NAOqi Vision - Overview | API | Tutorials | ALVideoDevice - Advanced
See also
The ALVideoDevice module is in charge of providing images from the video source (e.g. NAO’s cameras, Simulator, an ARV video file) to all modules processing them (e.g. ALFaceDetection, ALVisionRecognition) in an efficient way.
In addition, ALVideoDevice module is in charge of recording ARV timestamped video files when requested by the user. You will find more details on ARV and on how to change the video source in Recording and replaying .arv video files.
As the ALVideoDevice module knows at every moment the list of modules needing images and their requirements, it is able to set the minimum configuration for the video device that will fulfill the needs of all the modules while saving processing resources. Details on the internal mechanism and on how to switch between video sources can be found in Details on ALVideoDevice.
Performances
Best performance on NAO is achieved when processing directly the native colorspace provided by the camera: YUV422. For other colorspaces, a conversion is done by the ALVideoDevice module, so processing times are ranked as follow for the main colorspaces:
YUV422 < Yuv < YUV < RGB/BGR < HSY (close to HSV/HSL colorspaces in term of functionality, but faster to process).
YUV colorspace is convenient as it is more powerful than RGB:
luminance is in the Y chanel so don’t need to average the three RGB layers to get a grey level image,
chrominance is purely embedded in both U and V chanels, so it’s easier to work on colours compared to RGB for which luminance and chrominance are correlated.
Providing uncorrelated luminance and chrominance channels, it brings almost the same advantages than HSV/HSL without spending as much CPU time for processing it.
Limitations
Currently on ATOM CPU, requesting more than 5fps 1280x960 HD images remotely (getImageRemote) is bringing some frame drops. So we recommend not to go over 5fps HD images if you want to get them through the network. If all modules processing HD images are calling them locally (getImageLocal), there is no such limitation.
Here are the observed framerates when requesting uncompressed YUV422 images on NAO v4 (*).
| local | Gb Ethernet | 100Mb Ethernet | WiFi g | |
|---|---|---|---|---|
| 160x120 (QQVGA) | 30fps | 30fps | 30fps | 30fps |
| 320x240 (QVGA) | 30fps | 30fps | 30fps | 11fps |
| 640x480 (VGA) | 30fps | 30fps | 12fps | 2.5fps |
| 1280x960 (4VGA) | 29fps | 10fps | 3fps | 0.5fps |
On GEODE CPU, for some reason, processing images provided directly from the V4L2 driver (data are in kernel space buffers) and storing results in userland takes longer than it should, even when doing just a memcpy. A workaround is used by first copying manually the images in a userland buffer in a most efficient way than the memcpy (yes, in this particular situation memcpy is totally unefficient) and after processing this userland buffer.
If you want to develop your own vision module in C++, you might be interested in OpenCV.
It is a large and powerful library dedicated to vision processing. On NAO we are currently using OpenCV2.1 and plan to switch to 2.3 after extensive tests.
For further details about OpenCV, see http://opencv.willowgarage.com/wiki/.
See also the tutorial: Using OpenCV.
A free player for YUV422 images and videos.
NAO’s camera provides natively YUV422 color images. These images are used to create arv video files or ari image files that, in addition, are timestamped. As this format is not common, it can’t usually be opened by common viewers. This is where pYUV steps in.
pYUV is a multiplatform (Windows, Mac and Linux) freeware and can be downloaded from its main page: http://dsplab.diei.unipg.it/pyuv_raw_video_sequence_player_original_one or other sites.
This software can display several non common image and video formats. For arv and ari files, open pYUV, drop the file on, and set the different fields accordingly to what follows:
Main settings
Extra settings