Instructor: Trym Vegard Haavardsholm

All students should have access to the TTK21 blackboard site, where all announcements will be made.

This course gives a comprehensive introduction to visual simultaneous localisation and mapping (VSLAM), the art of using cameras to navigate the world by simultaneously building maps and estimating camera motion using these maps.

To build a map of the environment through exploration, we need to extract a good representation of the structure we observe. But in order to build a consistent map while we are exploring, we also need to know how we move through the environment. Combining local odometry with global map estimation is often called simultaneous localisation and mapping (SLAM). The SLAM map is based on additional constraints from recognising places that are revisited (loop closures), which is crucial for limiting the localisation error over time and building consistent maps with the correct topology.

Visual sensor systems are well suited both for measuring relative motion and recognising places. This means that these sensors can perform both short-term tracking for real-time pose estimation, as well as long-term tracking for detecting loop closures and making correspondences with previous maps. Systems based only on short-term tracking and mapping can be considered as reduced SLAM systems, and are often referred to as visual odometry (VO) systems. Combining visual and inertial sensors is advantageous because it makes the direction of gravity and scale observable, while also guiding image tracking and observing relative movement when image tracking fails. Systems based on this sensor combination are often called visual-inertial.

This course will cover how such systems work.

After completing this course:

• you should understand the fundamental technology that current VSLAM systems build upon.
• you should understand the details in how some of the most influential VSLAM systems work.
• you should be able to understand how to apply VSLAM technology in your own work.