This thesis describes how a system for detecting victims
in a disaster scenario can be implemented in software
using known methods from computer vision. The
goal has been to create a system that can process video
input from the camera on a robot. The first chapter
provides background on disaster management worldwide
and explores the cases where robots have been
deployed for rescue work after a disaster has struck.
To this day, robots have not played a central role in
search and rescue operations after any large disaster.
The purpose of the human detection system is to work
as a visual aid to a human robot operator. This can
hopefully increase the chance of locating victims if the
operator is burdened by mental fatigue as has been
the case during robot operations.
The final detection system consists of 3 detection modules
that all search for the human head or face in
images. The first detector uses the HOG descriptor
and an SVM classifier to determine if a human head
is present within a smaller region of the image. The
second detector uses the Viola-Jones object detection
framework to detect human faces. The third detector
uses template matching to search for the head using
quarter circle shapes.
The combined detection system achieved a detection
rate of up to 62 % on a varied set of image with humans.
On a simulated disaster image set the detector
has a detection rate of only 14 %.