Access Type

Open Access Thesis

Date of Award

January 2017

Degree Type

Thesis

Degree Name

M.S.

Department

Electrical and Computer Engineering

First Advisor

Nabil Sarhan

Abstract

Interest in automated video surveillance systems has grown dramatically and with that so too has research on the topic. Recent approaches have begun addressing the issues of scalability and cost. One method aimed to utilize cross-layer information for adjusting bandwidth allocated to each video source. Work on this topic focused on using distortion and accuracy for face detection as an adjustment metric, utilizing older, less efficient codecs. The framework was shown to increase accuracy in face detection by interpreting dynamic network conditions in order to manage application rates and transmission opportunities for video sources with the added benefit of reducing overall network load and power consumption.

In this thesis, we analyze the effectiveness of an accuracy-based cross-layer bandwidth allocation solution when used in conjunction with facial recognition tasks. In addition, we consider the effectiveness of the optimization when combined with H.264. We perform analysis of the Honda/UCSD face database to characterize the relationship between facial recognition accuracy and bitrate. Utilizing OPNET, we develop a realistic automated video surveillance system that includes a full video streaming and facial recognition implementation. We conduct extensive experimentation that examines the effectiveness of the framework to maximize facial recognition accuracy while utilizing the H.264 video codec. In addition, network load and power consumption characteristics are examined to observe what benefits may exist when using a codec that maintains video quality at lower bitrates more effectively than previously tested codecs. We propose two enhancements to the accuracy-based cross-layer bandwidth optimization solution. In the first enhancement we evaluate the effectiveness of placing a cap on bandwidth to reduce excessive bandwidth usage. The second enhancement explores the effectiveness of distributing computer vision tasks to smart cameras in order to reduce network load.

The results show that cross-layer optimization of facial recognition is effective in reducing load and power consumption in automated video surveillance networks. Furthermore, the analysis shows that the solution is effective when using H.264. Additionally, the proposed enhancements demonstrate further reductions to network load and power consumption while also maintaining facial recognition accuracy across larger network sizes.

Download

Share

COinS