University of Minnesota
CLA Media Mill
help@umn.edu
612-301-4357


CLA Media Mill

11/7/2006 - Portable Video Data Processor

The work presented in this talk aims to produce a traffic vision data processor - a PC-based system that will accept recorded video from intersections, weaving sections, and other areas of interest and produce desired traffic data or other interesting traffic events. Depending on the traffic domain, "interesting events" may range from phenomena that are clear-cut and easily pre-specified by the user to more abstract phenomena that are defined as deviations from normal behavior across some user-specified threshold. In general, for uncontrolled environments such as traffic sites, where target behaviors vary depending on the intersection type, location, camera-viewpoint, etc., it makes more sense for the system to learn normal target behavior patterns from previously observed target trajectories; this strategy increases the ease of portability.

Although several vision-based trackers with emphasis on outdoor scenes exist, most trackers make use of a single visual cue that can provide good target detection as long as certain constraints or assumptions are satisfied. As soon as the scene changes in ways that violate these assumptions, the cues fail to provide any useful information, thereby rendering the tracker inaccurate. The system that will be discussed in this talk incorporates multiple visual cues so that the range of successful operation of the tracker can be increased by reducing the scene constraints. Several experimental results from various traffic sites will also be presented.

ITS

ITS

Transcript not available for this clip.

Embedding not available for this clip.

Downloading not available for this clip.