Semantic Analysis of Motion Data

Abstract

The goal of this project is to develop effective algorithms to provide a semantic analysis of motion data from different perspectives. Although motion data can be generally represented by multivariate time-series (MTS), it has its specific characteristics which distinguish it from other MTS categories. In this work, I focus on these characteristics to find the commonalities and particularities in motion data which coincide to semantic concepts that are understandable by domain experts. Our algorithms fall into the following three specific categories of metric learning, dictionary learning and deep learning designs for motion data analysis.

Algorithms:

  • Distance-based LMNN
  • Large margin multiple-kernel learning (LMMK)



Algorithms:

  • Multiple-kernel dictionary learning for feature selection in motion data (MK-FS)
  • Confident kernel sparse coding (CKSC)
  • Multiple-Kernel Dictionary Learning (MKD-SC)
  • Non-negative kernel sparse coding (NNKSC)