Object Recognition of the Robotic System with Using a Parallel Convolutional Neural Network

Abstract

The work relates to research in the field of creating a collaborative robotic system—an assistant surgeon. At the request of the surgeon calling the required tool, the robot must find this tool and pass it onto the surgeon. This article covers only the part of the task related to the search by the robot of the desired tool. The task is performed using a manipulation robot, in the grip of which the camera is located. Information processing is carried out using a convolutional neural network, which includes two parallel convolutional neural networks. One of them highlights the areas of the desktop, on which the individual tools lie, highlights the individual areas characteristic of each tool, and determines the position and orientation of the tool necessary for its capture by the robot. The second one determines the type and name of the tool. The procedure of training a neural network is considered, after which the system is able to solve the “coordinate change” problem, etc. by the name of the tool, immediately find it on the desktop and determine its position and orientation. The results of experiments confirming the high efficiency of the proposed method are given.