Event detection for smart conference room using spatiotemporal convolutional neural network

• Main Author: Tan, Yi Jian
• Format: Final Year Project / Dissertation / Thesis
• Published: 2020
• Subjects: Q Science (General)
• Online Access: http://eprints.utar.edu.my/3922/1/16ACB01606_FYP.pdf; http://eprints.utar.edu.my/3922/
• Institution: Universiti Tunku Abdul Rahman

Conferencing room is one of the most important workspace in an organization regardless of its organizational domain and scale. Endless of fate changing organizational decisions are made in this workspace. Thus, in order for one to be triumphing over the crowd, one must make certain that the management of the conference room must be the most constructive when compared to their competitors. That being said, there is an increasing trend of organizations attempt to incorporate several of analytic tools in their conference room. The integration of different tools in a conference room is frequently described as a “smart conference room” and being abbreviated as SCR. As there are many different types of analytic tools, this project mainly focuses on the tools that are used to monitor the usage of the conference room. In the existing focuses on the tools that are used to monitor the usage of the conference room. In the existing SCR systems, the most common techniques used are based on occupancy analysis. Occupancy analysis is a technique aimed to detect the presence of occupants via various sensors such as infrared sensor. However, this technique lack of the capability to model more information about the conference room. In order to overcome this, this project aims to implement the current state-of-the-art human action recognition (HAR) techniques to detect on-going events in a conference room. The HAR technique selected in this project is based on two-stream network with ResNet-34 variant and (2+1)D convolutional blocks. Besides, current state-of-the-art object detection technique which known as You Only Look Once (YOLOv3) will be used for analytical purposes, for instance, counting people in the conference room. The model will be pretrained on Kinetics dataset and fine-tuned on Conference dataset. The Conference dataset is collected from Company X and will be annotated and pre-processed prior to the training process. Consequently, all the models will be integrated into the web service of SCR system in order to work with other modules in the conference room. Consequently, the system is able to detect on-going events based on the human activities and provide useful analytic insights for effective conference room management.