Stereo Vision Based SLAM Issues and Solutions

Abstract

In this chapter we have made an attempt to analyse the issues in stereo vision based SLAM and proposed plausible solutions. Correct sensor modelling is vital in any SLAM implementation. Therefore, we have analyzed the stereo vision sensor behaviour experimentally to understand the noise characteristics and statistics. It was verified that the stereo observations in its natural form (i.e. [u,v,d]) can safely be assumed to represent Gaussian distributions. Then several SLAM implementation strategies were discussed using stereo vision. Issues related to limited field of view of the sensor, number of features, spurious features, noise parameters and nonlinearity in the observation model were discussed. It was shown that the filter inconsistency is mainly due to inherent nonlinearity presence in the small baseline stereo vision sensor. Since UKF is more capable in handling nonlinearity issues than that of EKF, an UKF SLAM implementation was tested against inconsistency. However, it too leads to inconsistencies. This shows that even with implementations that circumvent the critical linearization mechanism in standard EKF SLAM as in UKF, the nonlinearity issue in the stereo vision based SLAM can not be resolved. In order to address the filter inconsistency a more elegant solution is currently being researched based on smoothing algorithms which shows promising results. In conclusion this chapter dwelt on some obscure issues pertaining to stereo vision SLAM and work being done in solving such issues.