Abstract
Extracting salient and most prominent features from a given video sequence is a critical step in Human Action Recognition (HAR). The work presented within this article proposes a new method for HAR, which efficiently addresses the issue of robust feature selection. The proposed method initially fuses three different feature categories based on their highest values, and later selects most optimal features using a novel Euclidean distance (ED) and strong correlation (SC) methods. Finally, it classifies the selected features using multi-class classifier. For experimentation, four publically available datasets including Weizrnann, KTH, UCF YouTube, and HMDB51 are used and results with improved classification accuracy, on average more than 94%, are obtained. Experimental results validate that the proposed approach outperforms the existing techniques.