Monocular SLAM Initialization Using Epipolar and Homography Model

The performance of the initialization module for any visual Simultaneous Localization and Mapping (SLAM) algorithm is crucial in determining the overall accuracy of the navigation solution. This paper proposes an automatic real time initialization module based on a heuristic algorithm which is a model selection method that provides robustness for both non-planar and planar initial scenes using the epipolar and homography geometry respectively. Initialization is made more robust by utilizing features comprising of lines, edges and corner points in a scene adaptive method so that there is a balance between the trade-off of computation speed and accuracy of the algorithm. The proposed algorithm has been effectively used to initialize map and showed to improve the pose estimation and tracking of the state of the art visual odometry algorithm by simulation over benchmark datasets. The executable code implementation of the proposed initialization module has been made open source publically to provide as a valuable reference for the students of SLAM at https://github.com/maazmb/LEP-Hybrid-Visual-Odometry.git.