Joint International Workshops on Structural and Syntactic Pattern Recognition
and Statistical Techniques in Pattern Recognition 2016, preprint
Learning Robust Features for Gait Recognition
by Maximum Margin Criterion (Extended Abstract)
Michal Balazia (0000-0001-7153-9984) and Petr Sojka (0000-0002-5768-4007)
Faculty of Informatics, Masaryk University, Botanick´a 68a, 602 00 Brno, Czech Republic
xbalazia@mail.muni.cz and sojka@fi.muni.cz
In the field of gait recognition from motion capture (MoCap) data, designing humaninterpretable
gait features is a common practice of many fellow researchers. To refrain
from ad-hoc schemes and to find maximally discriminative features we may need to
explore beyond the limits of human interpretability. This paper contributes to the state-ofthe-art
with a machine learning approach for extracting robust gait features directly from
raw joint coordinates. The features are learned by a modification of Linear Discriminant
Analysis with Maximum Margin Criterion (MMC) so that the identities are maximally
separated and, in combination with an appropriate classifier, used for gait recognition.
Recognition of a person involves capturing their raw walk sample, extracting gait
features to compose a template that serves as the walker’s signature, and finally querying
a central database for a set of similar templates to report the most likely identity.
The goal of the MMC-based learning is to find a linear discriminant that maximizes
the misclassification margin. We discriminate classes by projecting high-dimensional
input data onto low-dimensional sub-spaces by linear transformations with the goal of
maximizing the class separability. We are interested in finding an optimal feature space
where a gait template is close to those of the same walker and far from those of different
walkers. A solution to this optimization problem can be obtained by eigendecomposition
of the between-class scatter matrix minus the within-class scatter matrix. Obtaining
the eigenvectors involves a fast two-step algorithm in virtue of the Singular Value
Decomposition. Similarity of two templates is expressed in the Mahalanobis distance.
Extensive simulations of the proposed method and eight state-of-the-art methods
used a CMU MoCap sub-database of 54 walking subjects that performed 3,843 gait
cycles in total, which makes an average of about 71 samples per subject. A variety of
class-separability coefficients and classification metrics allows insights from different
statistical perspectives. Results indicate that the proposed method is a leading concept for
rank-based classifier systems: lowest Davies-Bouldin Index, highest Dunn Index, highest
(and exclusively positive) Silhouette Coefficient, second highest Fisher’s Discriminant
Ratio and, combined with rank-based classifier, the best Cumulative Match Characteristic,
False Accept Rate and False Reject Rate trade-off, Receiver Operating Characteristic
(ROC) and recall-precision trade-off scores along with Correct Classification Rate, Equal
Error Rate, Area Under ROC Curve and Mean Average Precision. We interpret the high
scores as a sign of robustness. Apart from performance merits, the MMC method is also
efficient: low-dimensional templates (dimension ≤ #classes−1 = 53) and Mahalanobis
distance ensure fast distance computations and thus contribute to high scalability.
The full research paper Learning Robust Features for Gait Recognition by Maximum
Margin Criterion has been accepted for publication at the 23rd International Conference on
Pattern Recognition (ICPR 2016), Cancun, Mexico, December 2016. arXiv:1609.04392
1