Exploration of Body Pose Algorithms for 3D Skeleton-Based Gait Recognition

Due to recent advances in machine learning there have been in the last few years a new interest in biometric identification powered by these techniques. Among them is Gait Recognition, which tries to use the characteristic way in which one walks as the identifying biometric element. Gait Recognition has the potential to be a major technology in person identification, as it can be performed at distance and without the need for the subject to be close to the camera. Two main ways to perform Gait Recognition have been developed: appearance based and model based. Appearance-based methods are currently more precise, but because it uses features directly inferred from videos, it can be subject to spurious correlations. Model based methods, in turn, promise to overcome the above limitation by using skeletons obtained via Human Pose Estimation techniques as inputs. More specifically, 3D skeletons would allow for models to be invariant to perspective and, in theory should be the most invariant to irrelevant visual cues. However, these methods suffer from a major setback, i.e., the lack of precision 3D Human Pose Estimation algorithms. The present thesis work aims to study the effect of this problem on the general Gait Recognition field, and to suggest some way forward.