Real-Time Action Classification using Intermediate Skeletal Pose Estimation

Abstract

The topic of Human Action Classification has attracted significant interest these past years, which can be attributed to the advancements in deep learning methodologies. Its applications range from robotics to surveillance and automated video categorization, as well as in the healthcare industry. The literature, however, has primarily focused on offline action classification, without significant attention being given to the constraints of an online real-time classifier. Using an intermediate skeletal representation of humans, while convoluted, provides a scalable means of tackling the action classification problem. This project discusses the existing literature, and adapts two of the state-of-the-art approaches for real-time analysis. An extensive analysis is performed to distinguish the advantages and drawbacks of each model. Lastly, an end-to-end system is implemented to experiment the efficacy of the classifiers in real-world environments.