Mahalanobis Motion Generation

Raquel Urtasun, P. Glardon, Daniël Thalmann, Pascal FuaInfoscience 2004

Abstract

Representing motions as linear sums of principal components has become a widely accepted animation technique. While powerful, the simplest version of this approach is not particularly well suited to modeling the specific style of an individual whose motion had not yet been recorded when building the database: It would take an expert to adjust the PCA weights to obtain a motion style that is indistinguishable from his. Consequently, when realism is required, current practice is to perform a full motion capture session each time a new person must be considered. In this paper, we extend the PCA approach so that this requirement can be drastically reduced: For whole classes of motion such as walking or running, it is enough to observe the newcomer moving only once at a particular speed using either an optical motion capture system or a simple pair of synchronized video cameras. This one observation is used to compute a set of principal component weights that best approximates the motion and to extrapolate in real-time realistic animations of the same person walking or running at different speeds.

Introduction

In traditional computer animation, representing motion as a linear combination of "basis motions" via Principal Component Analysis (PCA) is a standard technique. However, these PCA weights are abstract and difficult to tune manually to match a specific person's unique style. To achieve high realism, current workflows often require exhaustive motion-capture (mocap) sessions for every new subject at every possible speed—a process that is expensive, time-consuming, and hardware-intensive. This research introduces a method to capture a unique "style" from a single snippet of movement and realistically predict how that same person would look moving at different speeds.

Methodology

The proposed approach follows a two-stage pipeline designed to move from a general database to personalized animation.

1. Building the Basis

The researchers first established a high-quality database of walking and running motions across various speeds using a Vicon optical mocap system. Each motion cycle is represented as a high-dimensional vector of joint angles. PCA is then applied to identify the primary components of human gait, reducing the dimensionality while preserving the essential characteristics of the motion.

2. Personalized Extrapolation and Mahalanobis Distance

When a new subject is observed, their motion is projected into the established PCA space. To capture the newcomer's unique style, the system calculates the Mahalanobis distance between the newcomer and the subjects in the database at the observed speed. Unlike standard Euclidean distance, the Mahalanobis metric accounts for the variance of the data, effectively weighting PCA components based on their statistical importance. This allows the system to synthesize motions at target speeds by taking a weighted average of database motions, leveraging the previously calculated distances.

Figure 1: (a) Cumulative variance of PCA components showing the efficiency of the basis; (b) Visualization of distinct clusters in PCA space representing individual motion styles.

Key Insights and Results

Style Clustering

A significant finding of this work is that individuals form distinct, statistically measurable clusters in PCA space. This consistency across different realizations of the same motion confirms that "style" is a stable attribute that can be leveraged for synthesis. By using a normalized distance metric, the system avoids washing out subtle stylistic flourishes that a simple Euclidean distance might ignore.

Visual Fidelity and Accuracy

In cross-validation tests, the extrapolated motions proved nearly indistinguishable from actual recorded motions of the same person at different speeds. The "Interpolation Error" (the distance between synthesized and real motion) was consistently much smaller than the "Inter-variability" (the difference between two different people), indicating that the system successfully captures a person's unique signature rather than a generic gait.

Figure 4: Visual overlay comparing synthesized motion (light blue) with ground truth recorded motion (yellow), demonstrating high fidelity in the extrapolation.

Robustness

The system also demonstrated success in extracting motion weights from low-resolution video. Even when the subject's limbs were less than 10 pixels wide, the PCA model provided a strong enough prior to overcome noisy data and produce realistic animations.

Conclusion

By combining PCA with Mahalanobis distance metrics, this method enables the extrapolation of realistic, individual-specific motions from a single observation. While currently focused on cyclical motions like walking and running, the framework provides a robust foundation for reducing the need for exhaustive motion capture. Future work may explore non-linear dimensionality reduction tools to capture more complex, non-repetitive motion manifolds and adapt the system for non-treadmill environments.

Summary updated

Abstract

Representing motions as linear sums of principal components has become a widely accepted animation technique. While powerful, the simplest version of this approach is not particularly well suited to modeling the specific style of an individual whose motion had not yet been recorded when building the database: It would take an expert to adjust the PCA weights to obtain a motion style that is indistinguishable from his. Consequently, when realism is required, current practice is to perform a full motion capture session each time a new person must be considered. In this paper, we extend the PCA approach so that this requirement can be drastically reduced: For whole classes of motion such as walking or running, it is enough to observe the newcomer moving only once at a particular speed using either an optical motion capture system or a simple pair of synchronized video cameras. This one observation is used to compute a set of principal component weights that best approximates the motion and to extrapolate in real-time realistic animations of the same person walking or running at different speeds.
BibTeX
@inproceedings{urtasun2004mahalanobis,
  author = {Raquel Urtasun and P. Glardon and Daniël Thalmann and Pascal Fua},
  title = {{Mahalanobis Motion Generation}},
  booktitle = {Infoscience},
  year = {2004},
}
Figure