Feature Points Extraction Based on Improved ORB-SLAM

Abstract

In ORB-SLAM tracking threads of mobile robots, feature points extraction based on fixed threshold can cause large overlap and uneven distribution of feature points, and the number of feature points change sharply with the brightness changes. To solve this problem, a feature point extraction algorithm based on local adaptive threshold was proposed. Firstly, the algorithm filtered feature points by setting adaptive parameters and calculating each pixel threshold with dynamically local threshold, thus improving the extraction method of ORB feature points based on fixed threshold, overcoming corner loss or redundancy caused by improper selection of fixed threshold. Finally, by classifying the brightness of pixels, candidate feature points were screened layer by layer to achieve accurate extraction of ORB feature points. The experimental results demonstrate that the number of feature points of the improved ORB algorithm is not easy to fluctuate dramatically, and the accuracy gets improved, extracting time increases slightly, overall repetition rate is relatively high. These results are of reference value for the study of feature points matching, robot pose estimation and navigation.