Kinect-based Gait Recognition Using Deep Learning

Abstract

Accurate gait recognition is of high significance for numerous industrial and consumer applications, including finance, virtual reality, online games, medical rehabilitation, collaborative space exploration, and others. This thesis proposes two novel deep learning architectures for a highly accurate and robust Kinect-based gait recognition. First, the architecture of the Deep Learning Neural Network (DLNN) is developed using two unique view and pose invariant geometric features. Second, the end-to-end training of a residual learning-based convolutional neural network, named KinectGaitNet, is proposed to enable even higher recognition performance without the necessity of extracting domain-specific handcrafted features. The performance of the DLNN architecture and KinectGaitNet are evaluated on two publicly available 3D skeleton-based gait datasets recorded with the Microsoft Kinect sensor. It is experimentally proven that the accuracy, precision, recall, and F-score of the DLNN architecture and KinectGaitNet are superior to the recent state-of-the-art methods for the Kinect skeleton-based gait recognition.