Abstract:
A person’s activity in a library should be monitored to avoid any unwanted problems. In
this project, we have investigated a problem of image-based human action detection in a
library. It involves making a prediction by analyzing human poses, behavior, and actions
with objects from complex images instead of video. Comparing with all approaches, we
conclusively decided to use an algorithm YOLOv3 (You Only Look Once) which is latest
and more convenient. The algorithm utilizes anchor boxes, bounding boxes and a variant
of Darknet. We have created our own dataset collecting images from library and annotated
the dataset manually. During the research with this project, we have considered human
activities in a library into five section namely studying, phoning, using a computer, taking
book and sleeping. The proposed system provides not only multi-tasking knowledge with
classification but also localization of human and the equivalent actions instantaneously.
Interestingly, the proposed approach achieved a mean average precision (mAP) of 96.3%.
In the future, incorporation of real time data analysis will add value to this project.
