Segmenting sign language into motor primitives with Bayesian binning.

Area of Science:

• Biomechanics
• Robotics
• Machine Learning

Background:

• Human movement trajectories exhibit power-law relationships between velocity and curvature.
• Parameters of these power laws vary across action segments, suggesting potential for unsupervised action segmentation.
• Identifying movement primitives is crucial for understanding and replicating complex actions.

Purpose of the Study:

• To investigate the efficacy of Bayesian binning (BB) for unsupervised segmentation of human actions into movement primitives.
• To apply BB using a Gaussian observation model with polynomial time dependence to wrist trajectory data from Israeli Sign Language (ISL) users.
• To determine optimal polynomial orders for approximating kinematic data and segmenting actions.

Main Methods:

• Utilized Bayesian binning (BB) with a Gaussian observation model featuring a polynomial time-dependent mean.
• Applied the method to wrist trajectory data from ISL users to segment actions into movement primitives.
• Analyzed the trade-off between polynomial order, trajectory approximation accuracy, and segmentation performance.

Main Results:

• Bayesian binning successfully segmented ISL trajectories into distinct movement primitives.
• Polynomial orders between 3 and 5 provided an optimal balance between complexity and accuracy, aligning with minimum acceleration/jerk models.
• Higher-order polynomials were required to approximate natural kinematics not strictly adhering to power laws.

Conclusions:

• Bayesian binning is a viable method for unsupervised segmentation of human actions into movement primitives based on kinematic data.
• The optimal polynomial order for approximating movement trajectories depends on the adherence to power-law properties.
• This approach has implications for gesture recognition, robotics, and understanding motor control.