[Paper Review] Learning Robust Features for Gait Recognition by Maximum Margin Criterion

Michal Balážia, Petr Sojka|arXiv (Cornell University)|Sep 14, 2016

Gait Recognition and Analysis9 references4 citations

TL;DR

This paper proposes a machine learning approach that learns robust, discriminative gait features directly from raw 3D joint coordinates using a modified Linear Discriminant Analysis with Maximum Margin Criterion (MMC). The method outperforms eight existing gait recognition techniques on the CMU MoCap database, achieving state-of-the-art performance in class separability and rank-based classification metrics, including a 91.02% correct classification rate and 0.2996 mean average precision.

ABSTRACT

In the field of gait recognition from motion capture data, designing human-interpretable 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-of-the-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 so that the identities are maximally separated and, in combination with an appropriate classifier, used for gait recognition. Experiments on the CMU MoCap database show that this method outperforms eight other relevant methods in terms of the distribution of biometric templates in respective feature spaces expressed in four class separability coefficients. Additional experiments indicate that this method is a leading concept for rank-based classifier systems.

Motivation & Objective

• To develop a data-driven method for learning gait features directly from raw 3D joint coordinates, avoiding human-designed, interpretable features.
• To improve gait recognition performance by maximizing the margin between identity classes in the feature space.
• To address the limitations of ad-hoc geometric feature engineering by leveraging machine learning for optimal discriminative feature extraction.
• To evaluate the proposed method on a standard benchmark (CMU MoCap) and compare it against eight established gait recognition techniques.

Proposed method

• The method employs a modified Linear Discriminant Analysis with Maximum Margin Criterion (MMC) to learn a linear transformation that maximizes inter-class separation and minimizes intra-class variation.
• Input data consists of raw 3D joint coordinates from a gait cycle, normalized to a fixed length T, forming a high-dimensional vector of size 3JT.
• The MMC objective function is optimized to maximize the margin between different identity classes while minimizing within-class scatter.
• The resulting low-dimensional feature space enables efficient template representation and fast similarity computation using Mahalanobis distance.
• Feature templates are stored in a database and used for recognition via similarity matching with a rank-based classifier.
• The approach avoids reliance on hand-crafted geometric features and instead learns features directly from the data distribution.

Experimental results

Research questions

• RQ1 Can a machine learning approach based on Maximum Margin Criterion learn more discriminative gait features than traditional, human-designed geometric features?
• RQ2 How does the MMC-based feature learning method compare to eight established gait recognition techniques in terms of class separability and recognition accuracy?
• RQ3 To what extent does the proposed method improve performance on rank-based classification metrics such as CMC, EER, AUC, and MAP?
• RQ4 Can the learned features generalize to unseen identities, especially in surveillance scenarios with no labeled data?

Key findings

• The MMC-based method achieved the highest correct classification rate (CCR) of 91.02% on the CMU MoCap database, outperforming all eight compared methods.
• It recorded the best mean average precision (MAP) of 0.2996, indicating superior performance in rank-based classification systems.
• The method achieved the highest silhouette coefficient (SC) of 0.2227, indicating strong inter-class separation and intra-class compactness.
• It recorded the lowest Davies-Bouldin Index (DBI) of 96.2, confirming superior clustering quality and class separability.
• The method achieved a 1.278 FDR (Fisher Discriminant Ratio), the highest among all methods, indicating strong discriminative power.
• The approach demonstrated high scalability due to low-dimensional feature templates and efficient Mahalanobis distance computation.