An efficient classifier using machine learning technique for individual action identification

Abstract

Human action recognition is an important branch of computer vision and is getting increasing attention from researchers. It has been applied in many areas including surveillance, healthcare, sports and computer games. This proposed work focuses on designing a human action recognition system for a human interaction dataset. Literature research is conducted to determine suitable algorithms for action recognition. In this proposed work, three machine learning models are implemented as the classifiers for human actions. An image processing method and a projection-based feature extraction algorithm are presented to generate training examples for the classifier. The action recognition task is divided into two parts: 4-class human posture recognition and 5-class human motion recognition. Classifiers are trained to classify input data into one of the posture or motion classes. Performance evaluations of the classifiers are carried out to assess validation accuracy and test accuracy for action recognition. The architecture designs for the centralized and distributed recognition systems are presented. Later these designed architectures are simulated on the sensor network to evaluate feasibility and recognition performance. Overall, the designed classifiers show a promising performance for action recognition.