Fusion based Gait Recognition System using Ensemble Deep Learning for Biometric Security in Video Surveillance System

Abstract

Gait recognition has long been recognized as a potential biometric that is non-offensive and
capable of identifying a person from a distance. A novel ensemble learning approach for gait
recognition, namely multi-source gait recognition based on classifier fusion, is proposed in this
study. The vast majority of recent studies have relied on isolated deep learning systems for
tasks like data fusion and categorization. Effective data fusion can be achieved through the use
of a multisensor fusion system and an ensemble of deep learning algorithms. In contrast, a deep
learning ensemble approach and a decision fusion architecture-based multisensor classification
method for fusing gait posture images (GPI) and cumulative foot pressure images (CFPI) are
investigated in this article. We introduce FuseGaitNet, a novel deep convolutional neural
network structure, in this article. An SVM is used in place of the softmax layer in a DCNN.
After that, we use a random feature-selection-based DCNN-SVM ensemble system to generate
separate models for GPI data and CFPI data. We employ two diversity metrics to discover the
most diverse combinations of classifiers from the two ensemble systems in order to prevent
overfitting and closeness between the deep features and the classifiers. Finally, the acquired
different classifiers from CNN ensembles are combined using a decision fusion method. The
suggested architecture was created and tested with the renowned CASIA-D gait dataset.
Outcomes present that the suggested method works well compared to available methods and
achieves higher accuracy.