Segmenting Continuous Motions with Hidden Semi-markov Models and Gaussian Processes.

Area of Science:

• Robotics
• Machine Learning
• Time Series Analysis

Background:

• Humans naturally segment continuous information, like speech and motion, into recognizable units without explicit cues.
• This unsupervised segmentation capability is crucial for robots to learn diverse actions, gestures, and languages flexibly.
• Current methods may struggle with the complexities of real-world continuous data segmentation.

Purpose of the Study:

• To propose a novel Gaussian process-hidden semi-Markov model (GP-HSMM) for unsupervised segmentation of continuous time series data.
• To enable robots to learn and recognize unit actions from continuous motion data.
• To achieve accurate and robust segmentation without predefined segment points.

Main Methods:

• Developed a generative model combining Hidden Semi-Markov Models (HSMM) with Gaussian Processes (GPs) for emission distributions.
• Utilized forward filtering-backward sampling for estimating model parameters, including segment lengths and classes.
• Applied the GP-HSMM to analyze continuous time series data, specifically motion capture datasets.

Main Results:

• The GP-HSMM demonstrated comparable performance to existing methods on simple exercise motion data.
• On complex karate motion data, the GP-HSMM achieved a segmentation accuracy of 0.92, surpassing other techniques.
• The model successfully segmented continuous motion sequences into meaningful unit actions.

Conclusions:

• The proposed GP-HSMM offers an effective approach for unsupervised segmentation of continuous time series data.
• This method holds significant potential for enhancing robotic learning of actions and gestures.
• GP-HSMM provides a robust and accurate solution for complex motion segmentation tasks.