This 2012 project explored HAAR cascade detection, the pre-deep-learning workhorse behind many early real-time face detectors. The method scans an image at multiple scales and evaluates simple rectangular contrast features over each candidate window.
The reason HAAR cascades were fast enough for practical use is the integral image. Once the integral image is computed, sums over rectangular regions can be calculated very cheaply. A cascade classifier then rejects obvious non-face windows early and spends more computation only on promising regions.
What the video shows
The demo is a classical computer vision pipeline in action: grayscale image processing, scale-space scanning, trained cascade stages, and bounding boxes over detected features or faces.