Abstract
New bio-inspired sensors that measure brightness changes per-pixel have the potential to become
a novel solution to the pose estimation problem. The sensor generates a stream of events that
represent the position and the polarity of intensity change. This means that it can still work
well under the low light condition, even with unstable movement. Currently, event-based vision
sensors output compressed digital data in the form of events, reducing latency and having
higher temporal range than conventional image-based methods. These event-based cameras
encode visual information in an extremely efficient manner in terms of data reduction and energy
consumption. This is especially important in the field of localization, where responsiveness is
one of the most significant properties.
In this thesis, we explore approaches to perform feature detection directly on the event stream
without intermediate event accumulation. Event flows are obtained to track lines. We introduce
ASL-SLAM, the first line-based SLAM system operating directly on asynchronous event streams.
This approach maximizes the advantages of the event information generated by a bio-inspired
sensor. We estimate the local Surface of Active Events (SAE) to get the space-time planes
associated with each incoming event in the event stream. Then the edges and their motion are
recovered by our line extraction algorithm. We show how the inclusion of event-based line
tracking significantly improves performance compared to state-of-the-art frame-based SLAM
systems. The approach is evaluated on publicly available datasets. The results show that our
approach is particularly practical with poorly textured frames. We also experimented with
challenging illumination situations, including low-light and high motion blur scenarios. We
show that our approach with an event-based camera has natural advantages and provides up
to 85% reduction in error when performing SLAM under these conditions compared to the
traditional approach.