Sports Video Anonymisation and Accuracy Prediction

Authors: César Marrades Cortés, Keith Quille, Jelena Vasić, Seán McHugh

POLIBITS, Vol. 60, pp. 13-17, 2019.

Abstract: Video anonymisation, machine learning

Keywords: The anonymisation of people featuring in videos is required in many contexts. One of these is the physical education state exam in Ireland, where secondary school students are assessed as prescribed by the National Council for Curriculum and Assessment (NCCA), with the use of video recordings among other tools. For reasons of GDPR, all the video material presented for grading must not reveal the identity of the students. The work presented here was undertaken on consultation with the NCCA regarding their needs in this area and comprised two distinct tasks: (1) the implementation and testing of an anonymiser program in C#, supported by libraries Accord and EmguCV, each defining a different variant of the program; (2) the use of machine learning predictive models in Weka to investigate which of various factors (such as camera quality, camera angle, sport) affect the anonymisation program’s performance on sports videos. One hundred video inputs, resulting in 200 outputs (one for each of the two libraries, per input), were used and the best anonymisation success prediction model had an accuracy of 94% and a specificity of 95.2%. This work forms a base upon which a full automated video anonymisation system could be built, most importantly generating knowledge on what measures can be taken towards the optimisation of video anonymisation performance.

PDF: Sports Video Anonymisation and Accuracy Prediction
PDF: Sports Video Anonymisation and Accuracy Prediction


Table of contents of POLIBITS 60