Welcome, Commercial Drone Pilots!
Join our growing community today!
Sign up

Looking for a piece of software...

JoeDimwit

Active Member
Joined
Mar 29, 2019
Messages
42
Reaction score
11
Age
54
Location
Pontiac, MI
For the purpose of this conversation, let’s assume that I work for a large, well known automobile manufacturer, and flying a drone is only part of my job responsibilities. Let’s also assume that the drone program is still in its infancy (less than 3 months old), and we are looking for ways to expand our usefulness and value.

Given all that, I’m wondering if there is a way I can fly a path and take pictures over our storage lots, then use software to stitch those pictures together and get an accurate count of the vehicles in that lot. Bonus points if that can be broken down by colors, and even more bonus points if it can break down by color and sunroof.

Any ideas? Outside the obvious that I’m likely out of my ever-lovin’ mind?
 
The type of craft/camera required might not be of too much consequence if I can sell them on the idea. ?
 
I don’t have anything as yet. I am looking for something I can take to them as an excuse to get more flight time. I’m honestly not even sure they would be interested.

Just brain storming with you then ... a coworker of mine at the U of MN created a camera based system to count empty parking spaces at freeway rest stops. That way an electronic sign could be posted in advance that would say something like "Next Rest Stop 2 Miles, 8 Truck Parking Spots Available". For their system they mounted permanent cameras up somehow on the light poles and then used some computer graphics hackery/wizardry to figure out if a parking spot was empty or full. I don't know if or how much of that system is proprietary vs. open, but I could ask some questions.

What I was thinking about today on my drive home is that I needed a drone about 15' up looking down the roadway, figuring out which lane will be fasted up ahead. I think you could charge money for that ... at least until everyone owns the system and the whole lane switches over simultaneously and then switches back simultaneously, and repeats. Then you have to make the system lie a little bit and de-optimize everyone so that the lanes are better balanced so everyone gets home in a fair time. Then you could start selling the special version of the system that doesn't slightly lie to balance the lanes and charge even more money. You came to the right place here for business advice. I have no end to ideas! :)

Good luck. I think many of us are still in that awkward phase where hype and reality are trying to sort themselves out. There are lots of great applications and success stories, but also lots of ideas that didn't work, or maybe worked but weren't cost effective.

Ohhh ... I just found a link to a paper describing the computer vision / empty parking space system: https://rosap.ntl.bts.gov/view/dot/29437/dot_29437_DS1.pdf
 
Change detection software. That's what our customers call it.

We've flown parking lots, cities, etc. for change detection studies. We only did the acquisition though (manned aircraft).

Search for change detection software.
 
For the purpose of this conversation, let’s assume that I work for a large, well known automobile manufacturer, and flying a drone is only part of my job responsibilities. Let’s also assume that the drone program is still in its infancy (less than 3 months old), and we are looking for ways to expand our usefulness and value.

Given all that, I’m wondering if there is a way I can fly a path and take pictures over our storage lots, then use software to stitch those pictures together and get an accurate count of the vehicles in that lot. Bonus points if that can be broken down by colors, and even more bonus points if it can break down by color and sunroof.

Any ideas? Outside the obvious that I’m likely out of my ever-lovin’ mind?

It is certainly possible to do something like you describe though probably not with a single software package. The orthomosaic stitching would be one operation and then you would need some software and code to query by RGB values or something similar.
 
  • Like
Reactions: clolsonus
Pretty much all the available mapping software programs would do what you need, or you could send your images to one of several cloud based services to compile it.
 
Pretty much all the available mapping software programs would do what you need, or you could send your images to one of several cloud based services to compile it.

I know the mapping programs will stitch the pictures together, but they can give me object counts too?
 
I know the mapping programs will stitch the pictures together, but they can give me object counts too?
Pix4D, Correlator and a lot of others will stitch the images together.They do not do complex analysis. You'll need another software package to do that for you. We use ESRI's ArcGIS but there are others out there that will work equally well.
 
I know the mapping programs will stitch the pictures together, but they can give me object counts too?

Two quick comments.

1. I am imagining a huge car lot. If you have a ginormous lot filled with 1000's of identical Tucker 48's all parked in neat rows, then the dominate features in the images will repeat over and over again 1000's of times. I would be curious if popular stitching software handles this easily or if it could throw some challenges. You might have to play around with altitude/field of view to get consistent results?

2. I don't know ArcGIS, but when I've looked over the shoulder of other computer vision people ... typically I see that they combine various primitive operations in clever ways to get the answers they want. So there are things like edge detection, blob detection (by color or intensity), thresholding, statistics, and many more things beyond my small slice of experience. If the map stitch succeeds and you create a dense mesh of your area, maybe you could do some sort of blob detection by elevation. Anyway, if there is an existing solution then great, but if not, you might be able to get creative and piece together some computer vision or computer algorithm tricks to come up with an answer.

I know I asked before if you had a dataset available and you said not at this point, but if you would be able to go out and fly a car lot and get some sample imagery, and share that (even privately) I think seeing and working with the actual images might lead to some more concrete ideas or suggestions (or could point to where the process would need to be improved in order to get reliable results.)

Thanks,

Curt.
 
I know the mapping programs will stitch the pictures together, but they can give me object counts too?

Bentley Systems Context Capture and Harris Geospatial ENVI Deep Learning Module - both have deep learning/machine learning aspects. ENVI works with ESRI.

You'd have to train them on the objects, but those are the closest (commercially available) we've seen to accomplishing your task (we are only image acquisition and pass them onto folks that use these programs).

In our use these are referred to as change detection solutions - detecting/counting cars in parking lots/freeways during different times of day (these were for traffic studies and business park planning), looking a ground movement (landslides, encroaching on rights of way/pipelines), counting the ebb and flow of containers in ports, etc.

I have 94 images (10GB - Phase One 100MP aerial images @ 3cm/GSD) of a "harbor with boats and two parking lots" if that helps with your testing. Let me know and I'll send you a link.
 
Two quick comments.

1. I am imagining a huge car lot. If you have a ginormous lot filled with 1000's of identical Tucker 48's all parked in neat rows, then the dominate features in the images will repeat over and over again 1000's of times. I would be curious if popular stitching software handles this easily or if it could throw some challenges. You might have to play around with altitude/field of view to get consistent results?
There should normally be enough variation in the color mix of the vehicles that stitching shouldn’t be a problem if there is sufficient overlap in the photo set.

I can’t get a data set at present, because my company isn’t going to give me permission to fly over their lot without there being a value in it for them. But, if you can imagine a large asphalt lot with vehicles parked bumper to bumper, in rows barely wide enough to open the drivers door far enough to squeeze into the vehicle... you’d be on the right track.

Between the various lots they park vehicles in, there are probably somewhere between 5,000-10,000 vehicles waiting for delivery.
 
There are apps in DroneDeploy that count plants. Why not try it on cars?

Not sure why permission would be needed to fly over what is effectively a parking lot. Just get Verify or other on-demand insurance for the flight if you don't already have an aviation liability insurance policy.

13371558195841566.png1558195841566.png
 
Lensupthere has great equipment, PhaseOne using Bentley systems is top shelf; although, I am not sure you need it...less expensive equipment could be used with the same 3cm/pixel or less; however, the drone would be flown at a lower altitude....I understand the 1000s of cars, do you know the square area? 100 acres can be flown in less than 25 minutes with sufficient overlap and less than 3cm/pixel for stitching... the trick is using AI to "detect" either similarities/differences...
Is your goal to "identify all blue cars" or is it "how many blue cars have moved/relocated since last flight"?

We might be able to assist you, but will probably want to have to take it off line... PM me if you are interested.

Stephen
 
  • Like
Reactions: Lensupthere
I understand the 1000s of cars, do you know the square area?

~700 cars per 5 Acres. We also map areas in Benicia, CA where cars are offloaded after transit from Asia. We have some images of the area available if needed. Address/location is: 2850 Bayshore Rd, Benicia, CA 94510

Google Earth has recent pix - you can draw your own boundaries and counts as well.

If you have an alternative software solution, we're keen on learning about them.
 
Hi, to count cars on the storage lot it is not necessary to have orthophoto stitching in your processing conveyor.
It is not rocket science to detect cars in video stream and count them during flight. Knowing coordinates of the drone and attitude of the camera you can calculate coordinates of each detected car to avoid counting them twice.
I'm CTO of SPH Engineering, developer of UgCS - we did similar projects, but not to count cars.
 
Here is description of car counting project - Parking Lot Vehicle Detection Using Deep Learning

But I'm sure that for particular task of car's counting stitching of the hundreds of images is excess step. It adds additional complexity, requires lot of time for processing and can influence errors in some cases.

With on-line detection from video stream it is possible to generate list of detected cars with their coordinates during the flight and to have car's amount immediately.

We did a very simple test using video made near our office and YOLO classifier for cars. It works not very good (because trained to recognize cars not from above), but works. We will try to use another pre-trained classifier when we will have a bit of time.
 

Members online

No members online now.

Forum statistics

Threads
4,277
Messages
37,605
Members
5,969
Latest member
KC5JIM