An Integrated Feature Selection Strategy for Monocular Slam

For feature-based monocular bearing-only SLAM, how to select useful features for SLAM process is crucial. The reason is overwhelming feature number will not only seriously slow down the system but produce erroneous SLAM result due to feature mismatching. In this paper, we propose a novel method for feature selection. The method combines both bottom-up (visual saliency) and top-down (learned object database) approaches to select versatile features. We argue that using human’s visual saliency to guide the robot’s visual SLAM feature selection is practicable. The experimental result after 10 runs attested our perspective. Compared with SLAM without feature selection, the running time here is reduced to 62% and the localization errors in the SLAM process decrease to 89% in mean and 89% in standard deviation.