Real-Time Face Detection on High Resolution Smart Camera System

Yasir Mohd Mustafah (2010). Real-Time Face Detection on High Resolution Smart Camera System PhD Thesis, School of Information Technol and Elec Engineering, The University of Queensland.

Attached Files (Some files may be inaccessible until you login with your UQ eSpace credentials)
Name Description MIMEType Size Downloads
s41230033_phd_abstract.pdf Thesis Abstract application/pdf 77.35KB 2
s41230033_phd_finalthesis.pdf Final Thesis application/pdf 9.25MB 13
Author Yasir Mohd Mustafah
Thesis Title Real-Time Face Detection on High Resolution Smart Camera System
School, Centre or Institute School of Information Technol and Elec Engineering
Institution The University of Queensland
Publication date 2010-08
Thesis type PhD Thesis
Supervisor Prof. Brian C Lovell
Dr Abbas Bigdeli
Prof. Neil G Bergmann
Total pages 183
Total colour pages 15
Total black and white pages 168
Subjects 08 Information and Computing Sciences
Abstract/Summary Currently, video data in surveillance system is used predominantly as a forensic tool, thus losing its primary benefit as a proactive real-time alerting system. In addition, most of the current surveillance systems capture only low resolution video of the monitored scene making it difficult to analyze the data. To remedy these problems, National ICT Australia started a project called Intelligent CCTV for Proactive Security that aims to improve the counter-terrorism capabilities and sensitivities of surveillance systems. A high resolution (5 Megapixels) smart camera system is designed to be used as a component in a crowd surveillance system that has a face recognition capability. The smart camera is used to detect faces in the captured high resolution images as a way to reduce communication bandwidth and the processing load of a client processor that runs the face recognition software. In the early 2000s, Viola and Jones proposed a robust and very fast face detection algorithm that has attracted the attention of researchers in the face detection field. However, even the Viola-Jones face detection algorithm does not achieve real-time and produces many false positives when detecting faces on high resolution images. In order to achieve real-time face detection and minimal errors on the high resolution smart camera system, a two-stage face detection algorithm is proposed. The first stage of the real-time face detection is a fast region of interest detection algorithm to roughly estimate regions containing faces in the image. The second stage is an accurate face detection algorithm used to detect the precise face locations within the estimated face regions. Experiments show that this approach could improve the face detection speed on high resolution images tremendously and in addition reduces false positives in the detection. The region of interest detection is designed to be pixel based so that it can be processed at the same rate as the image sensor of the smart camera. The algorithm consists of three components: 1) Background subtraction algorithm, 2) Skin color detection algorithm, and 3) Region of interest classification algorithm. In the crowd surveillance problem, it is assumed that the cameras are static and produce color images. Hence, a background subtraction algorithm and skin color detection algorithm can be used to reduce the region of interest to the moving skin colored regions. For the face detection stage, we utilized a cascade of boosted face classifiers following the concept proposed by Viola and Jones. Since weak classifier evaluation is the main computation and the bottleneck of the face detection algorithm, we proposed a novel weak classifier or feature type called the Square Patch feature to achieve faster detection speeds. This feature requires fewer memory accesses and arithmetic operations compared to the Haar-like feature proposed by Viola-Jones. We also utilized the Realboost machine learning algorithm, as opposed to Adaboost, to produce a faster face detector with fewer weak-classifiers. To exploit the parallelism on hardware, we proposed a parallel cascade of face detector classifiers that can achieve much faster execution speed. Finally, the hardware architecture of the two-stage face detection is designed for the smart camera hardware. The region of interest detection hardware is designed to be implemented as a low level processing module of the smart camera. The low level processing is executed as soon as a pixel is available without having to buffer the whole image beforehand. Meanwhile, the face detection hardware is designed to be implemented as a high level processing module of the smart camera. The face detection algorithm hardware takes the output of the region of interest detection and reads the image buffer of the smart camera to perform face detection.
Keyword smart camera
background subtraction
skin color detection
hardware architecture
real-time system
Additional Notes Pages in Color: 24, 45, 52, 53, 74, 75, 80, 106, 122, 123, 125, 127, 128, 129, 132.

Citation counts: Google Scholar Search Google Scholar
Created: Mon, 28 Mar 2011, 23:43:03 EST by Mr Yasir Mohd Mustafah on behalf of Library - Information Access Service