Analyzing complex motion from video by variational inference for rehabilitation research - Rehabilitation Institute of Chicago

Skip to Content

Fri, Mar 18

Speaker: Gag Hua, Ph.D. (IBM T. J. Watson Research Center)

Title: Analyzing complex motion from video by variational inference for rehabilitation research

Abstract: Many emerging applications, including mixed/augmented reality rehabilitation for stroke and Parkinson patient, demand effective and efficient vision-based methods to analyze complex motions, such as articulated motion, deformable motion, and multiple motions. The fundamental challenges of this inverse problem come from two aspects: the high degrees of freedom in these complex motions, and the complications in the image measurements. The high dimensionality of this problem has plagued the scalability and efficiency of many existing methods.

In search for a new and scalable solution that overcomes the curse of dimensionality, we view this problem from another angle and conjecture that the complexity of such a problem can be approached by the collaboration among a set of low dimensional motion estimators. Therefore, it is natural to ask two fundamental questions: (1) what characterizes the optimal collaboration? (2) Does there exist an efficient computational diagram to obtain the optimum? This talk presents the exciting findings of our exploration to these questions under a probabilistic formulation, and many encouraging experimental results in several case studies.

In the end of my talk, along with demos of some of my other recent work on motion gesture based perceptual interface, I will present some potential use cases of our work on video based complex motion analysis for assisting the rehabilitation of patients suffered from stroke or Parkinson’s disease. Our ultimate goal is a marker-less non-intrusive vision based rehabilitation system based on a single camera which is cheap enough to be set up at a patient’s home. This has become a more and more realizable goal with the advancement of new types of vision sensors, such as Microsoft Kinect.

Host: William Z. Rymer