Constructing a Global Representation from Multiple Stereo Images
Gardner, Warren F.
MetadataShow full item record
The problem of using passive vision to construct 3-D descriptions of the environment is commonly encountered in computer vision and robotics. This paper focuses on the problem of merging multiple stereo images into a single global representation. The problem of how to fuse multiple data measurements is one which arises frequently in the field of robotics and computer vision. This type of problem is encountered while attempting to interpret data from a single source, but becomes more apparent when multiple sources or sensors are involved. This paper will focus on a tool, the Extended Kalman Filter (EKF), which is capable of fusing multiple erroneous measurements while reducing the uncertainty associated with these measurements. Section 2 introduces both the Kalman Filter and the EKF. The Kalman Filter is a linear, recursive, mean-square error filter. The EKF is based on the Kalman Filter, but is used for estimating nonlinear systems. Section 3 applies the EKF methodology to the problem of estimating the displacement between a pair of stereo images. A method of matching lines between images is discussed within the EKF framework. Section 4 introduces the error models used by the EKF. In Section 5 the problem of merging stereo images given an accurate estimate of their displacement is discussed. Section 6 examines the problem of divergence. The results of an implementation of the above systems is presented in Section 7. Section 8 contains recommendations for future research in the area of 3-D reconstruction with the EKF.