Category: Motion Capture

Configuring Eye cameras here.

Playing with the evaluation version of this.

Malcolm and I spent part of Friday experimenting with an evaluation version of a commercial motion capture package. I now have six PS3 Eye cameras and my VR machine has enough USB-3 ports and controllers installed to run them.

Our initial setup had three or four cameras attached and used the room lighting (pretty bright, ceiling mounted LED lights) for illumination. Initial results weren’t great but over the rest of the day we learned a few things.

• Need a more contrasty background and a less cluttered background. The bookcases behind the area we were using were better when covered with a piece of white fabric. The tan rug on the floor was less of a problem when the model put on black socks.
• Hands really aren’t handled by the package. No big surprise here as hands and fingers are rather small targets for these cameras.
• More lighting is better. I added two diffused studio lights I have around and a high intensity three light halogen light bar and things became more precise.
• A larger and more diffuse calibration target seemed to work better.
• Sliding the calibration beacon along the floor with periodic stops seemed to work better than touching it to the floor. This makes all floor level reference spots about equal (when I was touching it down, it took a little work to make sure we had the lowest spot in each arc).
• Aligning the reference person with the human figure in the images at the start helped quite a bit. The tool didn’t seem to do a very good job of this without manual help.

By the end of the session, we seemed to be getting a pretty good capture of arms and legs. Feet could still be a bit twitchy.

Malcolm is going to look at 3D printing mounting clips to attach the PS3 Eye cameras to light stands for more stability.

I ordered a couple of spare cameras to ensure that we don’t come up short if any of them fail and a couple of PCIe USB-3 cards to supplement USB controller availability.

Overall things turned out pretty well and I think we learned a bit more about making motion capture without dedicated beacons work decently. The price of the package is high enough that even a short time license would need us to have some substantial amount of motion capture to get done in order to make things make sense.

I proven that the cameras can run on Ubuntu and the RPi. I found a page with classic Unix/Linux style install instructions. I’ll be working on getting this set up on my biggest RPi machine and take a look at building code to red from multiple cameras and stream the data to a host. If I can run two cameras on a single RPi then I should be set.

I’ll probably also look at doing something similar for the RPi cameras on the RPi-2 machines. That might add a couple of additional cameras to my set.

I’ll then move on to building a simple LED beacon and look at some simple camera calibration code on an appropriate host.

Trying to build the OpenCV package on one of my RPi 3 machines. I think I’m running into heat issues. I’ve switched to a board that I can keep open and has a heat sink on it. Hoping that may be enough. I’ve also dropped the build scripts onto github to make them broadly available.

Pulling packages on this machine now.

Now I’ve got a fan blowing. I see from some web pages that the RPi is supposed to warn and throttle when the temperature spikes…didn’t see that with my black and silver RPi 3…it seemed to halt completely after a short span. Keeping this one cool up front and we’ll see how this goes.

Interesting…it looks as if the scipy build is eating all available physical memory on the RPi board (882 MB of 923 total). Nothing moving on the machine…not even the mouse cursor.

Ah, reference here to bumping the swap file for the RPi.

Monday morning dawns and it appears that I have a raspberry pi that is loaded up with an installed build of OpenCV. No time this morning to test this but tonight I’ll run through a few simple tests and then probably run the same process on one of my Intel NUC machines (should go faster and easier) to get a decently powerful system up and working with the same version.

Yesterday my five additional $6.00 PS3Eye 60 FPS cameras arrived. I’ve pulled my RPi machines and my two Ubuntu based NUC machines out and started checking out the cameras and making sure the systems are up-to-date.

The NUCs generally get more use and thus needed less updating. I found that using cheese I was Immediately able to run the cameras on both the Ubuntu NUC machines and the RPi boxes. Even the older RPi-2 machines ran the cameras…I’m not sure that the performance on those will be sufficient to make them useful though. They do both have RPi internal cameras connected though so that may be of interest. I’ve got to look at how to access those cameras from C++ code and see what sort of performance they have (and test them out to see what they can do).

Here are the cameras (the one sample I bought for testing and the five others that just came in). I may pick up a couple of spares to make sure that I have what I need in the longer run…at $6.00 each that isn’t a big deal). The beer pong balls look like the perfect size for a light diffuser on an LED…I’ll have to rig up a few LED/resistor/battery sets to try that out sometime soon.

On Sunday I’ll look at getting programmatic control of these cameras. If I can acquire images from two cameras at the same time and stream the data out over a socket, I’ll be ready to rock with these things.

Still to be managed is getting the cameras rigged to mount on light stands and getting three more light stands with ball heads to set them on. If the RPi-2 boards look interesting with their integral cameras then I may need some additional mounting options, but initially three to six cameras should be a good start.

I’ve installed the filmbox SDK on my home machines so that I am in a position to play with programmatic manipulation of *.fbx files. This feeds into the motion capture experimentation I’m doing and should be generally useful at times.

I also dug out my three RPi 3+ machines last night and got one of them booted up and connected to a PS3Eye camera…without actually taking images…just observing the power light go on and no other negative effects. I’ll likely move further over the weekend to try getting actual image capture working there.

I did find a github project that claims to be able to run one of these 60 FPS cameras from an RPi. This project appears to be motion sensing related, but I expect it can act as a source of sample code for my longer term purposes…

My sample cameras came in. I ordered two cheap webcams, a Sony PlayStation 3 Eye Camera and a Fosmon USB 6 LED 1.2 Megapixel USB PC Webcam as these were both under ten dollars and looked potentially interesting.

The Fosmon camera looks too slow and low on image quality to be useful for my purposes here so I’m shelving that one. The PS3 eye camera is challenging to interface to a PC, but at $6.00 per camera, I’m inclined to spend some effort here. There do appear to be Linux drivers out there and I’m going to look at a RPi mediated option with these (I’m willing to invest $30.00 for five more cameras to give this a try).

The other option that was presented as viable are Logitech C922x Pro Stream Webcams. I have two Logitech C930e cameras currently…these won’t run 60 FPS but have comparable image quality and a wider field of view. At $70.00 each, I won’t be picking up six c922x cameras any time soon, but I may purchase on or two if all goes well to see what they are capable of. They’re almost certainly better cameras than the PS3 Eye devices, but rather pricey.

I’m also picking up some CR2023 batteries and battery holders along with a box of LEDs and some 19mm ‘ping pong balls’ to try as diffusers.

This is all an experiment, so it may come to nothing more than an interesting diversion. The commercial motion capture options are far too expensive to be viable as a hobby thing and I’m expecting to have some fun putting this together and seeing where it can go.

I would be particularly happy to see an RPi solution work out as having one pi board for each pair of cameras (or even for each camera at a stretch) with an Ethernet back-haul to a full fledged PC for processing would be a flexible and relatively affordable solution with pretty good scalability to more cameras as well.