Learn, hack!

URL

http://events.ccc.de/congress/2005/fahrplan/attachments/647-slides-hopalong-casulty.pdf

File name

647-slides-hopalong-casulty.pdf

File size

1.1 MB

MD5

b07977a1cd0dca2cbe968ca6f556ad80

SHA1

cd66fe216b80167e74033cb606be0c720b06ad56

Automated analysis of surveillance videos has seen a lot of research in recent years. Face recognition and person tracking are widely available, more sophisticated behaviour analysis is coming. The aim of the current talk is an overview into the methods used for analysis, their current performance and limitations. Automated analysis of videos is a hot research topic currently, mostly fuelled (and funded) by interest in surveillance applications. Some of the work focuses on /identifying/ persons by individual differences in their motion patterns, e.g., the way the walk. Much current work tries to determine human interaction behaviour, e.g. whether two persons are standing and talking or whether they are fighting. A last big area is that of trajectory analysis, e.g. distinguishing persons walking straight across an open place from persons sticking around longer. This talk will give an overview into whats possible currently and then introduce some of the common methods of motion analysis with a focus on real-time capability. It will touch upon motion-history images, model-based tracking, graphical models for time-series and learning methods for classification. Throughout, pointers to toolkits that can be used to implement the methods presented will be provided. Of course, there are still a lot of problems, some of them quite fundamental, e.g. occlusions, crowds, influence of rain and wind, and the like. These problems, and their causes, will be explained under the assumption that the audience will be able to make creative use of this knowledge for playing with the system.