Human activity recognition using machine learning

In our modern world, the robots for service purpose are often equipped with one or more different kinds of sensing devices such as RGB-D cameras. They are required to detect and perceive humans and objects in the environment. In thispaper, we are proposing a novel system for the aim of recognizing the social activities performed by the observed humans from a stream of continuous RGB-D data. The majority number of works till this time have been satisfying in just recognizing the social activities from bounded videos in the relevant datasets. But, in the field of applications for robotics, there is importance for need of more realistic scenarios. In such cases, there is no manual selection of the activities. This is the reason for the usefulness of the detection of time intervals between which the social activities are performed by humans. The recognition of these social activities can have the applications such as to enable the interactions between human and robot; and also for the detection of potentially dangerous situations. In this research paper, a good system is introduced which recognizes social activities from the continuous stream of RGB-D data. This system is a combination of temporal segmentation and classification activities. It also includes a learning model for the proximity-based priors of social activities. A new dataset of RGB-D videos of both individual and social activities is made available. This dataset is used for the evaluation of the solutions that are being proposed. In the results, the good performance of the system is demonstrated, regarding the recognition of social activities from the continuous data of RGB-D sensors.