Hierarchical Activity Recognition: From Motion to Intention
Humans and other animals spend a good portion of their lives observing and modeling the behavior of those around them. Such information provides valuable input into an agent's decision-making process, since it is the basis for coordination, imitation, instruction, and a whole host of other cognitive processes. In this talk, I will provide an overview of my group's recent work on hierarchical activity modeling, where activities are modeled at a variety of temporal resolutions using decomposable parametric distributions. I will describe both generative and discriminative models of activity recognition, building on recent ideas in graphical models and kernel methods. I will show how computationally tractable algorithms for learning and inference in hierarchical activity models can be developed by carefully exploiting the structure of temporal hierarchies. The framework will be illustrated with examples from a variety of areas, including humanoid robotics, intelligent tutoring systems, multi-agent coordination, reinforcement learning, and weather prediction.