Riemannian Spatio-Temporal Features of Locomotion for Individual Recognition.

Area of Science:

• Computer Vision
• Biometrics
• Geometric Deep Learning

Background:

• Individual recognition via skeletal sequences is crucial for security and HCI.
• Existing methods often lack explicit quantitative individual differences.
• Gait analysis presents unique biometric characteristics for identification.

Purpose of the Study:

• Propose a novel 3D spatio-temporal geometric feature representation for gait recognition.
• Explicitly reveal intrinsic differences between individuals based on locomotion.
• Improve classification accuracy for biometric identification.

Main Methods:

• Constructing mean sequences on a Riemannian manifold by aligning motion data.
• Modeling differences as spatial state descriptors relative to the mean sequence.
• Applying a temporal hierarchy of covariance for higher-order statistical representation.
• Utilizing kernel metric learning for enhanced classification.

Main Results:

• Demonstrated promising results on the CMU Mocap and UPCV Gait databases.
• Successfully evaluated on a newly constructed database for running gait.
• Analyzed key factors influencing walking gait recognition.
• The proposed method effectively captures unique individual biometric characteristics.

Conclusions:

• The novel 3D spatio-temporal geometric feature representation offers a robust method for individual recognition.
• The approach provides explicit quantitative differences, overcoming limitations of prior work.
• This method shows significant potential for applications in public security, HCI, and surveillance.