Speaker: Dr Reid Porter
When: 11:00AM Friday 13th July 2007
Venue: General Purpose South Room 78-420, St Lucia Campus
Abstract:
Geographically referenced (geo-spatial) video acquisition systems are now in practical use. Wide area imaging sensors are placed on helicopters, balloons, small aircraft or unmanned aerial vehicle and geographically referenced video is communicated to a ground station in real-time. Compared to satellite imagery, which provides data at time scales of months or years, geo-spatial video provides data to observe and model temporal phenomena at time scales of seconds or minutes.
Geo-spatial video exploitation presents new challenges for computer vision researchers. First, many objects of interest (e.g. vehicles and people) cover very few pixels and therefore specific recognition is very difficult. Second, moving object detection in this imagery is an unsolved problem. Data arrives at about 1 or 2 frames per second, which means point-like moving objects move anywhere from 1 to 200 pixels. In addition, the oblique viewing angles and incomplete digital elevation maps mean buildings and other landmarks suffer from parallax. This introduces a large amount of motion clutter. Finally, registration is often required in real-time and is therefore approximate, e.g., stationary objects might move up to 30 pixels over a short period of time. All these factors combined lead to unique, and extremely difficult recognition and tracking problems.
In this talk we will describe recent algorithmic work which aims to address these challenges, which includes a novel combination of object detection, moving object detection, and anomaly detection. We present initial results for a variety of recognition and tracking problems and will discuss future work in this area.
Biography:
Reid Porter is a researcher at the Los Alamos National Laboratory in New Mexico, USA.
This seminar is presented by the IEEE Computer Society, Queensland Chapter
Contact: Dr John Williams