Classification of 40 Different Human Movements with CNN Architectures and Comparison of Their Performance

Abstract

Detection of human movements has become one of the current issues with the developing technology. Recognition of human movements is used in many areas such as security systems, human computer interaction, human robot interaction. Due to the increase in data stored in databases, deep learning methods have recently become one of the most frequently used methods. At this study, it is aimed to classify human movements by using Convolutional Neural Network (CNN) architectures. Images are classified with InceptionV3, Googlenet and Alexnet architectures using a data set with 40 different motion classes. The highest accuracy rate with 76.15% was obtained in InceptionV3 architecture. Increasing the amount of data in CNN networks is a parameter that closely concerns the network uptime. Since 40 different motion classes are used in this study, the results obtained in the related architectures are obtained in different times.