Generalized Multiple Correlation Coefficient as a Similarity Measurement between Trajectories

Julen Urain, Jan Peters

Introduction

In the fields of Human-Robot Interaction (HRI) and Imitation Learning, a fundamental challenge is the robust recognition of human gestures. Humans rarely perform the same motion identically; variations in speed, scale, and orientation are common. For a robot to learn effectively from demonstration, it requires a similarity metric that can identify "the same" motion regardless of these linear transformations.

Existing metrics often fall short. Signal processing techniques like Dynamic Time Warping (DTW) handle temporal shifts but struggle with spatial rotations. Topological methods like the Frechet distance focus on shape but ignore temporal sequence. Statistical measures like the RV Coefficient are invariant only to rotations, failing when motions are scaled or sheared. Distance Correlation (dCor), while powerful, often detects non-linear correlations, leading to false positives between fundamentally different shapes.

This paper introduces the Generalized Multiple Correlation Coefficient (GMCC), a novel similarity metric for multidimensional trajectories that is uniquely invariant to all linear transformations, including rotation, scaling, shear, and squeeze mappings.

Methodology

The GMCC is an extension of the Coefficient of Determination ($R^2$) and Pearson’s Correlation to multivariate data.

  1. Multivariate Extension: While standard Multiple Correlation measures how independent variables predict a single dependent variable, the authors extend this to cases where the target $Y$ is a vector (e.g., 3D coordinates over time).
  2. GMCC Formulation: GMCC is defined as a weighted sum of the Multiple Correlation Coefficients for each dimension of the target trajectory.
  3. Symmetrization: To create a valid distance metric, the authors define a symmetric version ($R_{sym}$) by averaging the correlation in both directions ($X$ to $Y$ and $Y$ to $X$). The final distance is calculated as $d = 1 - R_{sym}$.
  4. Linear Invariance: The metric is mathematically proven to be invariant to any non-singular linear transformation $H$. Specifically, $GMCC(X, HX) = 1$, resulting in a distance of zero between a trajectory and any linearly transformed version of itself.

Experimental Setup

The researchers evaluated GMCC using both synthetic and real-world datasets.

  • Synthetic 2D Trajectories: Linear and circular paths were subjected to Gaussian noise and various transformations (rotation, scaling, reflection, shear, and squeeze mapping).
  • Real-world 3D Trajectories: Using a KUKA LWR robot arm, 30 demonstrations were recorded across six motion types: lines, U-shapes, S-shapes, circles, triangles, and a "hello" gesture. Each type included variations in scale, starting position, and orientation.
Experimental setup using a KUKA LWR robot arm for recording 3D trajectories.

Key Results

GMCC demonstrated superior performance in both invariance and discriminative power.

  • Transformation Invariance: In synthetic tests, GMCC maintained a distance of 0.0 across all linear transformations. In contrast, the RV Coefficient and dCor showed significant distance increases when subjected to shear and squeeze mappings.
  • Clustering Accuracy: On the KUKA robot dataset, GMCC produced highly accurate hierarchical clustering, successfully grouping similar shapes despite variations in rotation or scale.
  • Discrimination: GMCC effectively distinguished between distinct shapes. For example, it yielded a distance of 0.87 between a line and a circle, whereas dCor yielded only 0.57, indicating that GMCC is less likely to confuse different motion classes.
  • Efficiency: Calculated via least-squares regression, GMCC is computationally efficient, making it suitable for real-time applications compared to the higher complexity of methods like DTW.
Clustering of 3D motion trajectories into distinct groups using the GMCC metric.

Discussion and Future Work

While GMCC is highly effective for linear transformations, it currently assumes temporal alignment. It cannot handle "time-warping" where gestures have different rhythmic variations. Future work involves developing a hybrid of DTW and GMCC to address temporal non-linearity. Additionally, the authors plan to integrate GMCC with Probabilistic Motion Primitives (ProMPs) to represent motions as probability distributions, further enhancing robot recognition capabilities in complex HRI scenarios.

Summary updated

BibTeX
@inproceedings{urain2019generalized,
  author = {Julen Urain and Jan Peters},
  title = {{Generalized Multiple Correlation Coefficient as a Similarity Measurement between Trajectories}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2019},
  url = {https://openreview.net/forum?id=sWar9QnwTy5},
}
Figure