An Approach to 3D Object Detection in Real-Time for Cognitive Robotics Experiments

Abstract

This paper presents a computer vision method that, taking information from an RGB-D camera, performs real time 3D object recognition to be used in cognitive robotics experiments, where the real time constraints are key. To this end, we have implemented and tested an algorithm that combines a deep neural network (YOLOv3 tiny) that processes RGB images and provides object recognition and 2D localization, with a point cloud analysis method to compute the third dimension. The proposed approach has been tested in real-time manipulation experiments with the UR5e robotic arm through ROS, and using a GPU to execute the method, showing that this combination allows for an efficient real-time learning using cognitive models.