Evaluation of methods and configuration for constructing a people detection system

Abstract

In recent years a steady transition from analogue to digital surveillance cameras has been observed. With this, computational power on cameras has been increasing, pushing the limit of applicable real-time solutions. One of these real-time solutions are people detection systems, which is the topic of this thesis. A wide range of methods and configurations utilized when finding people in overhead view footage will be concerned.
This thesis will also contain an in-depth comparison of these, hopefully aiding in future development of real-time classifiers. Haar features and histogram of oriented gradients are the two main features in this thesis, which have been thoroughly examined. A comparison between different classifier structures, such as cascades, trees and support vector machines is also presented. The final measurements performed during this master thesis demonstrates interesting tendencies, which are applicable in future development.

System Overview

All evaluations were performed in a system developed during this thesis. One of the main elements evaluated was different kinds of features. Features are used to describe images, which can be done in various ways.
One of the features examined was the Haar feature, often seen in real-time detection systems. Another more complex feature used was the histogram of oriented gradients.
The next step is to use these features in some form of structure. Two types of structures have been evaluated. One of the structures evaluated was a cascade, enabling fast detection of people in an image. This structure can be seen below.

Another more complex structure was the tree, seen below.

These structures, and many other methods and configurations can be read about in the report.

Publication

Evaluation of methods and configuration for constructing a people detection system
Marcus Danielsson and Marcus Johansson; Master Thesis, Lund University, May 2011.