Face recognition in an unconstrained environment for monitoring student attendance

Traditional paper based attendance monitoring systems are time consuming and suscep- tible to both error and data loss. Where technical advances have attempted to solve the problem, they tend to improve only small portions i.e. confidence that data has been collected satisfactorily can be very high but technology can also be difficult to use, time consuming and impossible especially if the overall system is down. Camera based face recognition has the potential to resolve most monitoring problems. It is passive, easy and inexpensive to utilise; and if supported by a human safeguard can be very reliable. This thesis evaluates a strategy to monitor lecture attendance using images captured by cheap web cams in an unconstrained environment. A traditional recognition pipeline is utilised in which faces are automatically detected and aligned to a standard coordinate system before extracting Scale Invariant Feature Transform (SIFT), Local Binary Pattern (LBP) and Eigenface based features for classification. A greedy algorithm is employed to match captured faces to reference images with faces labelled and added to the training set over time. Performance is evaluated on images captured from a small lecture series over ten weeks. It is evident that performance improves during the series as new reference material is included within the training data. This correlation demonstrates that the success of the system is determined not only by the on-going capturing process but also the quality and variability of the initial training data. Whilst the system is capable of reasonable success, the experiments show that it also yields an unacceptably high false positive rate and cannot be used in isolation. This is primarily because the greedy nature of the algorithm allows the possibility of assigning multiple images of the same person captured in the same lecture to different students including ‘no shows’.