Abstract

People all around the world are becoming more and more accustomed to watching 360-degree videos, which offer a way to experience virtual reality. While watching videos, it enables users to view video scenes from any perspective. To reduce bandwidth costs and provide the video with less latency, 360-degree video caching at the edge server may be a smart option. A hypothetical 360-degree video streaming system can partition popular video materials into tiles that are cached at the edge server. This study uses the Least Recently Used (LRU) and Least Frequently Used (LFU) algorithms to accomplish video caching and suggest a system architecture for 360-degree video caching. Two 360-degree videos from 48 users' head movements are used in the experiment, and caching between the LRU cache and LFU cache is compared by changing the cache size. The findings demonstrate that, for varied cache sizes, utilizing LFU caching outperforms LRU caching in terms of average cache hit rate.

In the first part of the research, we compared LRU and LFU caching algorithm. In the second part of the research, a suitable caching strategy model was developed based on user’s field of view. Field of view (FoV) is the term used to describe the portion of the 3600 videos that viewers typically see when watching 3600 videos. Edge caching can be a smart way to increase customer satisfaction while maximizing bandwidth usage (QoE). A 3600-video caching strategy has been developed in this study using three machine learning models that use random forest, linear regression, and Bayesian regression. As features, tiles' frequency, user's view prediction probability, and resolution were used. The created machine learning models are designed to decide the caching method for 360-degree video tiles. The models can forecast the frequency of viewing for 3600 video tiles (subsets of a full video). With a predictive R2 value of 0.79, the random forest regression model performs better than the other suggested models when the outcomes of the three developed models are compared.

In the third part of the research, to compare our machine learning algorithm with LRU algorithm, a python test bench program was written to evaluate both algorithms on the test set by varying the cache size. The results demonstrate that our machine learning approach, which was created for 360-degree video caching, outperforms the LRU algorithm.

Date of publication

Fall 12-9-2022

Document Type

Thesis

Language

english

Persistent identifier

http://hdl.handle.net/10950/4130

Committee members

Jounsup Park, Ph.D., Premananda Indic, Ph.D., Arun Kulkarni, Ph.D.

Degree

Masters in Electrical Engineering

Share

COinS