Indoor Visual Odometry Algorithm Based on Adaptive Feature Fusion

An indoor scene has both texture-less regions and dense repeated texture regions. When the VO algorithm based on image feature points is adopted to conduct positioning and mapping, the positioning accuracy will be low. Even the tracking will be missed due to problems such as insufficient effective feature points and too many mismatches. In this paper, we proposed a visual odometry algorithm based on adaptive feature fusion to solve the problems. The algorithm extracts both line and point features and matches them precisely using GMS (Grid-based Motion Statistics) algorithm and RatioTest algorithm, respectively. Then the two features are fused as per adaptive thresholds. The algorithm in this paper is adopted in the visual odometry section of ORB-SLAM2, and the result obtained from 18 sets of TUM data set sequences with different structures, and texture features are compared with ORB-SLAM2. The experimental result shows that in the dense repeated texture regions, the absolute trajectory error of ORB-SLAM2 is decreased by 8.87%. In the texture-less regions, the absolute trajectory error is reduced by an average of 9.96%. The positioning accuracy and robustness are effectively improved thanks to the algorithm.