ACTION RECOGNITION USING FEATURE FUSION

Abstract

Action recognition seems to be a very easy task for us humans but it requires a lot of information processing in terms of recognizing patterns when it comes to computer systems. Here, we try to devise a new way of action recognition for intelligent systems by fusing the shallow and the deep features from the data. Shallow feature extraction starts by identifying the motion salient pixels first, thus eliminating unwanted information and then extract the improved trajectory information from it. To get the deep features, we make use of Convolutional Neural Network (CNN). There will be separate classifiers for both the deep features and shallow features which will be fused in order to result in an efficient classifier for the action recognition. We are using HMDB-51[1] video dataset, one of the most challenging datasets for action recognition which consists of various actions of different kinds like clap, run, walk, box, etc taken from various sources like YouTube, movies and Google videos under various illumination effects, occlusion, camera angle variation and pose variation