Real-Time Structure from Motion Using Kalman Filtering

Aus International Center for Computational Logic
Wechseln zu:Navigation, Suche
Toggle side column

Real-Time Structure from Motion Using Kalman Filtering

Studienarbeit von Jeannette Bohg
Simultaneous localization and mapping (SLAM) is one of the core problems in

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.