Best Cameras for Computer Vision

[truell20]

I was wondering what are the more-popular/best-regarded cameras in the FRC community for computer vision.

My team is currently using two Microsoft Lifecams. The Lifecam maxes out at 30fps, but I don't see us needing more than that. It seems to have autofocus which is nice.

[ProfessorAlekM]

Pretty much any logitech camera is great for openCV, the most common one being the C270. You can get wider-angle cameras but they are nearly $100, so it's usually easier to just glue a phone camera lens on for a wider FOV.

[anushad]

The Pixy is really popular, and it has a bunch of code and libraries available online.

[Adnewhouse]

Our team is rather fond of the Playstation Eye camera that use to be an accessory for the PS3. It costs less than ten dollars and does 60fps at 480p (also 120fps at 320p, but we've never tried that). This is pretty ideal for vision tracking. The image sensor has ridiculously large photosites, so there is very little image noise to worry about. The only downside is that they don't work on Windows systems very well without a third party driver. On the raspberry pi, it works just fine with OpenCV. You can get them on ebay because they are not made any more.

[Commandman7]

Last season we had great success with the Logitech C920 / 922 / 922x. These cameras are great because they have built in H.264 Encoding, which really took a lot of stress off our vision processing hardware. In order to use the built in encoding, I had to dig up some ancient software (released 1999) in order to run it, but I'm sure there's a better solution.

[Jaci]

The Kinect's IR sensor has never let us down. If you have a Jetson board available, you can use the Kinect v2 to get both IR and RGB at the same time (Kinect v2 requires USB 3.0, while Kinect v1 doesn't). Alternatively, you can use the Depth stream as a cheap-mans IR sensor.

[cmastudios]

We have been able to use the Genius 120-degree (WideCam F100) for multiple years now, which has allowed us to see a greater portion of the field. We were also able to convert the camera to an IR camera by removing the glass filter inside the lens and inserting a clipping of floppy disk between the CMOS and the lens assembly. In 2014, with three of these, we were able to track the entire field, which I wrote about here. I recommend it.

[AutomationFTW]

Microsoft LifeCam HD-3000

https://www.amazon.com/Microsoft-Lif.../dp/B008ZVRAQS

[ahartnet]

Originally Posted by Commandman7

Last season we had great success with the Logitech C920 / 922 / 922x. These cameras are great because they have built in H.264 Encoding, which really took a lot of stress off our vision processing hardware. In order to use the built in encoding, I had to dig up some ancient software (released 1999) in order to run it, but I'm sure there's a better solution.

What was the ancient software you had to dig up?

Some of us remember partying like it's 1999 by the way....

[Justin Buist]

Originally Posted by Adnewhouse

Our team is rather fond of the Playstation Eye camera that use to be an accessory for the PS3. It costs less than ten dollars and does 60fps at 480p (also 120fps at 320p, but we've never tried that). This is pretty ideal for vision tracking. The image sensor has ridiculously large photosites, so there is very little image noise to worry about. The only downside is that they don't work on Windows systems very well without a third party driver. On the raspberry pi, it works just fine with OpenCV. You can get them on ebay because they are not made any more.

It also doesn't work with the roboRio, at least not out of the box. We've tinkered with them a bit this year. The most promising thing, to me, about them is that you can pull the IR filter out and replace it with Mylar (or the inside of an old floppy disk) to block most visible light but let IR through.

The pi works with the camera just fine though. Had two students "struggle" with getting it to work on their Windows and Mac machines but popped it into the pi and it was just fine. Not sure why the roboRio is crippled with only supporting the "uvcvideo" kernel module and none of the other USB webcams. I could understand it if they really nailed that one down but the Wpilib documentation and interaction with v4l2 is so bad even with that limited driver support that it's almost like they didn't try and forced people to get creative.

[Adnewhouse]

Originally Posted by Justin Buist

It also doesn't work with the roboRio, at least not out of the box. We've tinkered with them a bit this year. The most promising thing, to me, about them is that you can pull the IR filter out and replace it with Mylar (or the inside of an old floppy disk) to block most visible light but let IR through.

The pi works with the camera just fine though. Had two students "struggle" with getting it to work on their Windows and Mac machines but popped it into the pi and it was just fine. Not sure why the roboRio is crippled with only supporting the "uvcvideo" kernel module and none of the other USB webcams. I could understand it if they really nailed that one down but the Wpilib documentation and interaction with v4l2 is so bad even with that limited driver support that it's almost like they didn't try and forced people to get creative.

That's a real bummer. The camera is really great on a Raspberry Pi. The sensor is superb and that IR filter hack sounds really cool. I'm guessing recompiling the kernel on the roboRio to support the camera is a no-go because of national instrument's special sauce. We have at least 4 of them floating around because they're so cheap.

[euhlmann]

Last year we used Logitech c930e for its wide field of view

[Justin Buist]

Originally Posted by Adnewhouse

I'm guessing recompiling the kernel on the roboRio to support the camera is a no-go because of national instrument's special sauce. We have at least 4 of them floating around because they're so cheap.

It's all GPL code for the kernel, so there's no secret to it. I forget what kernel version the roboRio is running this year but it has to be one of these branches: https://github.com/ni/linux/tree/nilrt_pub/16.0/4.1

I think it's the 4.1 one but not sure. It's do-able, but tossing a pi in the mix and using NetworkTables to toss relevant vision data back to the roboRio sounds less risky.

Just my two cents.

[sparkytwd]

Originally Posted by Adnewhouse

That's a real bummer. The camera is really great on a Raspberry Pi. The sensor is superb and that IR filter hack sounds really cool. I'm guessing recompiling the kernel on the roboRio to support the camera is a no-go because of national instrument's special sauce. We have at least 4 of them floating around because they're so cheap.

They're not UVC based cameras so they need special drivers in the Linux kernel. I'm not sure if you can add modules on the Roborio, but in theory it would be possible to get it working with the kernel source and a crosscompiler.

[troy_dietz]

Shameless plug: If you have access to a 3d printer, I've made a minimalist mount for the LifeCam-3000 here: http://www.thingiverse.com/thing:1494829

The Microsoft LifeCam HD‑3000 is one of the most common due to its small size and low price, as well as being available from FIRSTChoice for several years.

Axis network cameras are also popular due to their many features and adjustable options.