PCRV Colloquim 21 Mei: Gertjan Burghouts

Op 21 mei aanstaande vindt het volgende PCRV onderzoekscolloquium plaats in Ravelijn (zaal RA 2503) van 15.30-17.00 uur. Aansluitend is er een borrel.

U kunt zich aanmelden via email: pcrs-gw@utwente.nl

De gastspreker tijdens dit colloquium is dr.ir. Gertjan Burghouts (TNO).

YOUTUBE AND OUR SEARCH FOR 48 TYPICAL HUMAN BEHAVIORS IN MOVIE CLIPS - DR. IR. GERTJAN BURGHOUTS (TNO)

Human behavior is complex. Usually we don't find behavior complex: we easily understand other people's behaviors, even within a second and without trouble. Yet, now imagine that you need to explain it to somebody else. How do you define 'give'? Yes, you might say: somebody has an object, moves his arm, and now another person has the object. But what if you don't see the object itself? I am pretty sure that you still recognize 'give'. So what is it then? Is it the fact that you see the arm moving? But we move our arms all the time. So, you might say: the person who gives the object needs to be close to the other person. But then the other person shouldn't move his arm, otherwise they may just be shaking hands. Alright. So, we are getting closer to the concept of 'give' - at least, when you consider 'give' from a perceptual point of view.

That's pretty hard right? Now imagine that you want a computer to understand it. For instance, to search for particular YouTube movie clips in the midst of billions of clips. That implies that we need the computer to interpret it - and that requires that we need to define human behavior. Like the example above. But that's only 1 type of behavior. We have analyzed 48 human behaviors and decomposed them into observable characteristics. We have programmed a computer to detect those observables in video clips, like 'one person is getting closer to another person' and 'one person moves his arm outward', and 76 other characteristics. The next challenge is to learn the computer which behaviors are described by which subsets of observables. We use a technique from Artificial Intelligence, the bridge between psychology and intelligent systems, to learn the computer 48 behaviors.

We competed with Stanford University, MIT, Berkeley, CMU and 7 other US universities. At the last evaluation of this research program, we achieved a number-1 ranking with our solution. We believe that the winning element has been the combination of psychology and artificial intelligence. In this talk, I will show example movie clips of the human behaviors, how we decomposed them into observable characteristics, and how we learned the computer to recognize the behaviors in 7.000 video clips.