Recognition and Classification of Human Actions Using 2D Pose Estimation and Machine Learning

Abstract

Recognition and classification of human actions is a fundamental but difficult computer vision task that has been studied by several researchers worldwide in recent years. Pose estimation is a widely used technology for recognizing human actions. It has several applications, especially in the field of computer vision, where it can be used to recognize basic as well as complex human actions. This research provides a novel framework for identifying and classifying human actions which include five categories: standing, walking, waving, punching, and kicking. The dataset used for recognition and classification purposes is generated using the videos that are recorded using a smartphone and a 2D pose estimation technique has been applied to extract the features from the human body. The machine learning (ML) classifiers have been trained on a custom-built dataset. While all algorithms nearly performed well in the classification task, the light gradient-boosting machine (LGBM) outperformed the rest in terms of accuracy (98.80%).