Bale collection/carting optimisation

[ptd006]

Hi,

I have been developing algorithms to optimise the collection of bales (i.e. best trailer placement etc). Enough people were interested that I made a usable prototype. The idea is that when baling you log GPX waypoints at the bale locations or maybe map them with a drone afterwards (only worth it for big fields really). You then upload the .gpx file of bale locations and the webapp generates a .gpx route you can follow when collecting the bales. It assumes a single tractor/trailer combo.

It's here-
http://fastbale.windridge.org.uk/pr...y=18&filekey=532eab359e744ae7a0d4e6ff12e487e2
[Currently GAE free tier so might die if lots of people click]

I'd be interested in feedback if anyone wants to try it as I am hoping to improve it before next season. It's a computationally difficult problem to solve exactly so currently uses clustering heuristics and straight line distances only.

Also if anyone happens to have a GPS track from collecting bales (with a single loader) that would be useful to benchmark algorithms against humans.

Cheers

 

[tr250]

How does your gps know where the bale falls off the back? I personally think that a park land sledge dropping 5s on headlands is where the most compaction and carting time could be saved

 

[ptd006]

Yeah, for now I had been assuming driver logs the location where (round) bale is ejected - length of tractor+baler (based on orientation). An app could be knocked together quickly to figure out when you stop, orientation and do the adjustment. Big square bale drop location realistically need extra hardware, say a bluetooth switch stuck on the rear and connected to your phone.

Agreed a bale accumulator is better but for smaller ops I thought this could work for near zero cost.

 

[e3120]

Fair play for the initiative. The human algorithm takes in more factors, though. e.g. trailer pointing downhill to stop rounds rolling off, path back to gate to avoid steep bits/soft spots/ulitise tramlines.

As a novelity, I did create some coverage maps when round baler door was open last year, with a view to crude visual yield mapping, but I think they're shp files.

 

[ptd006]

Thanks for the good points. The current bale clumping/clustering algorithm is tied to straight line distance and would take lots of work to change. The trailer routing could potentially incorporate OpenTopo data and if folks already have tramline GPS tracks for CTF that could be too. I guess anything that requires manual input from the driver is not worthwhile (from my simulations I guess there's only ~5 mins saved for a 50 bale field so there's not much incentive)

 

[Briantee]

What would be handy is knowing where the next 2 or 3 bales closest to you are and highlight them real time. Operator can decide which one to choose

 

[Andrew]

I’m guessing the yellow and pink lines are where the trailers run?
For compaction reasons I prefer to take the trailer to the far end of the field and turn round, then load up as you move down the field, getting a full load as you get to the gate.
We only run trailers in grass, in straw we use bale chasers.

 

[ptd006]

@Briantee - that would be easy from a GPX but could probably be done reliably with a 360 degree camera on the roof (i.e. no need to get bale locations in advance). I hadn't considered that before but did some experiments with aerial photos (YOLOv3 model if that means anything; intending to get a drone but mostly stock photos so far as I only started in late autumn). BTW coincidentally I watched your agrabot video- interesting .. and brave! My (wayyy future) idea is to use a Bobcat (skidsteer) as it can spin on the spot (easier navigation, lower compaction etc) and there's already a remote control kit. Also, the relatively small size is a kind of damage limitation. I never dreamed of retrofitting, say, a telehandler.... until I watched your video

@Andrew - thanks for the comment, yeah the yellow/pink/white (non-red) are supposed to be for the trailer. Perhaps I misunderstood but if there are no tramlines don't you simply want to minimise distance in the field to avoid compaction? (I guess 4-5t smallish tractor with loader/bale is not massively different ground pressure to a twin axle loaded trailer?)

 

[Andrew]

@Andrew - thanks for the comment, yeah the yellow/pink/white (non-red) are supposed to be for the trailer. Perhaps I misunderstood but if there are no tramlines don't you simply want to minimise distance in the field to avoid compaction? (I guess 4-5t smallish tractor with loader/bale is not massively different ground pressure to a twin axle loaded trailer?)

Thing is if you go the same way as the tramlines would go, if you make ruts you’re not then hitting those ruts every cut for the next 4 years.
Also we run 3 axles trailers so turning scrubs a lot, better to do turn before you load.


If you could make something with a 360 degree camera on the roof I’d be very interested - I’ll send you a PM later!

 

[Briantee]

Do have distance and created path to marked flag which also has a direction associated with it. Not a huge leap to pick the path to the nearest flag mark. Can also auto steer to it.

 

[ptd006]

Hey @Briantee that's cool. Re. direction/orientation of the flag/bale - I started playing with a CNN for my little robot. Just as a proof of concept I got about 100 images and tagged the orientation quadrant manually (tedious). A small holdout test (i.e. images not in the training sample) suggests it's not totally bonkers..

 

[Briantee]

If you used opencv and did a Hough transform, you could define it to 3 shapes. Circle - the way to pick it up, square you would be 90 degrees off or ellipse you are at like a 45 degree. Then light/dark etc would not matter and only need a contrast

 

[ptd006]

thanks for the suggestion, I actually naively tried doing that first to calculate the angle from fitEllipse. HoughCircles works ok when you're dead on. However I wasn't getting good results across the sample. Possibly my preprocessing steps (-grayscale -Gaussian blur) need tuning. I think it could be used as an extra verification method though. I never got good results for the other orientations.