|Institution:||KTH Royal Institute of Technology|
|Keywords:||Natural Sciences; Computer and Information Science; Computer Science; Naturvetenskap; Data- och informationsvetenskap; Datavetenskap (datalogi)|
|Full text PDF:||http://urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-166416|
The purpose of this study is to discuss and attempt to approach an answer to the question of how face detection could be used to measure attention in a lecture hall.The conclusion might help further studies in using face detection to provide teachers with tools which can be used to improve learning during lectures. Face detection in real time applications became possible in 2001 when Viola and Jones presented a new method several times faster than any previous attempt. In 2007 Liao et al. presented a method using multi-block local binary patterns (MB-LBP) for the purpose of overcoming the simplicity and limitations of the Viola-Jones method. Computer vision libraries such as OpenCV make it easy to implement such algorithms. It currently supports both the Viola-Jones algorithm and the MB-LBP algorithm. This study compared these two face detection methods to see how they perform in terms of sensitivity and precision and attempted to identified limitations of both methods when used to detect attention in a simulated lecture environment. The study was conducted using boosted algorithms and functionality provided by OpenCV. The input data consisted of a recorded simulated lecture with 6 subjects performing different poses, labeled either attention or no attention, during certain periods of time, each pose recognized from a previously recorded actual lecture as a commonly occurring pose. The most significant difference of performance identified in the study was that the MB-LBP method performed face detection in an image three times faster than for Viola-Jones which confirmed previous reported results. Both methods generated high sensitivity values for all poses, but low precision values for two of the poses.The ability of both methods to detect downward tilted faces contributed to a high number of false positives returned when subjects performed the two poses of subjects taking notes or subjects performing activities labeled as no attention. Due to the low precision values caused by this, both methods were not considered to measure attention effectively. It is therefore suggested to instead train a MB-LBP-based method for the specific task of measuring attention in a lecture hall by training it to reject downward-tilted faces and to accept only instances conforming to the chosen definition of attention.