Assessing the Performance of Handcrafted Features for Human action Recognition

Abstract

Recognition of Human action such as running, punching, bending, kicking etc. plays an vital role in futuristic applications like intelligent video surveillance, health care monitoring, robotics, smart automation system, computer gaming etc. This field relies on various approaches based on hand crafted features like PCA, HOG, LBPH, DWT, STIP, SWF, SWFHOG and deep learning techniques like CNN, RNN and their variants. Though many approaches have been proposed and implemented by researchers, the literature survey suggests that a detailed understanding of the approaches and a comparison of advantages and limitations is required to develop more accurate action recognition method. This paper focuses on this issue and gives detailed analysis of results obtained by implementing algorithms on standardize open source datasets of varying complexity namely Weizmann, KTH, UT Interaction and UCF sports.  The results are compared based on the classification accuracy as it is one of the performance measure for checking reliability of the method. The comparison shows that, SHFHOG feature gives the best classification accuracy as compared to other handcrafted features and also outperforms the simple CNN.