Defesa de Tese de Doutorado – Seyed Jamalaldin Haddadi – 20/5/2021

11/05/2021 16:10
Defesa de Tese de Doutorado
Aluno Seyed Jamalaldin Haddadi
Orientador Prof. Eugênio de Bona Castelan Neto, Dr. – DAS/UFSC


20/5/2021  13h  (quinta-feira)

Videoconferência (




Prof. Eugênio de Bona Castelan Neto, Dr. – DAS/UFSC (presidente);

Prof. Guilherme Vianna Raffo, Dr. – UFMG;

Prof. Henrique Simas, Dr. – EMC/UFSC;

Prof. Leandro Buss Becker, Dr. – DAS/UFSC.

Título Improvement of Visual-Inertial ORB-SLAM using Correction in State Estimation
Abstract: The fusion of monocular visual and inertial has gained a lot of attention from robotic systems. Recent results have shown that optimization-based fusion approaches outperform filtering approaches. However, poor initialization can lead to inaccurate state estimation in optimization-based visual-inertial Simultaneous Localization and Mapping (SLAM) systems. Therefore, due to the nonlinearity of visual-inertial systems, initial values (visual scale, gravity, velocity, and Inertial Measurement Unit (IMU) biases) play a crucial role. For this reason, this thesis aims to improve the initial state estimation using two adjustable parameters. Based on the difficulty level of the environment (texture, mid texture, and texture-less), the indoor room is categorized into easy, medium, and difficult, and then two adjustable parameters are regulated based on this difficulty level. This strategy has been tested in two types of implementation; Benchmarking with the public EuRoC dataset and real-world experiment. In benchmarking, by employing the right adjustable parameters, in some scenarios, we could attain satisfactory results compared to stateof-the-arts visual-inertial Odometery and SLAM in terms of positioning accuracy and reduction of accumulative error. In this part, also, from point of hardware’s view, some measurements are performed. While the proposed algorithm is being executed, the maximum CPU usage in each sequence is measured on a Raspberry-Pi singleboard and a Laptop. The results proved that the Raspberry-Pi 3 – because of poor hardware configuration – is under more pressure in terms of CPU usage. The second part is concerned with a real-world experiment in which a monocular-inertial RealSense ZR300 sensor is utilized. The outcomes were satisfactory so that the initialization time was very short and the proposed algorithm could quickly obtain the ORB features.