What it does ¶
ALVisionRecognition is a vision module in which the robot tries to recognize different pictures, objects sides or even locations learned previously.
How it works ¶
This module is based on the recognition of visual key points and is only intended to recognize specific objects that have been learned previously.
The learning process is described in the Choregraphe Video monitor documentation: Teaching NAO to recognize objects. With few minutes experience, the robot should be able to learn any new thing in less than 30s.
Like for all other extractor modules, recognition results are placed in the ALMemory. You can open the web page of your robot with your favorite browser, go to Advanced > Memory and look for PictureDetected in the search field.
When something is recognized, you see an ALValue (a series of fields in brackets) organized as explained here:
The “PictureDetected” key is organized as follows:
[ TimeStamp, PictureInfo[N] ]
with as many PictureInfo tags as things currently recognized.
This field is the time stamp of the image that was used to perform the detection.
TimeStamp = [ TimeStamp_Seconds, Timestamp_Microseconds ]
For each detected picture, we have one PictureInfo field:
PictureInfo = [ Label[N], MatchedKeypoints, Ratio, BoundaryPoint[N] ]
- Label: organized names given to the picture (e.g. [“cover”, “my book”], or [“fridge corner”, “kitchen”, “my flat”]).
- MatchedKeypoints is the number of keypoints retrieved in the current frame for the object.
- Ratio is the number of keypoints retrieved in the current frame for the object divided by the number of keypoints found during the learning stage of the object.
- BoundaryPoint is a list of points coordinates in angle values (radian) representing the reprojection in the current image of the boundaries selected during the learning stage.
BoundaryPoint = [ x, y ]
Performances and Limitations ¶
The recognition process is partially robust to distance (down to half and up to twice the distance used for learning), angles (up to 50° inclination for something learned facing the camera), light conditions and rotation. In addition, learned objects can be partially hidden for the recognition stage.
For a better performance on the robot, the database sent on the robot contains only essential information for the detection and not all additional data stored in the computer database.
- This module is based on the recognition of key points and not of the external shape of the objects, so it can’t recognize untextured objects.
- Currently it is not designed for recognizing objects classes (e.g. a cookie box) but objects instances (that cookie box).
- Currently every detected keypoint in the current image is matched with only one learned keypoint in the database. If scores for choosing between two objects are too close, the keypoint will not be associated to any of them. As currently the algorithm doesn’t vote for several objects, learning twice the same area of an object will reduce its detection rate.