Access Type

Open Access Dissertation

Date of Award

January 2017

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Electrical and Computer Engineering

First Advisor

Nabil Sarhan

Abstract

Computer Vision (CV) has been deployed recently in a wide range of applications, including surveillance and automotive industries. According to a recent report, the market for CV technologies will grow to $33.3 billion by 2019. Surveillance and automotive industries share over 20% of this market. This dissertation considers the design of real-time CV systems with live video streaming, especially those over wireless and mobile networks. Such systems include video cameras/sensors and monitoring stations. The cameras should adapt their captured videos based on the events and/or available resources and time requirement. The monitoring station receives video streams from all cameras and run CV algorithms for decisions, warnings, control, and/or other actions. Real-time CV systems have constraints in power, computational, and communicational resources. Most video adaptation techniques considered the video distortion as the primary metric. In CV systems, however, the main objective is enhancing the event/object detection/recognition/tracking accuracy. The accuracy can essentially be thought of as the quality perceived by machines, as opposed to the human perceptual quality. High-Efficiency Video Coding (HEVC) is a recent encoding standard that seeks to address the limited communication bandwidth problem as a result of the popularity of High Definition (HD) videos. Unfortunately, HEVC adopts algorithms that greatly slow down the encoding process, and thus results in complications in real-time systems.

This dissertation presents a method for adapting live video streams to limited and varying network bandwidth and energy resources. It analyzes and compares the rate-accuracy and rate-energy characteristics of various video streams adaptation techniques in CV systems. We model the video capturing, encoding, and transmission aspects and then provide an overall model of the power consumed by the video cameras and/or sensors. In addition to modeling the power consumption, we model the achieved bitrate of video encoding. We validate and analyze the power consumption models of each phase as well as the aggregate power consumption model through extensive experiments. The analysis includes examining individual parameters separately and examining the impacts of changing more than one parameter at a time. For HEVC, we develop an algorithm that predicts the size of the block without iterating through the exhaustive Rate Distortion Optimization (RDO) method. We demonstrate the effectiveness of the proposed algorithm in comparison with existing algorithms. The proposed algorithm achieves approximately 5 times the encoding speed of the RDO algorithm and 1.42 times the encoding speed of the fastest analyzed algorithm.

Share

COinS