A Novel Action Descriptor to Recognize Actions from Surveillance Videos

Abstract

Due to the increased application in the area of human action detection, automatic capture and action analysis becomes a great research area. This paper provides a novel method to detect human action by considering features from the positive space and negative space region. Usually, in the literatures features have been considered from the positive space of the subject. Positive space features alone cannot provide solutions to the challenges such as occlusion and boundary variations. Therefore, in this method we have also considered the surrounding regions of the subject along with the positive space features. Initially, the input video has been segmented using background subtraction. Then the features are extracted from both positive and negative space of the subjects. Later, action descriptor has been defined for each pose of an action and proposed a new way for detecting number of cycles required to describe an action. Later, nearest neighbor classifier has been applied to classify the actions. The proposed system is evaluated using Weizmann dataset, KTH dataset and the results seem to be promising.