hv.org :: Forums :: Computer Science

« Previous topic | Next topic »

Vision based navigation musings

Move Thread

LAN_403

rp181

Sun May 09 2010, 01:58AM

Registered Member #1062 Joined: Tue Oct 16 2007, 02:01AM
Location:
Posts: 1529

I was mulling over possibilities for vision based aerial navigation, and I came up with this. With a camera starting from above, a calibration sheet with multiple points is placed under it, with known distances. From this, The platform could calculate its height, as well as offset. Now, when the platform moves, it acquires new "significant" points, anything that can be tracked (simple as blob detection). From the previous calibrated sheet and the points remaining in view, it calculates the new distance for the new point. It continues in this web like fashion as it travels along.

Red is calibration points, blue dots are random blobs, and green is the calculated distance measurements. The black box is the cameras field of view.

The platforms tilt may be measurable, depending on the magnitude of the perspective distortion/ camera resolution. If not, its as simple as a stabilized camera platform.
Thoughts?

IntraWinding

Sun May 09 2010, 11:42AM

Registered Member #2261 Joined: Mon Aug 03 2009, 01:19AM
Location: London, UK
Posts: 581

It's too complicated for my tiny mind to tell if the scheme is flawed, but it sounds convincing for an observer at constant height over a flat terrain with a constant viewing angle (attitude) and a non distorting camera lens.

If you move to a real world environment with a 3D terrain and varying observer altitude and attitude (e.g. model plane) I'd have to resort to models and tests and maybe even a bit of maths etc.. I suspect it might be vulnerable to 'optical illusions' (like us), confusing one sort of change with another, but some added intelligence could help filter out consequential poor decisions. For example, as the observer moves in an approximately known direction the relative positions of the points should move in a way that's consistent with the model it has formed of their locations (a bit like moving your head from side to side to clarify things when some odd perspective confuses your visual centres).

(Just thoughts - I could easily be completely wrong).

Bjørn

Sun May 09 2010, 12:27PM

Registered Member #27 Joined: Fri Feb 03 2006, 02:20AM
Location: Hyperborea
Posts: 2058

There are a few things to note about this.

Errors will accumulate, soon the height and position will be very uncertain. Unless you fly in a special pattern you will get lots of conflicting information that has to be resolved. Finally it is not easy finding a reliable blob unless you litter the "unknown" area with blue bits of plastic.

You can greatly improve on these things if you have an accurate altimeter, then you can measure your ground speed accurately and from that true distance, that way errors wlil not accumulate.

rp181

Sun May 09 2010, 02:07PM

Registered Member #1062 Joined: Tue Oct 16 2007, 02:01AM
Location:
Posts: 1529

If this worked out with a decent success rate, I was planning to use it as another source of data in a kalman filter, to further refine data. I was planning on using a SCP1000 barometric sensor for altitude.

Now that I think about it, I recall seeing a similar system that targeted houses, and worked "even with 80% cloud cover".

At the ground station, I have access to satellite imagery, so a simple image matching program could provide semi-accurate coordinates when GPS is not available.

Bjørn

Sun May 09 2010, 02:21PM

Registered Member #27 Joined: Fri Feb 03 2006, 02:20AM
Location: Hyperborea
Posts: 2058

The sun keeps moving so simple image matching is not that simple.

Arkin

Sun May 09 2010, 02:38PM

Registered Member #2140 Joined: Tue May 26 2009, 09:16PM
Location:
Posts: 53

The sun affects everything similarly, so wouldn't a SAD algorithim still return the best match? The search area would be confined to a certain threshold based on time, last recorded speed, and last known GPS coordinate.

The Virtual Earth imagery has little shadows, but the USGS Urban Earth map has alot of shadows. Because in SAD, the order of the pixels in a block does not necessarily matter (a block matched with a block rotated would still match, if you were to alter the algorithm a bit), this would be a good map to compare to. I think the main problem here is the lack of up-to date imagery.

Bjørn

Sun May 09 2010, 03:29PM

Registered Member #27 Joined: Fri Feb 03 2006, 02:20AM
Location: Hyperborea
Posts: 2058

If you rotate the block 90 degrees so that it is perfectly aligned again then it will work, yes. If not you have made a completely different algorithm that you are not telling us about.

Carbon_Rod

Mon May 10 2010, 10:53AM

Registered Member #65 Joined: Thu Feb 09 2006, 06:43AM
Location:
Posts: 1155

Vision systems have many issues, but it really depends on how much CPU power you have.

a $50 GPS unit is a more common solution.

Moderator(s): Chris Russell, Noelle, Alex, Tesladownunder, Dave Marshall, Dave Billington, Bjørn, Steve Conner, Wolfram, Kizmo, Mads Barnkob