Improved Monocular ORB-SLAM2 Inspired By The Optical Flow With Better Accuracy

Abstract

ORB-SLAM2 is currently the best open source SLAM system with high positioning accuracy and map reusability. However, when using a monocular camera in a dynamic environment, the accuracy will be disturbed by the moving objects. Besides, even though there are no moving objects in the frame, there is space for further improvement in accuracy. This article improves the feature point selection based on monocular ORB-SLAM2 system, by creatively using the idea comes from optical flow and then using the K-Means algorithm to classify the matched feature point pairs. The existing open source datasets are used for evaluating the improvement. Under the pre-requirement that the improved system should ensure the real-time performance, the positioning accuracy of the improved system has been significantly improved.