Skip to Main Content
Feature subset selection is an important preprocessing step for pattern recognition, to discard irrelevant and redundant information, as well as to identify the most important attributes. In this paper, we investigate a computationally efficient solution to select the most important features for gait recognition. The specific technique applied is based on mutual information (MI), which evaluates the statistical dependence between two random variables and has an established relation with the Bayes classification error. Extending our earlier research, we show that a sequential selection method based on MI can provide an effective solution for high-dimensional human gait data. To assess the performance of the approach, experiments are carried out based on a 73-dimensional model-based gait features set and on a 64 by 64 pixels model-free gait symmetry map on the Southampton HiD Gait database. The experimental results confirm the effectiveness of the method, removing about 50% of the model-based features and 95% of the symmetry map's pixels without significant loss in recognition capability, which outperforms correlation and analysis-of-variance-based methods.