A. Gil, L. Payá, O. Reinoso, C. Fernández, and R. Puerto (Spain)
SLAM, Stereo Vision, Visual landmarks, Data Association
We consider the problem of building a map of an unmod ified environment using only visual information extracted from cameras. In order to build a map, we must estimate both robot’s location and a map of its surrounding environ ment. In general, this problem is known as Simultaneous Localization and Mapping (SLAM). It is an inherently hard problem because noise in the estimate of the robot’s pose leads to noise in the estimate of the map and vice versa. Past work on this area has centered on building maps using distance sensors (i.e. laser and SONAR sensors). However, in our case, we propose a method to build a map based only on visual information . While the robot moves along the en vironment, it extracts a number of interesting points from images (i.e. corners) and calculates a relative measurement vector Vr = (Xr, Yr, Zr) to each one of them using stereo vision. We are using SIFT features as the relevant points extracted from images. SIFT features, are said to be in variant to image translation, scaling and rotation and par tially invariant to illumination changes and affine projec tion. In consequence those points are suitable for localiz ing the robot in a particular environment. Our map consist of a number of L three dimensional landmarks referred to a global frame Sg. In addition, each of the 3D landmarks is assigned a SIFT descriptor that enables us to partially dif ferentiate that particular landmark from the rest. Our ap proach to SLAM is based on a Rao-Blackwellised Particle Filter. This permits us to separate the estimation process in two parts: On the one hand, we estimate the path of the robot using a particle filter and estimate the map condi tioned to each path of the robot. We present experimental results that validate our approach to vision-based SLAM in large unmodified environments.
Important Links:
Go Back
