Semantic human activity detection in videos

Many solutions have been proposed for human action detection in the past. Even though, almost all the solutions address only the detection of basic human activities such as 'shaking hands', 'sitting down' etc and all of them are based on the structure of the activity pattern. No considerable attention has been paid to detect more semantic activities (more meaningful activities) like 'smoking', 'fighting', 'riding', etc. Therefore existing solutions are not capable of identifying such semantic activities accurately. There are three main reasons behind this inability. First one is most activities do not have any identifiable common action structure in it ('talking'). Secondly even when there is such an identifiable structure that activity pattern does not follow every single instance of activity performing ('smoking'). Third reason is some activities are too complex to identify using such basic action pattern analyses approaches ('hurdling'). Nevertheless ultimate expectation of human activity detection is identifying more complex/meaningful activities. Therefore, it is essential to address this problem properly for implementation of more useful applications in the future. In this paper, we urge the importance of using contextual information associated with semantic activities to overcome above mentioned three problems.