An Approach to Estimate the Orientation and Movement Trend of a Person in the Vicinity of an Industrial Robot

Implementation of industrial robots worldwide -not only typical caged robots, manipulating heavy machinery and performing repetitive tasks, but also collaborative robots-, comes with the challenge of guaranteeing the safety of human operators, whether they work aside from the robot or in its vicinity. The UNE-EN ISO10218 safety standards for industrial robots set that robots and people can work in common spaces if robots have safety devices or are supported by them, to avoid hurt human operators. This paper introduces an approach that allows estimating the orientation and movement trend of a person in the vicinity of a robotized industrial task, aiming to avoid the collision between human and robot. As a previous requirement of the approach, a PointNet architecture was trained with the point clouds obtained from Depth images of a proprietary RGB-D image dataset. This task, in turn, required detecting people in the corresponding RGB images, through the application of a pre-trained saliency algorithm. To estimate the orientation of the detected person, a modified Biternion network was trained with the resized images from the same proprietary database. At evaluating the system, depth images captured by a Kinect sensor were used as inputs, then, for a set of four iterations (four frames), the movement trend was calculated since the orientation of the person was known for every frame. The prediction capability of the proposed approach was evaluated with three groups of images and resulted in a general precision greater than 0.5.