RoadNarrows CogniBoost
The RoadNarrows CogniBoost module provides an interface to the CogniBoost vision recognition board. Powered by
the CogniMem CM-1K neural network chip by General-Vision, visual objects are readily detected and
categorized from patterns previously trained. The recognized visual categories are then fed
into RoboRealm to guide goal decisions, object localization, platform movement and navigation, and
environment manipulation.
The CogniBoost RoboRealm module can be used to train patterns on the board. Recognized object categories are fed back into RoboRealm
as variables that can be used by other modules or transported to external application/devices.
The module sends images to be recognized to the CogniBoost board. This allows images to be prefiltered, oriented, etc. using
all of RoboRealms other modules in order to improve recognition results or focus recognition on specific image aspects.
Interface
Instructions
1. COM Port - Select the appropriate COM port that is connected to the
CogniBoost board. If the appropriate COM port does not appear ensure
that all connections are made and that the CogniBoost board is turned
on. Close the module and reopen to see if the COM port appears in the
dropdown list.
2. Train Pattern - Pressing this button will cause CogniBoost
to train on the currently viewed image. This image is stored in
CogniBoost and will immediately start detecting the presence of
that object. You should see the Recognition values being populated
as soon as an object is trained.
3. Pattern Width & Height - The CogniBoost works with 256 'pixel' values. In order to create
a two dimensional image a width and height of a pattern is required. If you find that your patterns
are better detected in a more vertical or horizontal shape as apposed to the default square pattern
you can change the width and height appropriately. Note, however, when you change these parameters
you WILL lose all currently trained patterns.
4. Upload Patterns To CogniBoost - uploads images from the PC to the
CogniBoost board. Note that these images MUST have the same dimensions
as what is currently configured in the CogniBoost board based on the ROI
settings. See Pattern Width and Pattern Height for what
this size needs to be. Note that images can be of any recognized image
format but it is recommended to use a non-lossy image compression
technique such as gif or png in order to store images. This ensures
that the trained pattern is exactly the same as the stored image.
Note that you may wish to "Forget All Patterns" before uploading
to ensure that only those images uploaded are contained with CogniBoost.
5. Forget All Patterns - Causes CogniBoost to forget all trained
images and set its pattern counter to zero. This should be used
before training on valid images to clear out any experimental
patterns that you may have trained on. This should also be used
prior to uploading patterns to CogniBoost to clear out existing
patterns. If you do NOT use this function prior to uploading
patterns new patterns will be added to the current list within CogniBoost
possibly duplicating existing patterns.
6. ROI - To focus on a specific area of an image you can change the ROI
(seen as a red square in the live image) to change what part of the image is focused on. Often
it is not possible to align the camera on the expected object position
such that the object completely fills the field of view. When this
happens adjusting the ROI to ignore non-object areas of the image can
help improve object recognition. Note that the ROI area is scaled
into the pattern width and height before sending it to the CogniBoost.
Note that changing these parameters may invalidate
image patterns that you have trained on if the aspect ratio changes from what
was trained. This is due to the scaling requirement to map the ROI into the pattern size.
X Start - the x (horizontal) start coordinate
Y Start - the y (vertical) start coordinate
X End - the x (horizontal) end coordinate
Y End - the y (vertical) end coordinate
Arrows - use the provided arrows to move the ROI as a whole
around the image. Note that moving the entire ROI will NOT affect
the CogniBoost image size and therefore will not affect the recognition of
previously trained patterns.
Zoom - use the provided zoom and un-zoom buttons to change
the field of view of the ROI. As the zoom will preserve the
image aspect ratio your previous image patterns will work.
7. Options
Classifier - K-Nearest Neighbor or Radial Basis Function. The KNN classifier discards the relationship between the
distance and influence field of a neuron. As a consequence, all the neurons fire
and their distance and category can be read in sequence per increasing
order of distance. The RBF classifier uses the Influence Field of the neurons at the time of
the recognition. A neuron fires only if the distance calculated between the
input vector and its vector in memory is less than its influence field.
Norm - Determines how to calculate the distance between the reference pattern or prototype
stored in a neuron and an input pattern. (Manhattan distance versus Euclidian distance))
Minimum Influence Field - Defines the "Uncertainty" zones in the decision space.
Maximum Influence Field - The default value of a newly committed neuron when no other neuron recognizes the vector to learn
8. Reset CogniBoost - You can reset the CogniBoost board
by pressing the "Reset CogniBoost" button. This will cause
CogniBoost to forget all patterns in volatile memory and reload all its
settings.
9.Save To EEPROM/Load From EEPROM - When training
or uploading images or changing camera settings all these
changes are stored in volatile memory on the CogniBoost board.
This means that if you turn off the CogniBoost board all these
settings are lost. To save these settings such that they are
used again once the CogniBoost board is turned on you need
to "Save to EEPROM". Likewise, if you wish to reload all saved
settings (to erase all current settings stored in memory) you
can press the "Load From EEPROM" button.
Variables
COGNIBOOST_NAME_X - The category name associated with the recognized
category. Note that this is a RoboRealm specific association and is
NOT present on the CogniBoost board. Up to four categories can be
recorded at one time. Note that they are ordered with regards to
distance with the first being the closest best match.
COGNIBOOST_CATEGORY_X - The numerical category CogniBoost has assigned
to the trained pattern. When the pattern is presented again to CogniBoost
this category number will become available as a recognition
identification code that the object is present.
COGNIBOOST_DISTANCE_X - A measure of recognition confidence or the "distance"
of the current recognized pattern to its associated pattern stored in
CogniBoost.
For more information
RoadNarrows CogniBoost
CogniMem Decision Space Mapping
|
