Real-Time Structure from Motion Using Kalman Filtering
Real-Time Structure from Motion Using Kalman Filtering
project thesis by Jeannette Bohg
- Supervisor Großer Beleg
- Wissensverarbeitung
- 21 März 2005 – 21 März 2005
- Download
robotics. Its central theme is the estimation of the location and motion of a
mobile robot situated in an unknown environment and the simultaneous construction
of a map of this environment.
In this work, the problem is tackled by using vision sensors. Therefore,
classical Structure From Motion (SFM) techniques, which have a long history
in computer vision, are applied to SLAM. In general, we can distinguish two
major approaches to SFM. Firstly, we have techniques based on the recursive
Kalman Filter where time dependend sequences of images are required. The
second kind of SFM-methods utilizes the constraints in multiple view geometry.
The sequence of the images considered is independent of time.
One of our main objectives is to obtain real-time performance which is required
in the field of robotics rather than offline processing.
We build upon the results by Tobias Pietzsch [24] who presented promising
results using the second kind of SFM algorithms specifically tailored for real-time
performance.
In this work we will analyze the usage of Kalman Filtering concerning the
task of SLAM and compare it with methods that utilize the multiple view constraints.
As a sensor we will first use a single camera and then compare the
results with the performance of a stereo camera. We track 3D point features as
landmarks.
