Goal
Goal of the IPSAR project (Image Processing for Search and Rescue) is to develop a complete, real - time system for detection of humans and other targets in order to use it in search and rescue applications. The work is conducted in the University of Split on the Department of Electronics (Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture) and Department of Polytechnics (Faculty of Science).
Motivation
Search and rescue (SAR) missions for lost persons are an everyday task in most countries. SAR missions are generally connected with non-urban (or at least sub-urban) terrain: mountains, forests, sea, deserts, etc. Usually, relatively large terrain should be searched. There are many resources that may be used for search missions: human trackers, searching and tracking dogs, aerial searching, and infrared cameras. Such missions can last for several days and therefore, significant financial and human resources are needed.
Any possibility of improving speed of the search as well as increasing chances of finding a lost person(s) should be investigated because eventual benefits are almost priceless. Focus of our research is aerial searching. Currently, aerial searching is conducted by naked eye and, when closer inspection of a particular area is needed, binoculars are used as an aid. Occasionally, generally on large missions with a lot of public attention, a series of photos of the searching area are taken from air to be visually inspected later.
Direct visual inspection from the air does not guarantee that a missing person will be found. Moreover, even photo-taking and later inspection does not guarantee finding the person. Namely, at least several hundreds, some times thousands of photos are taken in each mission. Visual inspection of so many pictures is hard and a time-demanding task, and still there is a possibility that lost persons will be omitted. Of course, this is not a real time approach, and finding a person several hours later may make the difference between life and death.
“Looking and not seeing” is a well known problem in human search. That problem will be, at least partially, avoided in an image processing system. It means that current and future observation systems provide a large amount of data which can hardly be completely processed manually: therefore, automatic systems are required for image interpretaton. Automatic human detection from long range images is a challenging problem since both precision and recall must be optimized. Humans must not be missed and false alarms must be reduced to a minimum.
