Robust human action recognition system for closely related actions in video

Abstract

With the availability of high tech electronic gadgets ultra fast Internet access and huge storage spaces at negligible cost the corpus of video that is accessible has grown tremendously over the last few years Simultaneously the demand for understanding of these videos has also exponentially increased The limited human capabilities of analyzing them in a natural way have necessitated the presence of intelligent systems that could analyze and recognize activities occurring in videos In essence the main goal is to automatically understand the action performed by human in videos and assign semantic labels to the video clips This process tries to bridge the semantic gap between low level representation and the high level descriptions given by humans This task is both challenging and compute intensive due to scale and illumination invariant partial/full occlusions the high dimensionality of poses cluttered background intra class variance and inter class similarity Recent progression in either handcrafted or deep learning methods extensively improved action recognition accuracy But there are still many open issues which keep action recognition task far from being solved Conventional methods for human action recognition normally use a local feature along with bag of words to model actions However due to the impacts of noise and camera movement using the bag of words model to obtain promising performance is still an arduous task newline