Activity recognition of interacting people

Abstract

With the widespread use of depth sensors, the recognition of human activities, especially in human-robot interaction, is of interest to researchers. The purpose of this work is to automatically recognize human activities using the joint coordinates of the three-dimensional skeletons obtained from the depth sensor. Our method computes the features to be used in classification automatically by deep learning methods. The results obtained are much better than the methods of recognizing human activity with hand-crafted features. In this work, we used a data set with multiple people in the images, allowing us to explore interactive human activities. Two different types of deep learning techniques and performance analysis were performed using different architectures. As a result of the experiments performed, it is seen that the network trained from scratch classifies with highest performance.