Hi all,
looks that with the spread of processing methods and technology,
we can make things that were available only to national security agencies, once upon a time.
Take a look at the maps of the flight made May 18th:
one day to fly, one day to upload, one day for processing, (one day for grill party), one day to download.
Few of you might remember the 3km long flight with EasyStar (a single line), stitched with Microsoft ICE.
The map above is 2.2x2.2km. It took 746 photos to make this map.
Funny fact: if the aircraft would fly straight, the surface covered would be >16km2 instead of just 4.5km2 (null side overlap).
The overlap used was 70% side, 80% along.
Comments
@Loljze "with good GPS and weather conditions a 3-5 cm absolute accuracy of flightline can be achieved."
What is the reference measurement? It could be you are using a GPS for a few KUSD.
"Maybe some additional IMU (in 5000€ range)"
...and a Hellfire or two ;-) This is more than half of Pteryx Lite pricing and a quarter of Pteryx Pro. At the current market conditions where large companies sit comfortably in a cessna and the newcomers believe you can make it of rubberbands and RC model, we are targetting the low price regions. We have the equipment for startups that want to challenge the larger companies with much less startup capital.
The only thing I'm confused about is why you think "All flying GPS RTK or not RTK have horrible accuracy compared to ground measurement strictly because the most uncertain value, teh altitude, cannot be filtered out accurately in the air." There's nothing magical about the GPS receiver or antenna being on the ground versus in the air. It doesn't know, it doesn't care. GPS is used on rockets and spacecraft. Altitude measurements don't suffer.
The professional photogrammetric companies use differential (carrier phase based) GPS for a reason. GPS works at the speed of light. The dynamics of anything humans fly through the atmosphere are irrelevant. Sure it may take ~10+ millseconds to compute a position in real time, but the data that went into that position has a very, very accurate time tag. You can either predict forward what the position is at a given instant - say when a camera trigger sends a voltage pulse, or better interpolate the position accurately later during post processing. The limiting factor here is your shutter speed ( crank it up :) ) and ground speed - or get into motion compensated cameras, but we're not going there.
While Pix4d doesn't have any real technical data on their site they do say that, of course, differential GPS helps.
In your example the 746 GPS measurements are essentially averaged and some of the error does tend to cancel, but it's not as much as you might think. There could still be a bias of a meter or two. With carrier phase differential there might be a bias of a centimeter or two. This is where the improvement comes. Of course there are other errors involved, but this helps.
In most situations the addition of a few high quality GCPs to the processing would be the thing to do, but again the idea of no GCPs is what intrigued me.I think you are mixing things. It doesn't matters where the aircraft is relative to flight plan and how it maneuvers, it is precisely about if it is able to know the position when making a photo. All flying GPS RTK or not RTK have horrible accuracy compared to ground measurement strictly because the most uncertain value, teh altitude, cannot be filtered out accurately in the air. This variable of the error elipsoid cannot be flattened anymore and spreads overall horisontal precision since all you are truly measuring is time phase, what impacts all dimensions. Once you essentially liberate time-averaged 2D problem into 3D problem (because of accelerations along Z-axis that you never do on the ground), the GPS position error fit 'explodes'.
Then, even this low accuracy, is not dominant factor towards IMU and mounting precision.
"When Pix4d does its adjustments the only way it can add georeferencing info is via the GPS coordinates you give it."
Exactly. In fact, it takes 746 measurement estimations out of 746 photos and the errors cancel out to some degree. It works that much that I am observing about 1m accuracy. But their initial accuracy per photo is a few meters at best, RTK or not. Military achievements suggest it is almost impossible to make it better and error contribution suggests there is no 80/20 rule, the stupid mounting screw and uneven lenses contribute as much as GPS accuracy.
The accuracy of RTK or the post-processed version is actually better in an aircraft compared to ground work. Excellent satellite visibility and little multipath give a GPS antenna in the sky an advantage over one on the ground. While it might not be a large advantage - when the ground isn't too cluttered - airborne RTK certainly isn't worse.
Now if you are talking about navigating a UAV precisely that will be more a function of your control system. If you've ever seen a rotorcraft maneuver with RTK vs autonomous GPS you will appreciate the difference.
When Pix4d does its adjustments the only way it can add georeferencing info is via the GPS coordinates you give it. If you can change their uncertainty from a few meters to a few centimeters that will help the situation.
PPL seem to not understand that from 300m AGL the main limitation is IMU accuracy, when we talk about non-military toys under 100KUSD or so. The projected precision is poor, about a few meters at best.
Also not everybody realizes that RTK from a moving plane is FAR from delivering the same precision as on the ground, because the true altitude is moving randomly up and down due to turbulence.
A simple projection errors are some 5-25m (occasionally probably 50m) are typical even for machines like Skylark which are military drones used for artillery fire targetting and similar. Therefore I don't agree with geodesic obsession of taking RTK from the air. The GPS amounts to maybe 20-30% error (you can shave maybe half of it due to 3D motion and accelerations or rather 3D jerks, certainly not to a few cm area), the rest are IMU and camera calibration+mounting. Don't make your camera mount of foam so it is different in each flight and you already win more than all those RTK processings. Sub-meter projection accuracy you cannot avoid without mechanical gyros, large scale camera calibrating range and similar things. In practice - forget it, a toy that uses it is very high quality military mini UAV and it is not going to be 10x cheaper in the next decade for the same reason that passenger jets are not 10x cheaper today than 2 years ago.
The precision is achieved statistically and by cross correlating GPS through optical matching.
I cannot say if the accuracy is sub-meter (no ground equipment), but is some 50 times better than Google Earth accuracy which tends to be somewhere around 50m 100km from larger Polish cities.
Another holy grail of photogrammetry is dispensing with the ground control points. Getting GCPs in an area with easy ground access is no big deal, but for harder to access areas skipping the GCPs and letting the UAV do all the work is kind of nice.
Certainly with a few accurate (cm - level) GCPs your overall absolute accuracy is going to be in the few decimeter range for the most part. There may be certain circumstances where the Pix4d methodology doesn't work so well, but with all the generated contrast GCPs distortion is greatly reduced and a non-metric (uncalibrated) camera works quite well.
I wonder what the absolute accuracy would be like with some highly accurate UAV positions? The hardware is cheap and relatively lightweight. A Ublox Lea4T attached to a Sparkfun Logomatic and you've got 10 Hz GPS carrier phase data that will process to a few cm accuracy. Yes, the "catch" is in the processing if you haven't done it before, but there are free tools. Update the UAV position data via time tags and what type of improvement do you see in Pix4d's accuracy? Are we sub-meter now?
Your actual application will determine if you really care, but sub-meter absolute position accuracy without GCPs and using inexpensive equipment would be pretty cool.
@Lojze The parameters found magically by pix4d are not published but its internal machinery. However I have in-house camera devignetting, whitebalance equalizer and lens corrector (geometrically calibrated), I found that it is cool to upload that to any application I have and it stiches better (in my ftp there are examples in directories _proc which are processed /lens corrected).
I have removed 3 or 5 slightly blurred photos from the image (there was some turbulence and there was only 1/400 exposition if I remember well).
I have no access to RTK GPS, I am not national geodesic research institute (yet).The camera is Canon S90 (Sometimes we fly with double payload, one IR one visual).
So far my major showstopper was the fact that flights longer than 1 hour usually have changing whitebalance and lighting conditions (you fly with fixed settings or else everything goes in all directions during automatic whitebalancing and exposition). Therefore you accept some gradual brightness falloff or rise and counter it after the flight if you really have to fly in the evening - and then the results are as on the image. But more than 2 h starts being difficult for those lighting reasons, camera battery endurance etc (small problems start to pile up).
BTW I would be happy to find a kind of automatic blur detector or a mathematical formula that is not fooled by changing landscape.
Absolute accuracy better than 1-2m on a few major crossings. Unable tell better (Garmin Geko).
Anyway the pixel size is 10cm (just to get something round) and since we are not expecting 10cm absolute accuracy, when you do comparative analysis, you might need to do optical matching first.
You must keep in mind that teh result is a projection of 746 coordinates from 10Hz GPS plus dead reckoning at 32Hz, flying both sides and IMU projection - it may go down to a decimeter accuracy in certain areas, but mostyl I woudl say 1-2m. There is least suqare fit report on that processing that says:
Table 2: Bundle block characteristics.
mean reprojection error 0.903891 [pixels]
Table 3: Localisation accuracy in metres without using ground control points.
geo localisation variance rho [m]
latitude direction (x) 2.28658
longitude direction (y) 2.04749
altitude direction (z) 6.34011
(altitude direction variance was a few times lower on flat terrain)
Thanks for the info. I'm sure the final image's relative accuracy will generally be in the 1 meter to a few decimeter range, but I was curious as to how good its absolute accuracy was without any ground control points.
This is getting more into the professional mapping side then the UAV development side, but typically you would check the accuracy by collecting some ground control points with survey quality GPS and look at the coordinate errors. Pix4d didn't seem to have much technical data on their site, but suggested that absolute accuracy might be in the 1 to 5 meter range and one example they had seemed to be a bit worse.
This becomes important when you want to overlay imagery collected at different times and do change detection analysis. Especially when you might be interested in areas of just a few, a few dozen, or even a few hundred pixels (using the great resolution you can get from a small UAV).
@John the miracle is, the final image seems to fit satellite photo within 1m just by shifting the map by hand, maybe much better. Those are multigig TIFFs and I don't have better referenced sat data at hand (we are UAV makers that try to understand every aspect of the user's work).
@Lojze impressive processing chain. You might want to contact pix4d about your processing idea, I bet they want to hear that.
-
1
-
2
-
3
of 3 Next