AASS Seminar - Deep probabilistic models for human activity modeling

The research centre AASS arranges a seminar with Judith Butepage, KTH, Stockholm.

Abstract

Human behavior is determined by many hidden factors such as intention, need, belief and environmental aspects. Therefore it is difficult to classify and anticipate human actions and to predict possible motion trajectories in 3D. Challenges are also posed by high noise levels in sensory channels and limited availability of data. In this talk, I will present how we approach these difficulties with help of deep probabilistic models, such as variational autoencoders. The first part of the talk will be concerned with human motion prediction and how probabilistic models can be used to anticipate many possible future trajectories. In the second part I will present recent work on combining high-level action modeling with low-level trajectory modeling. I will demonstrate how this approach is able to classify and anticipate semantic labels as well as low-level features over long time horizons.

Speaker's bio

Judith is currently a PhD student at the Robotics, Perception and Learning Lab at the Royal Institute of Technology, KTH, Stockholm. Her research focuses on the representation and interpretation of human actions and movements. Judith holds a Bachelor degree in Cognitive Science from the University of Osnabrück, Germany, and studied Machine Learning at KTH.