Person Identification Using Anthropometric and Gait Data from Kinect Sensor

Virginia Ortiz Andersson, Ricardo Matsumura de AraújoAAAI 2015 · DOI

Introduction

Traditional biometric systems, such as fingerprinting or iris scanning, offer high accuracy but require active cooperation from the subject. In contrast, "passive" biometrics like facial recognition or gait analysis allow for non-intrusive identification from a distance. However, high-quality gait analysis typically demands expensive multi-camera setups and complex processing. This research explores the feasibility of using low-cost, off-the-shelf hardware—the Microsoft Kinect sensor—to identify individuals by combining their body proportions (anthropometry) and walking patterns (gait).

Methodology

The researchers utilized a Kinect for Xbox 360 to track 48 skeletal points at 30 frames per second. To mitigate the inherent noise in the sensor's tracking, an Auto Regressive Moving Average (ARMA) filter was applied to smooth joint positions.

Two distinct feature sets were extracted for analysis:

  1. Anthropometric Attributes (20 features): Static measurements including the length of the neck, spine, thighs, and total height.
  2. Gait Attributes (60 features): Dynamic measurements such as joint angles (hips, knees, ankles) during movement, alongside spatiotemporal parameters like step length, stride length, and velocity.
The skeletal model used to extract anthropometric features such as limb lengths and spine segments.

To evaluate these features, the team compared three machine learning classifiers: K-Nearest Neighbors (KNN), Support Vector Machines (SVM), and Multi-Layer Perceptrons (MLP). The study employed a novel dataset of 140 individuals, providing a significantly larger "gallery size" than previous benchmarks.

Key Insights and Results

The study's most significant finding is that physical proportions are far more predictive than movement patterns for identification. Anthropometric data alone yielded approximately 85% accuracy, whereas gait data alone dropped accuracy to roughly 60%. While combining both feature sets achieved the highest peak accuracy of 87.7%, the gait data provided only a marginal 3.3% improvement over body measurements alone.

Comparison of classification accuracy across different feature sets: (a) All attributes, (b) Anthropometric attributes, and (c) Gait attributes.

Classifier Performance

  • KNN: Achieved the best overall performance (87.7%) for large galleries. While it was outperformed by SVM in smaller groups (under 20 people), it proved more consistent and scalable as the database grew.
  • SVM: Followed closely with 86.3% accuracy.
  • MLP: Recorded the lowest mean accuracy at 84.7%.

Feature selection analysis further revealed that "Stride Length" is the most critical gait attribute; its removal caused the most significant decline in system accuracy.

Conclusion

While an 88% accuracy rate is impressive for a consumer-grade sensor, it remains insufficient for high-security applications. The Kinect's limited range and sensitivity to occlusions (such as heavy clothing) are ongoing challenges. Currently, this system is best suited as a "soft biometric"—a secondary layer of identification to complement other techniques like facial recognition. Future research will focus on identifying more robust gait features that can withstand noisy environments and testing models in uncontrolled, real-world scenarios.

Summary updated

BibTeX
@inproceedings{andersson2015person,
  author = {Virginia Ortiz Andersson and Ricardo Matsumura de Araújo},
  title = {{Person Identification Using Anthropometric and Gait Data from Kinect Sensor}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9212},
}
Figure