Learning Models of Human Motion

Problem
What does it mean to dance like Fred Astaire? To answer this question quantitatively we must have a computational model of human movement which can capture its subtlety and complexity. One possibility is to learn models of human motion from data, in the same way that phonemic models of speech are trained from speech corpora. Commercial motion capture systems, which digitize the motion of an actor in 3-D in real-time, can provide a data source for motion learning. In this project you will experiment with learning models of human motion from motion capture data. You will use two modeling tools in your study: Hidden Markov Models and Switching Linear Dynamic System Models. Matlab code for learning and synthesizing from these models will be provided, so it will help if you are already familiar with Matlab. 

Here's what you need to do:

1. Read the following papers which discuss Markov models and their application to human motion modeling
   • Rabiner, L., "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition", In Proceedings of the IEEE, 77(2), 257-286, 1989.
   • M. Brand, "Voice Puppetry", Proc. of SIGGRAPH 1999.
   • Pavlovic, V., Rehg, J., and MacCormick, J. "Learning Switching Linear Models of Human Motion", In Proc. of Neural Information Processing Systems (NIPS), Nov. 2000.
2. Experiment with Matlab code (provided) for learning motion models from data (provided).
3. Compare the performance of HMM and SLDS models on synthetic and real motion data.
4. Make some simple animations by sampling from your learned model (time permitting). 

Evaluation

You will write a three page report describing your experiments. It should compare and contrast the performance of the models you examined on the data. Comment on the strengths and weaknesses of the learning approach.