Structured Time Series Analysis for Human Action Segmentation and Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Human motion recognition from monocular sequences is challenging due to arbitrary viewpoints and person variations.
• Existing methods struggle with high dimensionality and non-parametric nature of human motion data.
• Accurate segmentation and alignment are crucial for reliable action recognition.

Purpose of the Study:

• To develop robust methods for structure learning and action recognition in human motion sequences.
• To enable real-time, accurate segmentation and recognition of human actions from continuous, unconstrained motion data.
• To model latent structures in time series data for effective motion similarity calculation.

Main Methods:

• Proposed Kernelized Temporal Cut (KTC) for nonparametric, high-dimensional change-point detection in multivariate time series.
• Introduced a spatio-temporal manifold framework to model latent data structures.
• Developed Dynamic Manifold Warping (DMW) for efficient spatio-temporal alignment and motion similarity calculation.

Main Results:

• KTC achieved real-time segmentation with high action segmentation accuracy.
• The spatio-temporal manifold framework effectively modeled latent time series structures.
• Combined algorithms enabled online human action recognition using limited labeled data.
• Demonstrated effectiveness on motion capture and 3D depth sensor data, handling noise and occlusion.

Conclusions:

• The proposed KTC and DMW algorithms provide an effective framework for human motion analysis.
• The approach enables accurate, real-time action segmentation and recognition from monocular sequences.
• The methods are robust to noisy and partially occluded data, facilitating transfer learning.