Detecting Nodding in YouTube Videos: A Deep Learning Approach

Abstract

This thesis discusses the project of creating a human action detector to YouTube videos by deep learning and other tools and what is expected to be done in the future. The human action specifically studied is "shaking head", which was the focus of the detector. Most of the work is done by using existing ideas to make things work. In the late of the research, we also introduced OpenPose body landmarks to try to improve the efficiency of the model. In general, a detector is built and tested. There were 550 videos been detected and more than 5,000 moments were found. However, the accuracy needs to be further improved. The false-positive rate is 41.2%, while the false-negative rate is 10.4%. Our detection algorithm has the potential to detect 336 YouTube videos with 200 to 300 seconds in 1 hour. The detection algorithm simultaneously detects the video right after it is clipped, and it does not need to download the videos, which saves a lot of time.