| AI CAMERAS
Annual Showcase 2020 | Intertraffic World 151
can take days, weeks, or even months,
depending on the amount of data
being used. During this training,
the AI is repeatedly shown a set of
images that it is asked to categorize.
In the case of traffic applications,
the subject of these images would
be vehicles and the AI is tested with
a series of images where some do and
some do not contain vehicles. The AI
keeps running through these tests as
it gets more and more familiar with
each image and how they relate.
The AI has finished training
once the tests results show that the
set of images is consistently being
categorized with the correct results.
At this point, it is time to take the AI
out of training and begin working
through new images.
Using inference, the AI can deduce
if a vehicle is in an image because its
neural network was developed to
recognize what a vehicle looks like.
This is similar to how a person is able
to recognize a new car after having
already seen other cars throughout
their lifetime.
Embedding the AI
Normally, a camera just captures
an image that is then transferred
and used elsewhere. The whole point
of developing an AI to recognize
different vehicles is to put that
capability within the camera itself.
Instead of looking through all the
images that would be sent over the
network, the camera sorts through
each frame and only keeps the
images with a vehicle and a clear
license plate. By tracking a vehicle
through each frame, the camera
can compare and choose the image
with the most in focus license plate
characters. By paying attention to
only the information that matters
most, the Ls245R camera eliminates
the need for the large bandwidth
normally needed for sending image
files over a network.
The built-in processing power
of a camera with AI also enables a
streamlined workflow. The image
processing is done without the need
for external components to the vision
system. The device can use this
image recognition to identify
whether any of the images coming
in actually contain a vehicle, track
those to find the best result, and
remove any that do not.
After the tracker finds the
most in-focus image, the camera
then narrows the region of interest
(ROI) on the image and finds what
is actually important. For ALPR, it
begins with ensuring the camera
spots a vehicle because the rest of
the image is irrelevant. After that
the ROI narrows even further to
inspect just the license plate because
the rest of the car does not provide
any useful information.
Then, the ALPR efficiently works
through a much smaller area and
extracts the license plate data that
was originally a piece of a large
image file. Finally, the camera sends
over this text data without flooding
the network full of large files.
All-in-one solution
By utilizing the processing and
imaging power of an AI traffic
camera like the Ls245R, the result is
an all- in-one-solution for affordable
traffic imaging. Cameras no longer
need to feed all their image data in
order to be processed elsewhere, but
instead only useful information will
be extracted and processed using
the camera itself. By having a device
that captures images and provides
analysis, it enables users to focus
on other factors that can improve
the application. n
Above: The Teledyne
Lumenera AI camera
can extract license
plate data and
convert it into text
data to save space
on the back office
network