Long range computer vision for robotics

Bc. Michal Jankovič

Master's thesis

Abstract:

Práca rieši úlohu merania hĺbky na veľké vzdialenosti za použitia jednej kamery s motorizovaným zoom objektívom a skenovania nastavenia zaostrenia. Popisujeme vytvorenie datasetu zaostrovacích sekvencií s ground truth hĺbkovou anotáciou z LiDAR senzora na vzdialenosti do 100 metrov. Na našom datasete vyhodnocujeme presnosť viacerých metód na predikciu hĺbky zo zaostrovania, vrátane metód hlbokého učenia s konvolučnými neurónovými sieťami, ktoré predtrénujeme na syntetických dátach s umelou hĺbkou ostrosti. Analyzujeme vplyv viacerých faktorov na presnosť a dosah metód, vrátane úrovne zoomu a skenovanom rozsahu nastavení zaostrenia.

Abstract:

We solve the task of long range depth estimation using a single camera with motorized zoom lens by scanning over a range of focus settings. We create a dataset of multi-focus sequences with ground truth depth annotation from a LiDAR sensor, targeting distances up to 100 meters. We evaluate the accuracy of a number of methods for depth prediction from focus on our dataset, including deep learning with convolutional neural networks, which we pretrain on synthetic data with artificial depth of field. Impact of factors such as zoom level and the range of scanned focus settings on accuracy and range is analyzed.

Keywords

Depth Estimation, Depth From Focus, Deep Learning

Language used: English

Date on which the thesis was submitted / produced: 13. 12. 2022

Thesis defence

• Date of defence: 27. 1. 2023
• Supervisor: prof. Ing. Václav Přenosil, CSc.
• Reader: doc. Jan Mazal, Ph.D.

Masaryk University

Faculty of Informatics

Master programme / field:
Visual informatics / Image Processing and Analysis