Abstract
One of the biometric methods is human recognition based on ground reaction forces (GRFs) generated during a person's gait. Conventional methodologies for gait recognition have relied on the manual extraction of features from measured GRF signals, in conjunction with the utilization of machine learning algorithms. Recently, convolutional neural networks (CNNs) have become increasingly popular due to their ability to automatically extract features from signal data. However, the CNNs don’t always produce optimal results for human recognition. In this study, we emphasize a novel aspect of the approach: the use of an ensemble of homogeneous CNN classifiers, all sharing the same architecture but trained on different combinations of GRF components. This strategy leverages diversity originating purely from data representation rather than architectural variation, demonstrating that even identical CNNs can complement each other when exposed to distinct training data. The objective of this paper is to design a biometric system that recognizes humans based on GRFs and an ensemble of classifiers, in which the base classifiers will be CNNs. The study utilized a dataset for a total of 5,980 gait cycles from 322 individuals. The architecture of the base classifiers was consistent, and all possible combinations of GRF components were used to train the base CNNs. The optimal results were obtained when all six GRF components were used for CNN training, achieving a recognition rate of 96.57%. Combining seventeen base classifiers into a homogeneous ensemble further improved the performance, yielding a 99.57% correct recognition rate. This demonstrates the effectiveness of ensemble learning with identical CNN architectures in enhancing gait-based biometric recognition.