PROJECT TITLE :

On Learning Disentangled Representations for Gait Recognition

ABSTRACT:

One of the crucial modalities in biometrics is known as an individual's gait, which refers to their walking pattern. Gait features are typically represented as silhouettes or articulated body models in the majority of the existing gait recognition methods. When dealing with potentially confusing factors like clothing, carrying, and viewing angle, these methods have a performance issue that hinders their ability to recognize objects. In order to solve this problem, we have developed a novel Autoencoder framework that we call GaitNet. This framework can explicitly separate appearance, canonical, and pose features from RGB imagery. While canonical features are averaged in order to create a static gait feature, the LSTM integrates pose features over the course of time to create a dynamic gait feature. Both of them are used as classification features in various contexts. In addition, we collect what is known as a Frontal-View Gait (FVG) dataset in order to concentrate on gait recognition from frontal-view walking. This is a difficult problem to solve given that this view of walking contains fewer gait cues than other views. FVG also takes into account variations such as walking speed, carrying, and clothing, amongst other significant aspects. Our method demonstrates superior performance to the SOTA quantitatively, the ability of feature disentanglement qualitatively, and promising computational efficiency. These results come from extensive experiments performed on the CASIA-B, USF, and FVG datasets. We continue by contrasting our GaitNet with the most advanced form of Face Recognition in order to demonstrate the benefits of using gait biometrics for identification in specific contexts. These contexts include long distances with lower resolutions and cross viewing angles.