Deep Learning for Safe Human-Robot Collaboration

Recent advances in computer vision and deep learning have lead to implementations in different industrial applications such as collaborative robotics, making robots able to perform harder tasks and giving them consciousness of their environment, easing interaction with humans. With the objective of eliminating physical barriers between humans and robots, a security system for industrial collaborative robots based on computer vision and deep learning is proposed, where an RGBD camera is used to detect and track people located inside the robot’s workspace. Detection is made with a previously trained convolutional neural network. The position of every detection is fed to the tracker, that identifies the subjects in scene and keeps record of them in case the detector fails. The detected subject’s 3D position and height are represented in a simulation of the workspace, where the robot’s speed changes depending on its distance to the manipulator following international safety guidelines. This paper shows the implementation of the detector and tracker algorithms, the subject’s 3D position, the security zones definition and the integration of the vision system with the robot and workspace. Results show the system’s ability to detect and track subjects in scene, and the robot’s capacity to change its speed depending on the subject’s location.