Object recognition modular system implementation in a service robotics context

This paper presents an object recognition modular system implementation that allows a humanoid robot to recognize and manipulate objects in a service robotics context. The system consists of three modules: I) Object recognition, II) Object manipulation and III) Humanoid-robot interaction. The Object recognition uses A-KAZE descriptor and Growing Cell Structure (GCS) artificial neural network (ANN) to learn and recognize objects in a faster way and without extra computer resources needed. The Object manipulation allows to detect an object, estimate the 3D position and lift it up with out information about the objects properties. The Humanoid-robot interaction allows an user to ask for an object learned to search only with a label. The validation of the modular system is in two ways: 1) the object recognition module implementation with an own data base (10 objects, 10 images) obtaining results up to 90% and 2) through NAO robot simulation in Webots obtaining results about 71% of correct classification.