Fusion of histogram based features for Human Action Recognition

Abstract

The human action recognition (HAR) is framed as a machine learning problem. During HAR, speed of actions, shape of action and background noise play vital role. In this work, to represent speed of action, Bag of histogram of optical flow (BoHOF) is proposed. In this technique, the optical flow is calculated over the segmented human object. Further, the features are thresholded and bagged to compute the BoHOF. Along with BoHOF, sobel edge filter is used in horizontal and vertical direction to remove shadow effect. Median filtering is applied to suppress background noise. Histogram of oriented gradients (HOG) features are extracted from 3D projected planes and combined with BoHOF to extract maximum advantages of both the features. Finally, the multi-class SVM-based classifier with radial basis kernel is applied to recognize different human actions. The experiments are conducted on the benchmark KTH dataset and the experimental findings concludes that the proposed HAR technique provides better performance compared to the state-of-the-art techniques.