A New Mechanism for Collision Detection in Human–Robot Collaboration using Deep Learning Techniques

Human–robot collaboration is increasingly present not only in research environments, but also in industry and many contemporary day-to-day activities. There is a need for the automation of tasks ranging from the simplest to the most complex ones. The insertion of robotic arms provides a considerable step useful in achieving this goal. In this context, safety remains a concern, however. Among the most frequent issue in this collaboration context is human–robot collision. While focus has been on automation efficiency of the of the activities, there is a growing need to reduce or even prevent damage to the involved agents. As part of this goal, a new mechanism for detecting human–robot collisions is proposed in this article. It has been tested in a well-controlled scenario using equipment commonly present in collaborative scenarios for maintenance on a radio base station. The robot used is a UR5 robotic arm in addition to three 2D cameras and a network rack. In our experimental scenario, a person interacts with the network devices installed within the rack while conducting basic collaborative activities inserted in this context. For collision detection, deep learning models were used and evaluated. These were trained to detect overlap between humans and robots considering the view and perspectives from three different cameras. Finally, a new ensemble learning system is proposed in order to establish whether or not a collision took place. It receives as input the result of overlap detection through deep learning models. Results suggest that the proposed system is capable of detecting collisions with an average global accuracy of 89.81% of correctness in a well-controlled scenario. The effectiveness of the proposed ensemble is exposed in comparison with the use of only one of the cameras for decision-making. Finally, the proposed system is shown to detect collisions in real time and to achieve a low response time.