Learning Components for Human Sensing
Fernando De la Torre
Department of Computer Science, Carnegie Mellon University
Tuesday December 1, 2009
12:00 pm - SENSQ 5317
Hosted by Liz Marai
AbstractProviding computers with the ability to understand human behavior and infer people's mental states from sensory data (e.g. audio, video, motion capture) is an essential part of many applications that can benefit society such as human computer interaction, social robotics, and clinical diagnosis. In this talk I will give an overview of several ongoing projects in the human sensing lab, including our current work on depression assessment and deception detection from video, as well as hot-flash detection from wearable sensors.
A critical element in the design of any behavioral sensing system is to determine what constitutes a good representation of the data for encoding, segmenting, classifying and predicting subtle human behavior. I will propose several extensions of Component Analysis methods (e.g. kernel principal component analysis, support vector machines, spectral clustering) that are able to extract spatio-temporal patterns or components that are optimal for several human sensing tasks. I will show how the proposed techniques outperform state-of-the-art algorithms in problems such as temporal alignment of human behavior, temporal segmentation of human activity, learning image alignment for facial feature detection, learning a vocabulary for facial expression and a technique to discover discriminative events in time series. The talk will be adaptive, and I will discuss the topics of major interest to the audience.