Human Action Classification Based on Silhouette Indexed Interest Points for Multiple Domains

Abstract

Recent technological advances in the acquisition and dissemination of videos have driven the development of several applications in the context of human action recognition, such as automatic surveillance, strategic planning, crime prevention and traffic monitoring, among others. Despite this large number of applications, several techniques found in the literature are specialized for a particular purpose, working only for a limited scope of actions. To improve the current scenario, this work proposes and evaluates the development of a flexible descriptor and a methodology for identifying human actions in different domains. The classification process utilizes a judgement mechanism for iteratively refining its outcome in order to converge to a decision that best fits the recognizer. Experiments are conducted on five public datasets with different characteristics, from events containing few actions to more complex scenarios involving a large number of people and interaction with objects. Results have demonstrated that the proposed approach provides a proper balance between computational speed and accuracy rate. Therefore, the developed prototype represents a promising tool for real-time applications.