Gait Recognition from Markerless 3D Motion Capture

State of the art gait recognition methods often make use of the shape of the body as well as its movement, as observed in the use of Gait Energy Images(GEIs), for recognition. However, it is desirable to have a method that works exclusively with the movement of the body, as clothing and other factors may interfere with the biometric signature from body shapes. Recent advances in markerless motion capture enable full 3D body poses to be estimated from unconstrained video sources. This paper describes how one such technique can be used to provide improved performance for verification tests. The markerless motion capture algorithm fits the 3D SMPL body model to a 2D image. Joint rotations from a single cycle are extracted from the model and matched using a verification system trained using an automated machine learning system, auto-sklearn. Evaluations of the method were performed on the CASIA-B gait dataset, and results show competitive verification performance with an Equal Error Rate of 18.40%.