Human Action Recognition (HAR) aims to understand and identifies the human action based on features extracted from human poses in a video. The major difficulty in Human Action Recognition is to detect the foreground and identify the actions despite varying background conditions and also to process the action recognition task with the high dimensional data. In this paper, the background subtraction issue is resolved using an adaptive Gaussian Mixture Model which is combined with the contour saliency to detect the efficient silhouettes in dynamic backgrounds and also in identifying the human even if they are latent for two to three frames. Here the system is proposed to introduce an efficient transformation technique named Reduced Variant tSNE (rv-tSNE) to transform the high dimensional feature space data to a low dimensional feature space data. This method is inspired from tSNE where the crowding problem is eliminated but variation in the obtained low dimensional space is high. The proposed algorithm rv-tSNE reduces the variation and eliminates the Data discrimination problem. The proposed system also identifies the actions of two actors performing different existing actions. Finally the actions are recognized and classified using a multi class Support Vector Machine. Experimental results show the higher recognition rate achieved compared to the existing tSNE for various actions using the benchmarking datasets such as Kinect Interaction dataset and Gaming dataset. The proposed Human Action Recognition system finds its application in human-machine interaction, intelligent video surveillance, sports event analysis, and content-based video Retrieval and others.
Subetha, T., & Chitrakala, S. (2018). Silhouette Based Human Action Recognition Using an Efficient Transformation Technique. In Communications in Computer and Information Science (Vol. 804, pp. 153–162). Springer Verlag. https://doi.org/10.1007/978-981-10-8603-8_13