Prepoznavanje objektov v delovnem prostoru industrijskega robota s pomočjo globinske kamere

Bratina, Matic (author), Vrabič, Rok (mentor), Bračun, Drago (comentor)

V okviru magistrske naloge je bil nadgrajen projekt, ki združuje načrtovanje robotovega gibanja in prepoznavanje objekta z uporabo globinske kamere. Z globinsko kamero zajemamo barvno in globinsko sliko delovnega okolja robota. Na zajetih slikah uporabimo naučen model konvolucijske nevronske mreže, s katerim zaznamo ciljni objekt. Prepoznano območje analiziramo in določimo položaj ter orientacijo predmeta v prostoru. Z uporabo konvolucijskega modela Dex-Net na prepoznanem objektu določimo potencialna prijemna mesta in njihovo primernost. Lokacije prijemov in položaj predmeta so uporabljeni za načrtovanje gibanja. Načrtovanje gibanja robota Franka Emika Panda je izvedeno s pomočjo programskega paketa MoveIt in razčlenjeno na niz podnalog. Če načrtovalec gibanja najde rešitev za prijemanje in premikanje predmeta, je to realizirano v virtualnem okolju z robotskim simulatorjem Gazebo.

• Language: Slovenian
• Keywords: prepoznavanje objektov, industrijski roboti, robotski vid, pobiranje objektov v razsutem stanju, konvolucijske nevronske mreže, robotski operacijski sistem (ROS), MoveIt, Gazebo
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FS - Faculty of Mechanical Engineering
• Place of publishing: Ljubljana
• Publisher: [M. Bratina]
• Year: 2021
• Number of pages: XX, 72 str.
• PID: 20.500.12556/RUL-124539
• UDC: 007.52:004.932:004.896(043.2)
• COBISS.SI-ID: 51334915
• Publication date in RUL: 30.01.2021

BRATINA, Matic, 2021, Prepoznavanje objektov v delovnem prostoru industrijskega robota s pomočjo globinske kamere [online]. Master’s thesis. Ljubljana : M. Bratina. [Accessed 5 April 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=124539

Secondary language

• Language: English
• Title: Object recognition in the workspace of an industrial robot using a depth camera

Abstract

The master's thesis presents a project that combines industrial robotmotion planning and object recognition using a depth camera. The depth camera captures a colour-and-depth image of the robot's workspace. A convolutional neural network is then used to detect the target object in the captured images. The identified area is analyzed and the position and orientation of the object in space are determined. Using the Dex-Net convolution model, potential gripping points and grip probabilities are determined and evaluated. The locations of the grips and the position of the object are communicated to the motion planner. Motion planning of Franka Emika Panda robot is performed using the MoveIt software framework and broken down into a series of subtasks. If the motion planner finds a solution for the pick and-place movement, the movement is realized in a virtual environment based on Gazebo robotic simulator.

• Keywords: object recognition, industrial robots, robot vision, bin picking, convolutional neural networks, robot operating system (ROS), MoveIt, Gazebo