Publication: Asymmetric cascade face detector on multi-camera images
Loading...
Date
2023-09-01
Authors
Sirajdin Olagoke Adeshina
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Classroom attendance is one of the measures that ensure students’ presence and
punctuality in day-to-day classroom activities. Researchers emphasize its impact on
the academic performance of students. The traditional method of carrying out this measure
is of no standard in today’s classroom attendance system because of population,
time, sophistication, exhaustiveness, classroom size, and manipulative influence. The
conventional automated classroom attendance systems are designed for small classrooms
using a single camera. Therefore, a design and implementation of a multicamera
setup for a large lecture theatre attendance system is considered. the USM DK5
lecture theatre containing 14 volunteered participants (13 males and a female) of different
races was used for the exercise. A move-shift seating arrangement, producing
a total of 62 multiple-face datasets was obtained which was used for the evaluation of
the proposed algorithm. An asymmetric cascade face detector was trained using faces
and non-face samples. Selected Haar-like features were used based on their attention
focus, area and width-to-height ratio, and feature specialization to train the algorithm.
The asymmetric goal is set to minimize the False Rejection Rate (FRR) relative to the
False Acceptance Rate (FAR). A relationship between the two factors (FRR and FAR)
was established using a constant (l) as a trade-off between the two factors for automatic
adjustment during training. Consequently, a TPR comparison of the proposed
approach with the state-of-the-art tiny face (ResNet101) deep learning algorithm on
images captured in USM (DK5) lecture theatre produced 96.3% to 99.6%, with an
overhead constraint of 22.6s to 650.3s respectively. Tiny face as a deep learning approach with high level of accuracy, its limitations lie on the training data required,
computational complexities, and implementation on the resource constraint platforms.
Based on the experimental results on classroom datasets, the proposed approach shows
an improvement of 8% TPR (output result of low FRR) and 7% minimization of the
FRR. The average learning speed of the proposed approach was improved with 1.02s
execution time per image as compared to 2.38s of the original algorithm.