Towards invariance in gait recognition

Abstract

The gait of a person is a very attractive biometric as it can be captured from a distance, without the subject's cooperation. However, there is a large number of confounding factors that may dramatically affect its performance. The main objective of this work is to set a path towards full invariance in gait recognition. While recent deep learning techniques have shown promising results in tackling similar problems, they require large amounts of, often labelled, data. Having this in mind, a new multi-modal gait dataset is introduced, comprised of more than 6.5 million frames of real 3D human motion capture data and 2D image data. Based on this dataset, a data augmentation methodology is also introduced that provides the tools to synthetically generate data, while simultaneously controlling several different confounding factors. The way the data were collected and prepared, to introduce the augmentation methodology, provided the opportunity to answer a fundamental question regarding the source of identity information in gait recognition. This is the first attempt to answer this question. Statistically significant results suggest that the dynamics of the individual's motion is the main source of identity information in the most widely used features for gait recognition techniques. The potential of 3D volumetric gait information was also investigated using a Convolutional Deep Belief Network. The trained model was able to estimate the phase during a gait cycle, given a static 3D body volume. The ability of the model to learn high-level representations of 3D body shapes was also validated through an additional experiment, in which the full 3D body shape was successfully filled-in, given a 2.5D visible surface of it. The model was also used for gait recognition, using the collected 3D volumetric data, while introducing invariance to certain data capture conditions. The high-level features that the network learned were visualised in an informative way, showing the ability of the model to capture high-level representations of 3D motion patterns.