Automated Face Recognition's (AFR) popularity as a technology has been on a constant rise for the past few decades with new applications and usecases popping up every few years. A recent example of a possible application of this technology can be seen in Apple's Face ID and it's [sic] use of face recognition as a form of user identification. Due to the popularity of AFR there have been many research efforts to perfect the algorithms that make it possible. Despite this constant research relating to AFR we see a lack in literature that approaches the question of how to gather the large amount of data that would make an AFR system effective in terms of real-world situations, without the need for perfectly posed pictures.

In this thesis, we leverage C++, OpenCV and several other libraries to extract normalized data points (facial images) that allow for the construction of a database capable of supporting AFR.

While this form of database creation may not be perfectly suitable for identification purposes due to the missing layer of information, such as name and identity, tied to the images stored. We were able to create multiple systems that were able to successfully extract facial data in a real time environment without effecting [sic] the run time of a video stream. Each system extracted facial data bsed on differing parameters and needs, the first performed a single extraction per frame, as in it only returned the most prominent face in the frame. The second extracted any number of faces per image, a feat that proved to be somewhat taxing as it caused several low dips in the frame-rate. Finally, the last system preformed [sic] an extraction followed by a normalization of faces detected in the stream. There was also an attempt to use the facial data extracted as a kick-starter to train a better facial cascade, but that ultimately came to a halt, as training a cascade needs to control for lighting, pose and angle. A feat that would require the processing of large amounts of video streams, and then filtering out the images that do not fit the parameters. This proved too time consuming for our case.