Intelligent Image Compression Using Traffic Scene Analysis

Abstract

AbstractThe quantity of images generated at the edge of the Cloud is growing year-on-year, which puts an increasing strain on existing telecommunications infrastructure. There is also an associated increased cost for transmission bandwidth and storage of video images in the Cloud. In our modern society we tend to accumulate data, and are reluctant to throw it away, without asking “what is the value of this data?” and “do we need it?”. One of the major sources of video streams are the increasing number of traffic cameras, used to maintain the efficient flow of vehicles on our roads. In this work we focus on images taken from road traffic cameras, and show how their transmission bandwidth and storage requirements can be reduced. By analysing video feeds on a simulated edge device, we have shown that it is possible to extract objects of interest from the image, and discard or dramatically reduce irrelevant information in the content. Our technique also generates associated metadata, in the form of JSON-LD, which annotates the original image and maintains its semantic fidelity and provenance after compression. Our technique is compatible with conventional compression techniques, and thus the potential bandwidth savings would be incremental. We present the potential savings that can be made in the transmission and storage of unstructured data, as well as some of the challenges still to be overcome.