Automated Drinking Action Recognition in videos

Abstract

Today, human action recognition has been becoming one of the famous research areas in Computer Vision. This interest is motivated by a wide set of applications including security applications, video surveillance, video indexing, robot learning, etc. In terms of action recognition, recognizing actions that involve objects is a di cult task in current action recognition approaches. This research is to recognize such actions in another way to improve the accuracy by combining object detection and human action de- tection. In order to recognize the drinking action, a robust descriptor has been created to represent the body action by combining motion descriptor which has been calculated from optical ow and appearance descriptor which has been calculated from still video frames. This descriptor has been used in the action classi cation process. After detecting the action, object detection is conducted to verify the drinking action. Video clips containing drinking actions which have been extracted from movies, internet and some data sets were used in the training, testing and evaluation phases. In the evaluation, datasets were tested separately with action detection and object detection to identify the e ect of the veri cation process.