04 May 2020
14:00 Master's Defense Fully distance
Theme
LIFT-SLAM: the deep-learning feature-based visual SLAM method
Student
Hudson Martins Silva Bruno
Advisor / Teacher
Esther Luna Colombini
Brief summary
The problem of simultaneous location and mapping (SLAM) addresses the possibility of a robot locating itself in an unknown environment and, simultaneously, creating a consistent map of that environment. One of the main components of SLAM, called Odometry, is responsible for estimating the location of the agent and changes in position over time. Recently, cameras have been used successfully to obtain the characteristics of the environment to perform SLAM and Odometry, which are called visual SLAM (VSLAM) and visual Odometry (VO), respectively. However, the classic VO and VSLAM algorithms can easily be induced to fail when the robot's movement or the environment is very challenging. Recently, new approaches based on deep neural networks (DNNs) have achieved promising results in VO / VSLAM. Thus, we propose to combine the potential of feature descriptors based on deep learning with the traditional VSLAM based on geometry, creating a new VSLAM system for mobile robots. The experiments carried out on the KITTI and Euroc data sets show that the proposed approach was able to achieve results in the state of the art.
Examination Board
Headlines:
| Esther Luna Colombini | IC / UNICAMP |
| Paula Dornhofer Paro Costa | FEEC / UNICAMP |
| Paulo Lilles Jorge Drews Junior | C3 / FURG |
Substitutes:
| Hélio Pedrini | IC / UNICAMP |
| Alexandre da Silva Simões | ICTS / UNESP |