Gottschaldt: Implementation of a graph-based SLAM algorithm (Großer Beleg)

19.05.2020, 14.00 pm

Invitation to the presentation of Mr. Paul Gottschaldt

Topic: Implementation of a graph-based SLAM algorithm

Project: Großer Beleg

Supervisors: Gökhan Akgün, Ariel Podlubne, Pedram Amini Rad

Abstract: SLAM means simultaneous localization and mapping and represents a particular group of problems, in which someone wants to estimate both the map of the environment and its current position in it at the same time. This problem remains an open research area for at least three decades, and a solution was long time seen as a “holy grail” for the mobile robotics. In this thesis, a graph-based SLAM algorithm was implemented. Graph-based SLAM is one possible solution to SLAM, which utilizes a graph to represent the problem and later transforms it into an optimization and solves this. It was designed for the needs of the formula student autonomous race car of Elbflorace to be integrated into the autonomous software stack. A SLAM algorithm at Elbflorace is mainly required for a special discipline called ‘trackdrive’, in which the car is required to drive 10 laps as fast as possible on an unknown track. Estimating a map with SLAM in the first round enables the self-driving vehicle to drive on the racing line and drive much faster. The algorithm uses then the provided system state of the car and utilizes the preprocessed observation data of the lidar and camera pipeline. All the data is fused, and a list of all observed track cones is estimated as a map and provided for further usage. Everything needs to be processed online because the car utilizes the results online to follow the track. Because Elbflorace already had an existing FastSLAM solution, this second graph-based approach was mainly implemented to evaluate, which of both approaches performs the best in the environment of our car. Also, the worse approach can be utilized as a backup algorithm, if one approach would fail in the real environment. Both approaches are currently not evaluated on the car itself and could therefore cause problems at an event. The achieved performance of the graph-based SLAM in the estimated map was compared against the previous solution in a Gazebo-based simulation environment called FSSIM. As part of this thesis, the online capability of the graph-based approach was further evaluated, and a profiling of different components of the implementation was done to estimate possible optimization targets. The possibility of hardware acceleration was also shortly revised.