For all practical purposes, it is impossible in the current state of security/surveillance video footage. The example described in the paper in your link is us

For all practical purposes, it is impossible in the current state of security/surveillance video footage. The example described in the paper in your link is using cameras that are setup to provide very high resolution images of people. It would be like using full-frame portrait images for face recognition, and then expecting that to translate to real-world scenarios where you might only have 20 pixels on a face, and the person is off-axis to the camera. Gait detection has been discussed for a while, and may definitely be a thing one day, but right now we are barely at the point where pose estimation is a thing in security video. Very far from being able to do high precision pose recognition and sampling over successive frames to model something that would qualify as "gait".