Key Feature Extraction from Unlabelled Video Frames for Hidden Markov Models
Author(s)
Chng, Zhe Ming
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Wearable computers have gained traction in both industrial and consumer applications, providing large amounts of unlabelled data that require expensive processing and labelling to be effectively used for state-of-the-art machine learning algorithms. To this end, we investigate the use of Hidden Markov Models (HMMs) to label the timestamps of boundaries within video sequences, and in particular, the extraction of key features from video sequences to be used within HMMs. We look at the use of ArUco markers and hue-saturation values as potential key features for the HMM, as well as post-processing techniques that can be implemented, and compare their performance. We find that the use of ArUco markers along with forward-filling and Gaussian convolution post-processing techniques result in the lowest error achieved by the HMMs.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis