Entropy-based action features selection using histogram intersection kernel

Current approaches of local spatio-temporal interest point detections provide compact but descriptive representations for human action recognition. However, unavoidable noisy points interfering with video representation lead to bringing the accuracy of recognition down. This paper proposes an efficient approach to select human action features in videos. We combine entropy with histogram intersection kernel incorporating method of feature distance measurement in similarity to compute histogram significance. The accuracy of our method tested on the KTH dataset using 3D-Harris detector and 3D-HoG descriptor is 83.52%. Experimental results demonstrate that our method with distance of Histogram Intersection to build visual code words has a positive impact upon selecting features which are beneficial to classification.