This is a face tracking system built around the
face_recognition library, which is built around
dlib. The point of it is to be able to easily
add new people to a library of detectable people
by providing an interactive live video feed from
which screenshots can be taken with known
detections, and then take actions based on these,
for example marking live video and controlling
door locks. The detected faces are saved to a
directory specified by the name of the
individual. Run this:
python capture.py Colin
This will make a Colin directory under /images/
that contains images captured from the live
video feed if you press the c key while there is
an affirmative face identification.
There is another utility, preview.py that simply
reads all of the images from /images/ and
associates the images with the directory name as a
label (e.g. the Colin directory). It validates
the detection to the extent of making sure that a
face is detected. If not, the training example
is deleted. All detected faces in the live
video are compared against the known faces and
labeled. Invoke it like this:
python preview.py
The last utility, door.py is an interface with
Kisi door locks. This does not have a display,
so you won't physically see who is in the frame. In
this application, if one and only one known face is
identified for a number of consecutive frames, the
door is unlocked. Regardless of other logging that
may or may not occur, you probably want to log what
is being done. This can be done like this:
python door.py > log.txt 2>&1
Obviously you need to have Kisi door locks to
use this. Nothing fancy going on here with
regard to stopping and starting, it just
runs in your terminal until you quit. To get the
desired automation around it you can just control
the launch of the program. You might have to change
the KeyboardInterrupt condition for quitting if
the situation warrants it.
You will need to copy the example configuration,
config.yaml.example to a real configuration,
config.yaml. If you are using Kisi door locks
you will need to fill in credentials and which
door to unlock. The rest of the configurations
shouldn't need messing with.
One thing to note about the implementation is the
abstraction for image input. The Capture class
requires an object to fetch images. One such class
is the LocalCamera class that is an interface to
the OpenCV VideoCapture interface for local
video devices. You can easily add new functionality
for things like IP cameras by simply implementing
the image fetch in a new class.