Clothing-invariant Gait Recognition Using Convolutional Neural Network

Gait recognition is recognizing human through the style in which they walk. However, the recognition task can become complicated due to the existence of covariate factors (e.g. clothing, camera viewpoint, carrying condition, elapsed time, walking surface, etc). Amongst all the covariate factors, clothing is the most challenging one. This is because it may obscure a significant amount of discriminative human gait features and makes it much more challenging for human recognition task. In recent, there has been significant research on this problem. However, conventional state-of-the-art methods have mostly use hand-crafted features for representing the human gait. In this work, we explore and study the use of convolutional neural networks (CNN) to automatically learn gait features or representations directly from low-level input raw data (i.e. Gait Energy Image (GEI)). Evaluations on the challenging clothing-invariant gait recognition of OU-ISIR Treadmill dataset B, the experiment results shows that our method can achieve far better performance as compared to hand-crafted feature in conventional state-ofthe-art methods with minimal preprocessing knowledge of the problem are required.