Abstract
The paper presents an approach for feature selection in human activity recognition. Features are extracted based on spatiotemporal orientation energy and activity template, while feature reduction has been studied thoroughly using various techniques. Due to high dimensional data from extraction phase, a model with less features which are important and significant can build attractive, interpretative and accurate model. Finally, activity classification is done using SVM. With experiments to classify six activities of the KTH Dataset, significant feature reductions were reported with optimal embedded selection recorded for Gradient Boosting and R-Square techniques. The results show a reduction in time and improvement in accuracy. The Comparison to related work were given.